ZeroWidth Journal and Research

Decision capture

2026-10-31T00:00:00Z

We give an AI assistant a work conversation and ZeroWidth's tools, then check what it records. A good assistant records each decision that was made, once and with the right details, and records nothing when nothing was decided.

Store manager: I have a three-step plan to get more visitors who buy, and I'll watch the conversion rate. Step one is reshoot every product on a white background. Step two is rewrite the product descriptions. Step three is add customer videos.
Store manager: I'm only agreeing to step one: reshoot every product on a white background.
Store manager: Steps two and three are not decided, so leave them as ideas for now.
right: Reshoot every product on a white background
right Goal: More visitors who buy
right Metric: Conversion rate, up
right why: Only the step that was agreed, recorded once.
wrong: Reshoot every product on a white background, rewrite the product descriptions and add customer videos
wrong why: Records two steps nobody agreed to.

Each model runs every conversation in two setups:

Tools only The assistant has ZeroWidth's tools for recording and looking up decisions, and nothing else from us.
Tools and instructions The assistant also reads our short instructions on when to record a decision and what to include.

The models: Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6 Luna, Gemini 3.8 Flash and DeepSeek V4.1 Flash.

An answer is right when the assistant records the decision once, or records nothing when nothing was decided. Recording something that wasn't decided is a false positive. There are eight kinds of conversation:

A decision made The person decides something. Record the decision.
A decision reversed The person undoes an earlier decision. Find the earlier one, then record the change.
Someone else decided The person passes on a colleague's decision. Record the decision, with the colleague as the one who decided.
A plan partly agreed Only the first step is agreed. Record that step only.
A choice left for later "We'll decide after the board meeting." Record nothing. A note that the choice was left for later is allowed.
A choice handed to someone else "Sam will pick the vendor." Record nothing about the choice.
Not a decision An idea talked through and never agreed. Record nothing.
Off the record The person asks to keep a decision out. Record nothing.

In the pilot, every model recorded every decision it should have. Where the models differed was the other half of the job: recording things that weren't decisions. Our instructions cut those sharply.

Scoring

Every conversation is written from a fact sheet, so the right answer is set before any model runs.

Each record is checked against the fact sheet, detail by detail: what was decided, why, the goal, the metric and which way is good, when to check, and who decided. A number the person never said counts as made up. If the assistant calls tools six times in a row without replying, the conversation counts as failed.

When a choice is left for later, the assistant may note that the choice is still open. That note counts as neither right nor wrong. The strict chart counts it as wrong.

Every rate comes with a 95% range. Both setups run on the same conversations, so the gap between them is the effect of the instructions.

Private test set

From the first full run, we keep a quarter of the conversations private, so no model can be tuned to them. Each model is scored on both sets, and the page flags any model that scores much higher on the public set than on the private one. The pilot has no private set yet.

On Stratified Attention, Part 3 of 3, The Teachable Hive

2026-10-13T00:00:00Z

Part 3 - The Teachable Hive

In Part 1 of this 3-parter, I talked about how the notion of Stratified Attention was a kind of stealth problem-solving capacity that requires some measure of patience and trust - using a recent example from my own novel-writing experience as an instance where a form of subterranean heed came to my rescue in unsnarling some literary trouble I was having. In Part 2, I described the pitfall of the Division of Dignity Fallacy, which holds that humans = higher brain function; bots = grab a shovel, dumdum, and how ZeroWidth co-founder Pete Binggeser was a sound cognitive model for the kinds of orchestrated, observed thinking that’s going to maximize impact and effectiveness in this AI era. So far, all of this series has concerned itself with the individual scale of AI and its impacts, as well as the adaptations we believe will serve people well in folding AI-enabled technologies into their work. Let us now turn to the enterprise side of this Cognition Ledger.

To review a couple of the truths - truths that, at the individual level, are mostly taken to be self-evident:

Thinking is Work. We all know this. We all feel this. The trouble with the culture of too many organizations is that, given the reality that this form of work is not observable, it cannot therefore be supervised. As such, it too frequently goes unrecognized, too often gets dispensed with in terms of a company’s planning, forecasting, scheduling, etc. There can be this weird blind spot where the leaders of an organization or department will wish obviously for thinking to have taken place, but no allowances of time and space and personnel are made for it to take place. A misread of this reality might hold that such leaders are dimwitted themselves so have precious little experience with the act of thinking, so have no basis for including it in their dopey, dopey projections. And while, yes, there is no shortage of dim bulb so-called leadership out there (who among us has not at some point found ourselves asking of a misguided supervisor “HOW do you have this job?”) it is usually more accurate and fair to say that such lapses are more likely attributable to other pressures - what the competition’s up to, fast-looming deadlines, the expectations of higher-ups, the weight of economic demands both internal and external, etc.

Thinking is HUMAN Work. As tempting as we for some reason find it, there is no technology that can substitute for our ability to think. Probing analysis, genuine ingenuity, impactful synthesis - these are all manifestations of human capacities.The fallacy that any technological platform or system can step in and step up to supplant or even replicate these capacities must be resisted and refuted. Wherever it takes hold, companies are bound to yield to (very real) economic pressures of scale and expediency and permit these human capacities to be taken on by technological proxies that lack these fundamental abilities. The resulting “thinking” will prove naive, or deluded, or solipsistic, or derivative, or otherwise deficient in a dozen different ways. We at ZeroWidth have said this elsewhere and will continue to holler about it till it’s come to be widely ingrained in the public mind: AI is an INGREDIENT technology, an undeniably helpful and time-saving thing in certain respects, but offers NO adequate substitute for human thinking.

As a well-grizzled American person, in (what passes for) my adulthood, I have been present for at least three seismic, society-scale upheavals - the attacks of 9/11, the global financial meltdown of 2007-08, and the Covid pandemic. Each of these in their time were purported by many to represent a fundamental recasting of the nation’s spirit, a profound and irreversible course-correction of the path we were on, a tectonic remaking of the landforms underfoot. And while each of these crises had the potential to bring about this kind of foundational national transformation - of our character, of our priorities, of our aspirations - not one of them (to this observer, at any rate) succeeded in effecting anything like change of this depth.

In the wake of 9/11, a moment that could (and in my view ought to) have represented a period of reflecting on our roles as citizens and neighbors to each other and the rest of the world, our so-called leaders failed us. At what should have been a once-in-a-century moment to rally us, to galvanize by summoning the loftiest and most expansive of our aspirations, we were exhorted by not to elevate our sense of who we are and wish to become, but to constrict this sense, to regard ourselves foremost as cogs in the nation’s economic machinery. We were urged, in other words, to hit the mall and resume our places at the conveyor belt of heedless consumerism. Which turned out to have just been cover for a mess of bad policy.

Then when the global financial crisis torched a coupled generation’s worth of accumulated and attempted wealth, there was more bad policy to step in and save some of the very institutions that had caused the crisis in the first place. We were all urged to embrace new less material pleasures - to take up hobbies and volunteer, to devote our time to pursuits that were less oriented toward chasing money. We were enjoined to revise downward the scale of our prospects, to defer (or abandon altogether) our plans to buy first homes or retire in comfort. As individual consumers and citizens, we obviously are not too big to fail - investment banking and insurance firms are - so rescue measures are extended to these, while we all get urged to reduce the scope of our futures.

And when the Covid pandemic sequestered us all away from one another and caused many of us to lose our livelihoods, we were instructed to stave off the madness of our isolation by learning to crochet and approximating social gatherings and live performances via video conference. Millions across the country and world necessarily adopted remote work - there was a learning curve, obviously, but we did pretty well with it. The moment the plague-tent quarantines were eased, all the rhetoric about prioritizing employee wellness, all the “at Globotech, we’re family” sentiments dried up quickly when companies were confronted with the rude reminder about their financial burden of tons of downtown commercial real estate, a burden they were obligated to pay for whether there was workforce inside or not. Return-to-in-person-work mandates become widespread and workers - many of whom had discovered that for a variety of reasons (young kids at home, caregiving for elder relatives, their own mental and physical health, etc.) they preferred to work from home - were compelled to return to office whether they wished to or not. But even more salient to the point we’re attempting to make, here: the question of corporate culture and policy. Where the so-called leadership structures within organizations were based upon compliance and direct supervision, companies demonstrate a foundational lack of trust in their workforce.

A common thread I see in all these crises is that the onus of adaptation falls to individuals. Institutions - corporate, governmental, nonprofit - after (often begrudgingly) implementing temporary accommodations to upheavals of circumstance will revert to business as usual, shifting the work of long-term acclimation to permanent changes in prevailing conditions from the institution to individual workers and customers. This is reductive and wrongheaded and unfair, to be sure - but it is furthermore ineffective over the long term. It creates and perpetuates conditions where workers burn out and customers grow wary. If loyalty may be said to be one of the currencies an enterprise needs in order to prosper, then the ways it treats its people must factor into cultivating this currency. In order to earn trust and loyalty, organizations must demonstrate a willingness to and ability to adapt along with the individuals who work for and buy from them, a recognition that in the face of permanently changed conditions, they are capable of keeping pace with the changes their people must contend with.

Speaking for myself, my memory in this regard is long. Where I see companies that violate principles they’ve professed to hold, whether it’s bailed-out financial institutions engaging in predatory lending or consumer brands name-chacking national tragedies in cynical marketing or using my money to back fearmongering political candidates, tax-exempt churches politicking from the pulpit or companies whose rhetoric claims to value its people while its policies are dehumanizing and exploitative - I take such factors into account when deciding where to spend my finite dollars and where I devote my finite time and attention. I’m certainly not alone in this.

A vital concept to incorporate into our understanding of this present moment is that AI is unlike any preceding technology - it is its own CATEGORY of technology, in that it is the first to be legitimately considered to interact with us, to respond and adjust to us, to learn from and revise in relation to us. The precedents we have, our well-established tendencies to revert to our business-as-usual defaults, the “factory presets” on how we function as organizations cannot pertain. If we agree that the advent of these technologies heralds the arrival - not the promise, not the prospect, the arrival - of a new age of work, then the organizations we lead must demonstrate a readiness to be as adaptable as every individual in confronting this new reality.

So it follows that if companies resolve to retain talent, to prevent brain drain, they would be well served in the long run by devising and adopting policies that are as mindful of worker wellbeing as they are of their bottom line. In Part 1 of this series, I talked about the subterranean nature of some forms of cognition, and in Part 2 described how ZeroWidth co-founder Pete Binggeser engages in what I call Observed Cognition - both of these aspects of cognition have this in common: they are accretive. Each instance where my Novel Writing Brain is assured that its trust in the Unseen Depths is not misplaced results in a less stressful encounter with comparable future difficulties; each time Pete logs a Deep-Noticing aspect of his cognitive processing, it gets added to the overall sense of things, a new page in the operating manual, so to speak. You and he and I are Learning Animals, and as such have a native tendency to fold new inputs into the way we do things - everything from tips-n-tricks type stuff up through broad concepts and principles.

IF, then, we hold the above ideas in mind - that Thinking is Work, and furthermore that Thinking is HUMAN Work, and consequently that these forms of work have VALUE to an organization, then it makes operational and economic sense to create environments that allow for and reward this work. The role of trust becomes indispensable here, because thinking, crucially, is not readily observed. We can infer that someone is or is not thinking, but cannot empirically prove that they are. And so many of the systems we have in place that have shaped our work lives have to do with the collection of the trackable, the countable, the overt. The number of observable minutes we’re clocked in at our work stations perhaps makes sense if we’re on an assembly line or working in a foundry - settings where there is effectively a one-to-one relationship between physical presence and location-dependent effort and work product, where outputs are the direct result of time- and place-based activity.

But in work where ideas are the outputs, it makes far greater sense to make allowances for the idiosyncrasies of those responsible for generating them. As noted in Parts 1 and 2 of this series, ideas can require various things to fully develop - setting a problem down and turning your attention elsewhere, say, or having a conversation with a colleague about a seemingly unrelated topic, deep engagement with some other arena of activity, etc. In order for an organization to foster the kinds of excellence and innovation they all claim to want, they MUST cultivate the climate of trust and flexibility of approach that permit thinkers to think.

Which is a transition far bigger than the open-floor-plan-free-snacks-foosball-table approach that’s become a cliche of so-called creative workplaces, it necessitate a profound shifts to the culture of your organization - to the values you enact, the priorities you pursue, your means of assuring that the methods by which your organization operates are consonant with the principles it professes to hold. And we’re not talking about drafting some static Code of Conduct doc that’ll get tucked into a binder and never referred to again, but a lively, dynamic, responsive ecosystem that accommodates multiple modes, styles, cadences, and likely as not varying degrees of physical presence.

If (as we must) we accept the premises above as key to the realities of an individual in this changed technological landscape, which again, are::

Thinking is work.

Thinking is exclusively human work.

AI systems CAN have transformative power in helping us conceive and make stuff, BUT that

AI systems are INGREDIENT technologies only able of providing (important, yes, even vital) ASSISTS to human cognition that CANNOT substitute or supplant human thinking. (See the Distributed Cognition “Mindset Myths” posted last year to help frame and interrogate your own organization’s often unexamined beliefs on this question, and the default policies that arise from these beliefs.)

And that the following condition must therefore pertain to any organization that resolves to be as adaptable and responsive to the shifting technological landscape as it expects its employees to be: it must create and preserve an ethos that is thinking-friendly.

This will involve operational considerations, obviously - an elevated tolerance for remote work, say, or the cordoning off of Solitary vs. Communal spaces, etc. - whatever ways people concur would improve the prospects for the human work of thinking.

Organizations must become willing to remake our workplaces to reflect these realities. The “observe and report” and/or “surveil and control” models of compliance-based performance expectations are dumb. And boring. And a needless time suck. The weirdest element of the tenacity with which these models persist is that just about everybody agrees that they’re counterproductive. Yet the “feet flat, eyes front, at work station” mold - a structure better suited to a Victorian classroom or a guard towered prison than a 21st century workplace.

Also: the decades-old fallacy of “knowledge work” does not pertain - all work demands knowledge of various sorts and making divisions about which forms of knowledge have greater worth is a counterproductive form of snobbery. (co-founder John Cain ID’d a concept he calls the “Dignity Divide” HERE) Thinking about human work this way renders an organization unready and ill-suited to dealing with the realities of layering AI technologies into their processes - if there is the grit of this kind of snobbery in there, the “smart workplace” pearls you’re trying to grow will be flawed, they’ll reflect a warped set of assumptions and the processes arising from such deficiencies in planning with be full of snags and frustrations, your results will fall short of expectations that lead to more work and time spent, not less.

In short, for organizations to reap the benefits of these technologies, they must be as nimble and responsive to the changes they bring about as any team or individual on the payroll, any category of customer. The organizations that prosper in this upended landscape will be those that onboard the use of these technologies as mindfully as they would any key staffer.

An organization, in other words, must undertake some of the human work of thinking.

Shims

2026-09-22T00:00:00Z

One decision {demo=one-decision}

Most software is full of small judgment calls:

Which team should read this message?
Which filter did the shopper mean?
Does this clause need a lawyer?

Each one has a handful of possible answers, and each is made thousands of times a day.

Today those calls are typically made one of two ways:

You write keyword rules, which break the first time a customer phrases something differently.
You send the text to a large model in someone else's datacenter, which costs money on every call, adds a round trip, and means your users' words leave your product.

A shim is a third way. It is a decision model small enough to live inside the app: a few kilobytes of weights that answer one narrow question, from a set of answers your product already owns, in about fifteen milliseconds, without calling a server. A shim can be built in a few seconds from a list of examples. When the product changes, you throw it away and build another. A shim is the thin, disposable tool that sits between your application and your user, making the product fit a bit better without requiring a complete overhaul.

One is running here, live in your browser. In this example, a clothing-store search box aims to categorize what occasion the shopper is looking for, and it has five answers to choose from. Type your own input and it will instantly rank all five categories, with a number beside each for how sure it is. That is the whole of what a shim does. Everything is decided in your browser, and nothing you type is sent anywhere.

Since shims are fast to build, and fast to run, layering them and combining them can yield interesting results and novel solutions to often frustrating user experience problems.

Is this a classifier?

The part that decides, yes. A classifier sorts input into a fixed set of buckets, and that is what the head of a shim does. When building a shim, our compiler tries three standard methods, logistic regression, nearest neighbor, and a plain average of each answer's examples, and keeps whichever measures best. All three sit on an encoder someone else trained and published. If you have fitted a classifier before, you already know how this part works.

A classifier on its own is typically not something you can put in front of a customer. It answers every question it is asked, including questions outside its subject. Its confidence is a number between zero and one that is not the probability it looks like. It gives no sign when the input is outside its subject, cannot read a paragraph, cannot tell you how many answers it should have, and cannot tell when the world it was fitted to has changed.

Fitting a classifier takes an afternoon. Deciding what it may act on, what it must refuse, when to stop trusting it and how to tell takes much longer, and is usually skipped. A shim is a classifier with that work already done and measured.

What is inside one {demo=where-text-lands}

An encoder is a type of model that turns text into a list of 384 numbers. Those numbers are a point in a space where sentences that mean similar things sit near each other, and that is the whole of what the encoder does. It is frozen. Every shim on this page shares a single base encoder, and nothing about your examples changes it, retrains it, or edits it. It is a published model that anyone can download, and can be swapped for other encoders as new ones are released or preferences change.

A shim adds two things on top of the encoder:

A head: the small piece that turns the encoder's 384 numbers into one of the shim's answers.
A compressed copy of the shim's examples, one byte per number. That copy is what lets the shim tell familiar input from foreign.

That is the whole technique. A shim is not quite a model, but a method of leveraging an encoder and limited examples to build an ultra lightweight, extremely fast, and specific confidence based classifier. A shim can only separate things the encoder already places apart. Examples set where the lines are drawn, but no example can create a distinction the encoder does not make. That is where most of a shim's limits come from.

How the compiler chooses a head

A head is the part of a shim that decides. It takes the 384 numbers the encoder produced and returns one of the shim's answers, with a score for each. There is more than one way to build that piece from a list of examples, and none of them is best for every shim, so our compiler builds three and measures them:

A linear head. One direction per answer, fitted by logistic regression. Deciding is a dot product and a comparison. It is the smallest of the three and usually the most accurate once each answer has a few dozen examples.
A nearest-example head. The input is compared with every stored example, and the answer of the closest few wins. It costs no extra weights, because the shim already carries its examples for the familiarity check, and it is the one to use when an answer's examples sit in several separate places (a "red" that means blood, traffic lights and fruit), which a single direction cannot hold.
A class-average head. One prototype per answer, the plain average of its examples. The weakest of the three with plenty of examples, and often the most stable with very few.

The compiler fits all three on the shim's own examples, scores each on examples it held back, and ships whichever measured best. When two are within noise of each other, it prefers the one that can act on more input. The choice is recorded in the report beside the shim, and it is made again on every rebuild, so a shim that starts as a class average with twelve examples can become a linear head at fifty without anyone changing a setting. Every shim on this page went through that choice. Of the fifteen, six shipped as a linear head, eight as a class average, and one as nearest examples, which is about what you would expect from shims built on a few dozen examples each.

:::aside If you have built retrieval for a language model

All of this is familiar. The same kind of encoder turns text into the same kind of vector. A store of vectors sits beside it. A new input is placed in the same space and compared with what is stored, and the nearest match is the interesting one. The dashed line in the plot is exactly the query a vector database answers.

Retrieval fetches the passages behind the nearest vectors and typically hands them to a large model, which then writes an answer or generates some text. A shim's stored vectors are labeled examples, and the label of the nearest one is the answer, so there is no passage to fetch and no model to hand it to. When the compiler picks the nearest-example head, the shim is a vector search over a few dozen labeled rows with the label returned in place of the text. When it picks a linear head, the examples have been folded into one direction per answer, and the search happens at build time rather than on every call.

Familiarity comes from the same place. Every retrieval system computes the distance to its nearest match and most throw the number away once the passage is in hand. A shim keeps it, scores it against how far the examples sit from each other, and uses it to decide whether to answer at all. The number says how far this query is from anything the store holds, and a retrieval system has it available on every call.

A shim is small for the same reason an index over eighty rows is small. A retrieval system has a language model on the other end of the index, and a shim does not, and that difference is behind most of what a shim cannot do.

:::

Which encoder {demo=encoder-table}

The current encoder that comes with our shims-sdk is bge-small-en-v1.5, published by the Beijing Academy of Artificial Intelligence: a 33-million-parameter BERT that reads English only and returns 384 numbers per text. The page runs the ONNX conversion with its weights quantized to 8 bits. A shim is tied to the encoder it was built on, and every figure on this page is with this one.

The work started on all-MiniLM-L6-v2, which is the default small encoder in most browser demos. Eight candidates were then run on three test sets with the same head and the same math, shown in the table with this section:

32 hand-written charter clauses labeled by purpose, with no examples given, so the encoder alone does the separating.
The 16 clothing-store categories.
The 53 materials, the hardest separation we have.

The last two test sets are model-written, so their levels are optimistic and only their ordering means anything. All eight ran at 8 bits, because that is what ships.

The four newer small models sit within a few points of each other on the generated sets. On the hand-written set, bge-small is clearly ahead of the others. The two base models buy five or six points on the hardest set for three times the download and twice the time per text, and a product with a 32MB budget cannot make that trade. MiniLM-L12 has twice the depth of MiniLM-L6 and scores no better, so depth on its own bought nothing.

The table compares bare heads. Rebuilding the page's actual shims on bge-small gained less on the charter clauses than the table suggests, because the shim was already getting 22 of 24 right on MiniLM, and gained more where it mattered: the shims became more confident on the clauses they got right, and on 128 real clothing-store searches the category shim went from 21% to 36%. Changing an encoder also moves every distance a shim relies on, so every shim has to be rebuilt on the new one, and a shim's weights record which encoder they belong to so the two can never be mixed.

Sizes

The encoder is 32MB, downloaded once and cached by the browser. A shim with three answers and two dozen examples is about 18KB in memory, and half of that is the examples. A shim carrying a vocabulary, where the answers are names of things rather than kinds of message, can run to several hundred kilobytes, because a lookup is as big as what it looks up. The fifteen shims on this page total 1.6MB on disk.

A short query takes about 15 milliseconds in a browser and a support-message-length input about 30. Both figures are from headless Chrome on an M-series laptop, single-threaded, and neither has been measured on mid-range hardware.

What an encoder buys

Every number on this page is also measured against having no encoder at all. Keyword matching, BM25 with textbook settings and no tuning, is under a megabyte and takes microseconds. Given the same examples, the shim beats it by 8 to 15 points on the three intent benchmarks, and the margin is widest when examples are fewest, which is the situation a new product is in. On sentiment, keyword matching cannot move off chance and the shim sits 30 points above it.

We also tried the cheapest possible encoder: a lookup table distilled from bge-small itself, every word in its vocabulary embedded once and a text scored as the average of its words. It is 11MB and about four hundred times faster than the model it came from. It also loses to keyword matching on every benchmark, by 46 points on the banking one. Averaging words throws away which words appeared together, and that is most of what separates one banking question from another. Below 32MB, keyword matching is the better tool. The transformer's job is to read words in context, and that is what the 32MB pays for.

Confidence you can act on {demo=calibration}

Whether a shim acts, offers, or stays quiet is a decision about that number beside the winner, so it matters that "70% sure" means right about seven times in ten. A number that means what it says is called calibrated. Out of the box, a head's number is not. On a shim with 150 answers, the head reported under 10% confidence on every decision while being right nine times out of ten.

The demo shows where the number comes from. The head is the part of the shim that decides: one row of weights per answer, and the search's 384 numbers scored against each row. Those scores are turned into percentages that add up to 100. The first view shows those percentages as the head computes them, and they sit close together, because the compiler keeps the head's weights small to stop it memorizing its examples. Small weights mean small gaps between scores, and small gaps mean a flat spread of percentages, however clear the winner is. The more answers a shim has, the flatter the spread: across 150 answers the winner's share is small even when the ranking is right.

The compiler corrects this with one number per shim, called a temperature. It stretches or squeezes the gaps between the head's scores before they become percentages, without changing their order, and the compiler fits it on examples the head did not see during training, so that the percentages match how often the head was actually right. The shim applies it for you; the second view of the demo is what the shim reports. The temperature is not a dial to set by feel. What you can do is refit it later on real outcomes: the SDK's recalibrate() takes the accepts and corrections a product collects and rescales confidence against the traffic the shim actually meets rather than the examples it was built from.

The gate

Once the number can be trusted, the compiler reads one more thing off the examples it held back: the confidence above which the shim was right nine times in ten. That is the gate, the amber mark in the demo. Above it the shim acts on its answer. Below it the answer is only offered, for a person or a larger model to confirm. To act less often, raise the gate. It is a policy you can state and audit, which a temperature is not.

"Above this confidence the shim is 90% right" is read off test examples in which every answer is about equally common, and real traffic is not. Uneven traffic alone turns out not to matter: a steeply skewed stream left the promise intact on all three tasks we tried. What breaks it is the hard answers also being the busy ones. Then a gate promising 90% delivered 80 to 84.

The repair needs no labels. From about a hundred unlabeled inputs the shim estimates how common each answer really is, from its own calibrated probabilities, and re-reads the gate for that mix. The SDK's refitGate() brought it back to 89 to 92%, and the cost is that it acts on less. What it cannot see is an answer it gets confidently wrong, because the estimate is made from its own predictions. There, only real outcomes help.

:::measured

The standard measure of calibration groups decisions by how confident the shim claimed to be and, in each group, compares the claim with how often it was right. The average gap across the groups is the expected calibration error, so 0% is perfect. Before and after the fitted temperature, on three public benchmarks: CLINC150 went from 36.3% to 1.0%, Banking77 from 34.9% to 1.9%, and MASSIVE from 30.9% to 5.3%. Accuracy is unchanged, because a temperature cannot reorder answers. On small sets of examples the fit can make things worse, so the compiler keeps it only when it improves calibration on the examples it held back; three of the fifteen shims on this page ship with no temperature for that reason.

:::

Declining to answer {demo=refusal}

A classifier always answers. Ask a clothing-store model about a boiler repair and it will say "outerwear", with real confidence, because something has to win. That behavior is a large part of why small models are hard to put in front of customers.

So every shim ships its examples alongside its weights, and measures how close the input is to the nearest one. That distance, scored against how close the examples sit to each other, is what we call familiarity. Far from everything it has seen, the shim declines.

The demo with this section is a shim that sorts clothing-store searches into categories. Ask it about a pharmacy and the head is still sure of some category, because a head has no way to say "wrong shop". Familiarity is what says it.

The floor is fitted per shim at build time, aiming to wrongly refuse about one search in twenty. For a while we had it as a hand-set constant instead, and that constant was turning away a third of legitimate searches. Then we measured the fitted floor against 300 searches real shoppers had typed, and it turned away between a third and five in six of those too. The floor is set by asking "how unfamiliar are the least familiar one in twenty of the inputs this shim should accept?", and we had asked that of the wrong inputs: the shim's own examples rather than real traffic. The repair needs no labels. refitFloor() in the SDK takes a few hundred inputs your app treats as in scope and refits the floor to them. This is the weaker of the shim's two gates, and the refit is one you have to do.

Confidence and familiarity are different numbers

Confidence is how strongly a shim prefers one of its own answers to the others. Familiarity is whether it has ever seen anything like this. Only the second can say "not mine".

Confidence cannot, because a shim can only answer in its own categories. Ask a billing shim about closing an account and it will tell you, at 99%, that this is a failed payment. That makes familiarity comparable across shims in a way confidence is not.

:::measured

On CLINC150, a public benchmark of 150 intents with 1,000 requests that belong to none of them, a fitted floor caught 85% of the out-of-scope requests while wrongly refusing 6.5% of the real ones. The old constant caught 99% while refusing 32%. Measured as a single score for how well familiarity separates the two groups across every possible floor, where 1.0 is perfect and 0.5 is a coin flip, that benchmark scores 0.96.

Against the 300 real clothing-store searches, the same score is between 0.71 and 0.88. The mechanism carries over and the benchmark's numbers do not. The benchmark's out-of-scope set is a different topic, and real queries are terse strings that resemble other terse strings whatever they are about. These figures have not been re-run since a later change to how text is encoded; the accuracy figures, which were re-run, moved by less than a point.

:::

While someone is still typing {demo=typing}

The demos above assume the input is finished. In a live product it usually is not. A shim cheap enough to run on every keystroke gets asked most of its questions halfway through a word.

The answer wobbles while the text is being typed, and we measured how much. Take one character off the end of a finished message and the answer changes on one input in nine. That figure is not portable: the same test on a different task gives one in twenty. Anything that shows a suggestion live has to expect visible flicker and measure its own rate, because ours differed twofold between two public benchmarks.

A shim running per keystroke has already encoded the last few prefixes and thrown them away. Keep them, and ask whether they agree. When the last three prefixes land on the same answer, that answer is wrong about one time in forty. When they do not, about one time in four. That is a five- to eight-fold difference in error rate, available on every input, for no extra work.

Whether the answer has stopped moving is a third way to know you should not act, alongside how sure the shim is and how familiar the input is. It is independent of the other two, because a shim can be fully confident and fully familiar on a word it has not finished reading.

It is also a clear example of what running locally buys. A decision that costs nothing can be made speculatively and thrown away, and the ones you throw away leave a signal behind. A hosted model would bill three round trips for this number.

:::measured

Two benchmark tasks, 400 messages each, the last six prefixes of every message decided one by one. The answer flips on the last character in 10.8% of CLINC150 messages and 5.0% of Banking77 ones, costing 7.5 and 0.8 points of accuracy. The last three prefixes disagree on 16.0% and 12.3% of messages. Where they agree, the answer is wrong 2.4% and 6.0% of the time; where they disagree, 20.3% and 28.6%.

Averaging the last two or three vectors, the obvious free fix, never beat simply using the full text by more than one message in four hundred, and cost three points on CLINC150. The full text is the most informative prefix, and averaging it with shorter ones only dilutes it.

:::

When the input is longer than a sentence {demo=long-input}

The inputs above are a sentence long. In practice, a support message rambles, a contract clause runs ninety words, a review is a paragraph.

Every encoder of this kind has a fixed reading length, and none of them complain when you exceed it. They read the opening, drop the rest, and hand back a vector that looks healthy. On a message of about 140 words, the truncated vector still resembles the full one (a similarity of 0.78 on a scale where 1 is identical) while its resemblance to the decisive final sentence is 0.04. Nothing reports that anything was lost. This catches people who call an encoder themselves; a shim measures the input first.

The encoder on this page reads about 1,700 characters in one pass, and a high ceiling costs nothing, because the cost of a read follows the text you give it rather than the ceiling you allow.

The number that matters is smaller. Past about 420 characters a shim stops reading the message whole and reads it in pieces, well before it has to, because reading in pieces measurably beats reading whole on the kind of message that buries its point in one line. Averaging a document's words drowns the sentence that matters, the same way averaging its sentences does.

The shim does all of this itself. An app calls decide(text) exactly as it would for a sentence; the shim measures the text, splits it at sentence boundaries into pieces of about 160 characters, encodes each piece, decides on each, and combines those decisions into one answer before decide() returns. Nothing about the length of the input changes the code that calls it.

Combining the pieces into one decision can be done two ways, and this is the one choice in the process that is left to the developer:

Read the most familiar piece and let it decide. This is the default, and an app that never sets it gets this.
Weigh every piece by how familiar it is. An app asks for this per call, with decide(text, { reduce: "pooled" }).

The compiler does not make this choice, because the right answer depends on the shape of the documents an app receives rather than on the shim's examples. Weighing every piece can go wrong in one particular way: a piece the shim finds very familiar gets a full vote even when it is off the point, because a sentence being familiar and a sentence being on the point are different things, and only the first is being measured. Reading the best piece avoids that and misses evidence that is spread across several sentences.

On the public benchmarks the two finish close. Weighing every sentence is worth about ten points on documents that spread their evidence, and reading the best piece is worth about fifty on a document with one decisive line; over a mix of both they finish within half a percent of each other, which is why the default is the simpler one and the other is a per-call option. A contract reviewer asks for the weighted reading, because a clause carries its risk across a proviso and a main limb. A support inbox keeps the default, because a rambling customer usually does bury the point in one line.

:::measured

Two shims with 77 and 150 answers, over a hundred documents of each shape: documents with one decisive sentence among filler, and documents that spread their evidence across every sentence. Reading the best piece scored about fifty points higher on the first shape; weighing every piece scored about ten points higher on the second. Across an even mix, the two methods finished within half a point of each other.

:::

One encoder, many shims {demo=three-questions}

A single shim answers one question. Most products have several questions to ask of the same text, and this is where a shim's shape pays off. The encoder is the one piece that is large, and it reads the text once. A shim is a few kilobytes on top of that reading, so a second question costs a fraction of a millisecond, and so does a tenth.

Three are running live with this section, each asking something different of the same support message: where should it go, how urgent is it, and what tone was it written in? Each one was built on its own, from its own examples, and each reports the two numbers you have now met: how sure it is, against its gate, and how familiar the text is, against its floor. Type your own message, or try the banana bread recipe and watch all three fall through their floors while still fairly sure of an answer.

A shim can be sure enough to act, unsure enough to only offer, or so far from its examples that the right answer is silence. Because every shim reports the same two numbers, a product can wire several together and apply one policy to all of them.

Wiring several shims together {wide=deep-tree demo=wiring-file}

The bank above asks its questions side by side. Real features also need questions in order, where the answer to one decides which shim is asked next. A support message needs a team, then a reason. A clinic message needs an emergency check first, by rules and never by a model, and a person at the end for anything the shim will not commit to.

The developer writes that wiring, as a JSON file next to the shims; the one shown with this section is the file the figure runs from. It lists the steps in order: a rule to run first, a shim to ask, a route that picks the next shim from the last answer, and what to do when a shim is unsure or the input is unfamiliar. Each shim in it is built on its own, from its own examples, and has no link to the others. Because the wiring is a file, it can be reviewed in a pull request, and the compiler checks it at build for:

an answer with nowhere to route
a route to a shim that does not exist or is still a draft
a route on a later step
a route for an answer the shim never gives
an invalid rule pattern
a step routing on a score that is still provisional

At runtime the app hands that file to the SDK's System and calls run(text). The system encodes the message once and walks the steps, asking each shim in turn about the same vector.

Running a deep tree is cheap. The expensive step is reading the message, and that happens once no matter how many questions are then asked of the result. Five questions cost one reading and five small pieces of arithmetic. Being right at depth is harder.

A tree does not need to be symmetrical. A refund earns a third question, whether it was a duplicate charge, a service problem or a change of mind, while a feature request is finished after one. Forcing every branch to the same depth invents questions nobody asked, and each invented question is another place to be wrong.

The failure that appears at this depth is in the route as a whole, even when every question clears its own bar. Five questions at 90% each leave the whole path at 59%, so a message can pass four confident checks and still end up somewhere absurd. The system multiplies the routing confidences along the path and reports the product, and the wiring file sets a minimum for it, so a route that is weak as a whole is reported as a failed classification even when every step passed, and the app decides what to do with it.

Try "where is my parcel". This company does not ship parcels, so no answer on the tree is right, and every shim still picks one, because a shim can only answer in its own categories. The path confidence is the only number on the figure that speaks for the route as a whole. On this build it is not low enough to stop the message, and it reaches a specialist's answer with a confident-looking path. The path bar catches some of these. The rest can only be caught by noticing, across many messages, that a question is arriving which no answer covers.

Familiarity as a tiebreaker {demo=sibling-vote}

Without a tiebreaker, a routing step that fails its confidence gate stops the tree. Every shim below that step is never asked, and the message ends with no classification. The app decides what happens next; in the wiring file above, that is a hand-off to triage.

Stopping there throws away information the tree already has. The shim that failed was built only on the coarse categories at its level. The shims below it were each built on one category in detail, and one of them will often find a message familiar that the step above could not place. Familiarity is the number that says so, and because every shim scores it on the same scale, the scores of different shims can be compared.

The tiebreaker uses that. When a routing step is under its gate, the tree asks each shim it could have routed to how familiar the message looks. The message goes to the shim with the highest score if two conditions hold:

The score is above that shim's own familiarity floor.
The score beats the runner-up by a set margin.

If no shim meets both, the tree stops as it would have without the tiebreaker. Nonsense text, which no shim finds familiar, is never routed. The check costs no extra encoding, because every shim reads the vector the failed step already produced.

The result is a route that would otherwise have been lost. It is not a confident classification, and the tree reports it as a rescue rather than a normal step, so the app can treat it differently. But it is the specialist's own reading of the message, which is better than no answer at all.

Learning from the people using it {demo=memory}

The hardest part of building a shim is the beginning. Nobody has labeled examples of their own product's decisions. A shim needs only a handful per answer to compile, and they can be written by hand in an afternoon. We have also experimented with having a larger model write the first examples from a one-line description of each answer, which is how the shims on this page were made; it works, with the caveats in the last section. Either way, the examples a shim starts from are a rough draft, and the shim gets better from use. Every accepted suggestion and every dismissal is a label from the person using the product.

The SDK handles a correction in two stages. It stores the correction immediately and applies it to any later input worded closely enough, so the user sees the effect at once. Changing the model itself is a separate step: the developer adds the corrections to the shim's examples and rebuilds it, and that change is reviewed like any other.

We tried the obvious alternative, retraining the head in the page as corrections arrive, and it breaks things it was not asked to change. Live retraining is the design most people reach for first, and the measurement below is why the SDK does not do it.

How far a remembered correction reaches is set by how similar a new input has to be to the corrected one. Rephrasings of the same search score a similarity of 0.83 to 0.97 with the original, and unrelated searches stayed below 0.79, so the match threshold is 0.82.

:::measured

On three shims, retraining the head on a few dozen corrections moved accuracy on everything the corrections did not touch by +6.3 points, −4.2 points and −38.7 points. Remembering the same corrections, as the SDK does, changed nothing elsewhere. Adding them as examples and rebuilding moved the same three shims by +3.1, −6.3 and +1.8.

:::

Where the labels come from

Users rarely confirm anything. Explicit feedback widgets collect very little, and what they collect skews toward the annoyed. So the loop cannot depend on the user knowing they are teaching. Instead, the user's next action is the label. The shim proposes, the user acts, and the difference between the proposal and the action is the training signal.

Four outcomes cover it:

Corrected, with the right answer.
Rejected, with no right answer attached.
Accepted, meaning the user proceeded without changing it.
Ignored, meaning nothing observable happened.

The last two differ, and treating them alike teaches a shim to be timid, because a suggestion scrolled past was never refused. Which user actions count as which is up to the application. A filter the user edits after the shim set it is a correction. An undo within a few seconds is a rejection. A ticket a human reassigns after the shim routed it is a correction, and the strongest kind.

The signals carry different weight. Corrections carry the accuracy. Acceptances only keep the boundary from drifting, and one on its own says very little. Weighting them equally is the most common way to get this wrong. Abandonment cannot teach the right answer, but a pattern of users abandoning decisions of one shape is what makes a shim uncertain in the right places, and uncertainty is what the gate routes on.

Noticing a question you have no answer for {demo=radar}

Refusing and correcting both assume the answer exists and the shim missed it. A harder failure is when the right answer is not one of the options. A feature shipped last Tuesday, a policy change, an outage at a payment provider. No amount of correcting helps, because there is nothing to correct it to.

A single shim has no way to report a missing answer. A stream of inputs can show it. One unfamiliar message means only that the shim should not answer it. Several unfamiliar messages that resemble each other are a topic, and the shape of a topic is visible without knowing what it is about.

There are two ways to watch for that, and we measured both on a week of tickets we wrote for the demo. Flagging every unfamiliar message finds every ticket about the new topic, and it also flags seven ordinary tickets, and seven false alarms a week is enough for someone to switch the alert off. Flagging an unfamiliar message only when it has unfamiliar neighbors misses two of the ten and raises no false alarms. The useful signal is how close a message is to the other unfamiliar messages, and a cluster of unfamiliar messages is also what tells you what the missing answer is about.

The output is a short list handed to whoever owns the product, saying that these arrived together this week and nothing here covers them. That is the moment to write a new answer, and then the feedback loop has something to learn.

The list does not have to wait for a person. The cluster is already a set of example messages, so a larger model can be asked to name what they have in common, the cluster becomes the first examples of a new answer, and the compiler rebuilds the shim. Every build produces the same report, with accuracy and recall per answer, so the new build can be checked against the old one before it replaces it: the new answer has to be found, and every existing answer has to score as it did before. A rebuild that fails that check is a rebuild that does not ship.

How good is it, and compared with what {demo=escalation}

The table compares every way we could think of to answer the same question, on four public benchmarks. Every row gets the same examples and is scored on the same test items, which none of them saw while being built. Each number is the average of three runs with different random samples of examples, English only. The two hosted rows run on a server, and they are the yardstick.

	CLINC150	Banking77	MASSIVE	SST-2
Keyword matching, on device	84.6%	79.4%	66.3%	54.0%
Static word embeddings, on device	73.9%	33.5%	55.1%	50.4%
A shim built from descriptions alone, on device	86.1%	65.6%	73.8%	86.0%
A shim, 32 examples per answer, on device	95.0%	90.8%	80.2%	85.4%
An off-the-shelf zero-shot classifier, hosted	52.0%	47.0%	—	—
A frontier model given the answer names, hosted	96.5%	88.2%	92.3%	96.0%
A frontier model given 8 examples per answer, hosted	97.9%	93.2%	—	—

The last row settles one question. Hand the large model the same examples the shim was built from and it wins everywhere: 97.9% against 95.0% on CLINC150, 93.2% against 90.8% on Banking77. A frontier model is more accurate at this, and nothing on this page claims otherwise.

Everything above the bold row runs on the device and scores lower. Everything below it is hosted. That gap is the claim this page makes.

Keyword matching is the row to worry about most, because it is free, under a megabyte and instant. A shim beats it by 10 to 15 points on routing and by 31 on sentiment, and the margin narrows as examples arrive, so the encoder earns most when you have least data. Keyword matching with a few dozen examples beats a shim with none, which is the verdict on the no-examples row.

Static word embeddings are the row we most wanted to work. A lookup table of word vectors, no transformer, 11MB and a hundredth of a millisecond. It loses to keyword matching on all four tasks and collapses on Banking77, whose answers are separated by which words appear together. Averaging context-free word vectors throws that away, and it is the clearest measurement we have of what the encoder is for.

The off-the-shelf zero-shot classifier is a published model that sorts text into any labels you name, with no examples and no training, so it is the fair comparison for a shim built from descriptions alone. The shim wins, 65.6% against 47.0% on Banking77, at a twelfth of the size and a hundredth of the latency. At 402MB and two seconds a decision, the zero-shot classifier is in a different product category regardless of the score.

The no-examples row is the noisy one. Eight generated sentences per answer is a thin enough footing that a very small change in the vectors changes which head the compiler picks, and between two runs its cells moved by more than three points in opposite directions. Treat that row as accurate to a few points; the row with examples moved by less than one.

When to send a decision away

A shim can be less accurate than a frontier model and still be the right choice, if its confidence marks the decisions it should not be making, and those are handed on. Confidence that means what it says makes that possible: the number is comparable across decisions, so "anything under this goes to the big model" is a policy you can set and audit.

The gate can be measured rather than guessed at. Score every message twice, once with the shim and once with the frontier model, then split them at a gate: the shim keeps what clears it, the rest goes to the frontier model, and each half is scored on the decisions it would actually have made. The result is the accuracy of the whole stream at that gate, which can be set against two fixed points: sending everything to the frontier model, and the shim alone.

On CLINC150, a gate of 0.9 keeps 75% of decisions on the device and scores 97.3% overall, against 97.9% for sending every decision away. Three quarters of the traffic becomes free, instant and private, and the whole stream gives up about half a point. The decisions the shim keeps are 98.6% right, higher than the frontier model's own average, because a confident shim is keeping the easy ones.

Banking77 is the harder task. A gate of 0.8 keeps 73% on the device and scores 92.4% overall, less than a point behind sending everything away. The shim alone is 88.4% there against the frontier model's 93.2%, a wider gap than on CLINC150, and differences of that size between tasks are what to expect on your own.

An 86% accurate head is only useful if it can tell you which 86%, and the gate is what tells you.

:::measured

The split is measured on 188 in-scope messages from CLINC150 and 250 from Banking77, the ones the frontier model was run on. Each message carries the shim's calibrated confidence, whether the shim was right, and whether the frontier model, given the same eight examples per answer, was right. Nothing is extrapolated; every score at every gate comes from decisions that were actually made.

:::

Where shims work well {wide=use-cases}

To see whether shims hold up outside the benchmarks, we built six small proofs of concept, each for a situation a real product might have. For each one we wrote a short description of its answers, had a larger model write a few dozen examples from those descriptions, and let the compiler build the shims, which takes a few seconds. They are running below, on the same inputs we tried them with.

What it cannot do {demo=limits}

A shim cannot write anything. It cannot pull a value out of text, and for sizes, prices, dates and names, plain rules beat a learned model anyway: on the one extraction task we measured, rules scored 77 and a learned head 59, on a 0 to 100 scale that counts both misses and false finds. It struggles with negation and with who did what to whom, because sentence embeddings capture what a text is about more than how it is arranged. It is weak where the difference between two answers is a verb rather than a subject, which was 35% of all the errors on one benchmark.

It holds no facts about the world. This is the largest limit, and it is easy to mistake for a problem with short input. Cut a sentence to its first two words and a shim falls from 89% to 13%. Give the same two-word fragments to a frontier model and it scores 20% against the shim's 21. The information had left the sentence, and the shim was doing as well as anything could with what remained.

Real short queries are a different matter, because they are short and complete. On 147 clothing-store searches typed by hand, a frontier model given only the answer names got every one right and the shim got fewer than half. On grocery searches, 94% against about a third. "Anorak", "henley?" and "pinafore denim" are not ambiguous and not short of information. Knowing that an anorak is outerwear is a fact about the world, and a 32MB encoder with nine examples per answer does not hold it.

Most of that gap closes, at a price. Give the shim examples written for vocabulary, the words a shopper might type rather than the ways they might phrase a request, and let it answer from the nearest one, and the gap on hand-typed searches roughly halves. Only the nearest-example head benefits. Averaging forty-eight garments into one direction makes a vaguer direction, and those heads got worse as words were added. A lookup is as big as what it looks up, so the weights grow two to three times.

:::measured

Scored by holding out a fifth of the hand-typed searches at a time, the clothing shim went from 43% to 83% once its examples covered the store's vocabulary, and the grocery shim from 35% to 67%. The weights grew three times and 2.3 times.

:::

So the rule of thumb for where a shim belongs is whether answering takes reading or takes knowing, and length has little to do with it. On messages, commands and clauses a shim is within a few points of a frontier model. On catalogs it has to be handed the vocabulary first. The search box that opens this page asks what the shopper is dressing for, which is a reading question. Asking which garment they named would be a knowing one.

It also cannot judge things nobody can check. We built a demo that flagged risky contract clauses. The flags looked plausible and nobody involved could verify them. The same shim was fine at filing clauses by type, a question with a checkable answer, and that distinction is the useful one.

It works in English. The encoder covers no other language, and a two-kilobyte character-n-gram library tells languages apart better than a shim can, so use one.

Scores move between model-written text and text people type, and in a predictable direction. Text written by a model is about twice as easy to classify as text a person types, so a score measured on a generated test set is an upper bound. A shim built from generated examples runs the other way: one reported 49% when tested on its own generated examples and scored 72% on searches a person typed, because the generated test set was harder on it than real searches were. The number that matters is the one measured on the traffic a shim will actually see, and the SDK's observe() and outcome() are there to collect it.

Where this stands

What is established: on the benchmarks we ran, a shim beats every cheaper way of making the same decision, and it is the only one of them that is both small enough to ship and accurate enough to act on. A short query costs about fifteen milliseconds and nothing in money. The confidence it reports means what it says, and a gate fitted on nothing but generated text, promising 90%, delivered 84 to 96% on text people wrote.

What a shim does not do: it does not out-score a frontier model given the same examples, on any task we tried, 97.9% against 95.0% on CLINC150. The case for a shim is narrower than that, and it holds. At a confidence gate of 0.9, three in four decisions never leave the device and are 98.6% right, against 97.9% for sending every one of them away. It earns its place because its confidence marks which decisions it should be making.

Using it {demo=build-a-shim}

Everything on this page runs on one published package, the same one you would install. It holds the runtime an app imports and the compiler that builds a shim, with nothing to wire and nothing that phones home.

npm install @zerowidth/shims-sdk

A shim is a JSON file: a question, its answers, and a few examples of each. You write the examples; the compiler needs at least three per answer, and does better with a few dozen. The same source is filled in with this section, and the compiler that builds it runs in your browser, so you can build this one, change it, and build it again.

{
  "name": "urgency",
  "question": "How quickly does this need attention?",
  "context": "Messages customers send to a B2B invoicing product's support inbox.",
  "labels": ["now", "soon", "whenever"],
  "examples": [
    { "text": "the whole site is down for us", "label": "now" },
    { "text": "invoice #4410 has the wrong VAT number", "label": "soon" },
    { "text": "any plans for a dark mode?", "label": "whenever" }
  ]
}

The compiler turns it into a weights file and a report. Commit both. The source is what a teammate reviews, because they can argue with a label in a pull request. The weights are a build artifact nobody reads, like a lockfile, and the same source rebuilds to the same one.

npx shim-compile machines

The report is the half a reviewer reads, and the builder here hands one back:

accuracy on examples the compiler held back, with its interval
recall per answer
the confidence gate the compiler fitted, and how much traffic clears it
the familiarity floor
which head it chose
notes about anything that looks wrong with the task itself

A shim with an answer the compiler cannot learn fails to compile and says so.

An app imports the weights like any other file. The three branches are the whole interface.

import { Shim } from "@zerowidth/shims-sdk"
import urgency from "./machines/urgency.weights.json"

const shim = new Shim(urgency)
const r = await shim.decide("the whole site is down for us")

if (r.action === "act") route(r.answer)          // clears the gate it earned at build
else if (r.action === "suggest") offer(r.answer) // sure enough to show; something else confirms
else handOff()                                   // nothing like this in its examples

There are no thresholds to pick. The shim decides which of the three it is, from the gate and the floor the compiler fitted on decisions it had not seen. The encoder runs in a worker the SDK starts itself, so the first decision waits for the download and none of them land on the UI thread. preload() warms it early, and preload({ remoteHost }) serves the encoder from your own host instead of the Hugging Face hub.

Several shims on one message share the reading. Bank holds a set of shims, runs the encoder once, and asks every shim about the result. System takes a wiring file and runs it, rules first, then shims and routes in order. observe() records every decision, and outcome() reports what happened next; refitFloor(), refitGate() and recalibrate() use those records to refit a shim's gates on real traffic.

On Stratified Attention, Part 2 of 3

2026-08-19T00:00:00Z

Attendant to the rise of AI has come what we’ll call the Division of Dignity Fallacy - which holds that the human actor gets to do all the fun, sexy thinking - your decisionings, your judgmentations, etc. - and that the AI agent gets stuck with all the mental scut-work - your computings, your collationings, etc. This is perhaps broad-brush true, but is at best reductive. Such a (falsely) sharp divide ignores the symbiotic character of this newly hatched co-cognition. disregards Much of the difficulty we face when endeavoring to fold AI into our work has to do with grossly overestimating what it can do, broadly misinterpreting what it is for. A month or so ago on LinkedIn*, I wrote up this sketch of co-founder Peter Binggeser’s workflow/mode of working/mindset about work - as a model of the kind of stratified attention we’re talking about here, as relates to AI use, you could do a lot worse than adopting something of Pete’s approach. His effectiveness - in an array of areas from woodworking projects to chicken care to consulting with corporations and nonprofits to building new capabilities into zv1, the ZeroWidth AI workbench, to reconfiguring his scratch-built synthesizer - is due as much to Pete’s eclectic-and-growing skillset as to his impressively plate-spinning cognitive style.

* NB: If you find you sometimes don’t wish for the Entire Big Honkin’ Multi-Course Buffet of these posts, following ZeroWidth on LinkedIn is where you can find our more Tapas-Sized Small Plate Posts - bon appetit!

Pete’s is what I’ve come to regard as a profoundly Observed Cognition, meaning some portion of his considerable brainpower is dedicated to witnessing what is of interest, what things on his present docket are puzzling and needs solving, what’s a looming problem that he can’t yet even describe. This Observed Cognition is not Policed or Surveiled Cognition of the anxious or shame-based variety - seizing on “wrong” thinking and castigating oneself for having thought it - nor is it the Compulsive Cognition of the fixated who tumble into a thought-trench and have trouble scrambling out again. Pete’s is an Observed Cognition is more like there’s an interested bystander always present, a tolerant and keen-eyed witness, a bemused well-wisher, maybe, who takes note of everything, keeps a running tally of relative importance and urgency, which contingencies are connected to what outcomes, which components are dependent upon resolving other factors in order to progress, etc.

When Pete walks us through a demo of new features to the zv1 workbench, say, he’ll hit the inevitable glitch or a snag or some little irritant impeding process or functionality, and will simply fold that new item into his perpetually unfurling to-do list and carry on without breaking stride. As with my plotting quandary cited in Part 1 of this 3-Part series, there is the sense of the item in question not disappearing from Pete’s mental view, just receding to some more dimly lit corner of consciousness where it’ll get tended to later.

In fact, there is an accretive effect of these (seemingly) stray bits and bobs that begin as small-scale frustrations and misfires - these can aggregate inside the machinery of a Pete-brain into larger-scale conceptions - recently this tendency coalesced into the seeds of fundamentally reimagining ZeroWidth’s ENTIRE approach. In rapid consultation with John, the two of them essentially concluded that the enterprise must adopt an altogether different position within the fast-moving AI landscape, and that the direction the company had been heading would not be the course it continued pursuing. In a future post, I will tell the story of this transformation, but it’s salient here to note that story arose not from analysis carried out and recommendations made by AI, but from the deeply noticed realities of process and priority throughout the company’s work, a meta-cognitive act that constitutes the highest form of the Stratified Attention we’re talking about - a form of attention not fixated on task or project or plans, but that funnels all that upward to consider one’s place in the world and one’s orientation and bearing in the map one is making.

Mindset Sidebar: Pete trained as an architect, John as a designer - both of which are without question creative fields. But I’m struck over and over again by the comparative lack of ego in both of them - my background is in the arts, where the Self-Doubt Meltdown and Various Forms of Flying Off the Handle are commonplace, usually attributable to some form of ego-fueled maladjustment. The orientation I see in Cain and Binggeser centers squarely on the project at hand, the obstacles in need of overcoming, the optimal pathway to best outcomes. As such, there is little to no ego investment in “ownership” of ideas, or in the eager credit-claiming/blame-deflecting that’s so commonplace in the arts. True, yes, performing in a play that requires raw, volcanic emotion is inherently more vulnerable-making than plotting out the HVAC for a shopping center or whatever, so the actor’s neuroses are more comprehensible, maybe, but are no less draining - of time, of colleagues’ good will, of focus, etc.

Takeaway: Your existing capacity for Stratified Attention is an indispensable asset when folding AI-enabled platforms into the performance of your work. This capacity permits you to deploy AI across multiple timescales and degrees of urgency, have it stress-test approaches and prototypes while you consider next moves, solicit the feedback and input of others while you map out something else. This is a distinction between ability (which you have) and skill (which you might not yet have, but can learn); between what we have (intrinsic) and what we acquire (cultivated).

In another piece on the ZeroWidth LinkedIn page, I describe something of Pete’s considerable facility with Stratified Attention (“Mind as Mixing Board” - HERE) that have him toggling pretty seamlessly between Virtual and Actual, between Home and Work, between multiple Arenas of Inquiry and Activity. What we don’t make much mention of that piece is a phenomenon Pete notes specifically pertaining to a highly particular degree of difficulty associated with this emergent form of work, namely the Decision Fatigue that attends this kind of executive function co-cognition with AI-enabled platforms.

In short, there is a cognitive toll, a specific form of mental fatigue that attends this brand variable thinking - a form of thought that requires defining the nature and scope of problems, mapping potential solution pathways, identifying steps, devising assessment criteria, testing and revising all components - an essentially constant cycle of invention, evaluation, and refinement. Understandably, this cycle is mentally taxing, a cognition sport that’s incredibly demanding, one that’s both a sprint and an endurance trial - when Pete’s in full development mode like this, he reports that he can go at it for three or four hours at a stretch before he’s brain-fried.

It is worth pausing here to emphasize: mental fatigue is real, it is universally experienced, and we ought to factor more prominently and deliberately it into the way we conceive of work. In Part 3 of this Stratified Attention series, we’ll consider some of the workplace implications of folding AI into the way we operate.

We’ve all experienced that moment after a period of sustained, intense focus where our brain goes rubber-legged and abandons all sharpness. AI, obviously, never gets winded, never flags or needs downtime. But since it is we who are the conductors of this cognitive orchestra, we who must drive and shape and direct the flow and outcomes of this collaborative endeavor, we who are the lead car on the thinking train, we must become adept at:

Recognizing when our brains have reached their limit on a given session of decisioning, and to be OK with the need to step aside for a spell.

Harnessing the tirelessness of AI to have it persist in carrying out tasks while we power down - this is where the Observed Cognition cited above becomes indispensable: noting our own finite capacities, tracking the tempo and progression of the cycles of the ongoing activity we’re orchestrating with AI, and sync-ing our fatigue level with an ability to pass assignments to our AI-enabled Non-Napping, Ever-Ready colleague that it can complete whilst we rejuvenate.

Becoming adept at hopping back into work mid-stream, picking up not where WE left off when we needed rest, but flash-assessing what the AI’s been up to in our absence and recalibrating all our markers of assessment, making updates and adjustments on the fly.

Next time, we’ll conclude this series by contemplating what all this might mean for the world of our workplace and the ways we need to adjust our expectations, and the (potentially wide-ranging) impacts on operational realities.

On Stratified Attention, Part 1 of 3

2026-07-31T00:00:00Z

“Learning ‘how to think’ really means learning how to exercise some control over how and what you think. It means being conscious and aware enough to choose what you pay attention to and to choose how you construct meaning from experience. Because if you cannot or will not exercise this kind of choice in adult life, you will be totally hosed.” - David Foster Wallace, This is Water: Some Thoughts, Delivered on a Significant Occasion, about Living a Compassionate Life.

At this dawn of the Age of AI, an age of disruption and upheaval at the hands of unruly new technologies, an age where we struggle to determine what role these already-powerful yet swiftly-evolving technologies have to play in our lives and work, along with what adaptations we as individuals and organizations must undergo in order to unlock the potential of these technologies we find confounding and exhilarating in equal measure. And just as importantly, how we’re to unlock this potential without getting overcome by the technology itself - not simply in the much hand-wrung-about AI-will-take-all-our-jobs sense, but in the unwittingly-remake-our-minds-to-suit-its-purposes-and-methods sense.

We’re prey, I think, to a nagging idea that AI stands apart from preceding technologies in the sense that more is required of us in order for it make actual sense to us and for us to feel it is fully useful and beneficial, the sense that there’s more to this than learning keyboard shortcuts, say, or some other form of tips-n-tricks level of low grade know-how. We have the sense - the correct sense, it feels important to emphasize - that for these technologies to be as transformative for us as a proliferation of companies and pundits and policymakers are claiming they’ll be, we must become somehow more than passive recipients of them, that these are not simply “plug-n-play bystander” technologies - “switch on the lamp, the lamp throws light,” nor are these simply some “new wrinkle” technologies that execute an existing function in some mildly novel way - “Clap on, clap off! The Clapper®!”

We understand that AI is something altogether new, something that (seems, at least) to learn from us and talk to us, and (where permitted) to operate independent from us. What we’re less clear about, probably, is how we must remake or reconceive our means of engaging, how, in other words, we are to be in relation to these technologies.

Over the course of a recent series of conversations, ZeroWidth co-founders Pete Binggeser and John Cain, along with myself, have identified what we believe to be at least one important way in which everybody already knows how to fold these technologies into our thinking and doing and making: our capacity for Stratified Attention. A vital distinction, we think, worth making: the phrase “divided attention” connotes the blitzed, beleaguered condition that comes of having too much to do and not enough time, resulting in a proliferation of Partially Completed-Then-Quickly-Abandoned Tasks. Whereas with Stratified Attention, we alternate between centering things and allow them to recede, to waft upward in our minds to gain our full notice, then retreat to a lower plateau of our attention where lower urgency or some measure of unresolvedness warrants.

A concrete example from my own recently lived experience:

Like many folks who are socially maladapted, I am currently at work on a novel. Not too long ago, like every novelist, probably - but especially those whose minds are of the hectic and/or ricocheting variety, as mine tends to be - I hit a snag. A bad one. A weeks-long plotting quandary I could not crack.

Quick sidebar about process: generally speaking, approach-wise, there’s a fork in the novel-writing road - in one direction, there are the Outliners, the sensible people who create for themselves a map for where their story is headed and where it’s meant to wind up; in the other direction, there are folks we’ll call the Intuitioners, those recklessly hapless goobers who just start clacking away and see where it takes them. As an avowed goober, I am in this latter camp - get cracking on something and instead of puzzling things out in advance and proceeding with the confidence that comes of clearly knowing what your objectives are, I (for reasons I can’t really name) trust that solutions - to plot, to character, to structure - will emerge as they are needed. Just where these solutions are meant to emerge from I cannot say - the primordial depths someplace? My fractional share of the collective unconscious? The “yes, and” spirit of improv?

Like many “creative” people (these scornful-seeming quotation marks are NOT directed at creative people generally but only at myself, since like every artist, I’m prey to an unhealthy skepticism about my own capacities to be even minimally inventive or interesting,) I harbor scads of superstition about the sources of my (meager) output. So despite my dubious claims to having something resembling smarts, I cling - in the absence of any evidence - to a conception ultimately of my own devising as being some fundamental truth that drives and shapes the direction of my creative work. In some other person, I would urge them to abandon - or at a minimum to examine - such a set of beliefs, as this sort of magical thinking can add nothing of practical utility or value to the complex work of generating original material. In myself, though, I have this odd, unnamed dread of even looking too closely at that this frankly quite weird belief that, in order to avoid unspecified (imagined) consequences, I must place my full trust in some primordial stratum of brain activity toiling away below the level of my consciousness that will feed out a tattered ribbon of theme and incident and so forth in some unwitnessed inexplicable way.

And despite the flimsy and probably spurious underpinnings of this non-method, I nonetheless DO place my trust in this substrate of my cognition, and so far, knock wood, this trust has been well(ish) placed. Because I was able to place this recent plotting quandary, this bottleneck that had bedeviled me for weeks, on some lower shelf of my mind-pantry and let it sit there in the darkness, working away on itself like sourdough starter or something. And after an interval of a few days, the solution to the plotting dilemma presented itself - it just emerged. (From? Who can say?)

Gene Fowler famously said: “Writing is easy. All you do is stare at a blank sheet of paper until drops of blood form on your forehead.” And, setting aside the anachronistic paper-not-screen construction of it, this is undeniably true - mentally and emotionally, writing is a costly enterprise. But it is ALSO true that there is a category of writing operation that takes place at a depth below effort. And it was here, in that mental layer where sweat is not a requirement, that the troublemaker of a plot pivot was resolved. Which is not to suggest that there’s some magical intracranial Narnia where woodland sprites tend to our business - there is a form of our attention being devoted to a complexly thorny conundrum, it’s just a form that adheres to other rules than the front of mind one we’re used to. It’s like lower gear in the transmission - same gearbox, just more specialized and less frequently required. A much smaller-scale version of this phenomenon happens in my writing every day all the time - I’ll be searching for a phrase or puzzling over what might be a helpful image - the line or paragraph level of constructing text, not the foundational basis level - and I’ll downshift to checking my email or playing a dopey puzzle game for a minute, while the problem turns over in some less explicit quarter of my mind - not a different room or down-some-flight-of-stairs level of my mind-house, the way the plotting quandary seemed to, but in some handy but less directly observed spot, like a drawer in my mind-desk, say.

Shifting your attention back out of my dinky little writer-troubles and onto your own experience* - every day, all the time, there are multiple simultaneous, often conflicting claims upon your attention: some fleeting, some immediate and pressing, some well down range in Important But Not Urgent territory. Mostly, I think, these many claims on our attention, claims of varying urgency and differing timescales exist in our minds like cream poured in coffee - a swirl that’s differentiated but quickly becomes incorporated. True, when stressed, the addition of a Single Other Damn Thing can set us off, but mostly we just fold the new thing into what’s ongoing. When we’re kneading dough and a bit of it breaks off, we don’t typically run around shrieking in terror and confusion, we just add it back into the whole. The way we rank and shuffle, focus on and disregard the many claims on our attention is fluid, is constantly evaluated and updated - a process where we toggle between reflexive/instinctual and conscious/deliberate all the time.

* Meta-Cognition Sidebar: This, right here - in the current moment you’re presently having as a reader, where you’re capable of tracking the threads of A) our observation about interacting with AI, B) garnished with my anecdote about the Novel Plot Quandary, while seamlessly C) shifting mental gears into your own lived experience, AND D) hopping off all these other topics to consider this brief pertinent-yet-tangential-seeming moment of Meta-Cognition WITHOUT LOSING ANY OF THE REST.

And we at ZeroWidth wonder if it is this capacity for Stratified Attention - among other capacities, surely - an inborn quality in ourselves that we sharpen and expand upon over time, is it this quality in us that might permit AI-enabled tools to become such a singular and radical means of assisting us in our work?

Given the now longstanding tendency for monetizing hype and often unhinged rhetorical overreach to cluster around all topics relating to AI, it feels worth re-emphasizing, as we always will, this ordinal relationship between ourselves and this technology: WE do not remake ourselves to suit IT; IT exists in order to further OUR objectives, NEVER vice versa. In order for this to remain so, though, we must develop greater clarity on what we expect and demand of AI. And a vital element of developing this clarity must involve deeply noticing what use of these tools does to our thinking, how it shapes and redirects our “natural” means of navigating problems and processes, what existing aspects of ourselves we may draw upon and what new skills or mindsets we must adopt.

Because it’s nearly always true that there are methodologies and approaches we must master in order to put any newly acquired tools to good use. This is as true of a belt sander as it is a forklift, a screwdriver or set of tweezers - the vital difference between between analog tools and AI-enabled ones is that instead of demanding physical skills of dexterity and precision, procedural steps and safety measures, the capabilities of AI are more in the realm of time spent, attention devoted, decisions that need making. So it follows that that the adjustments that we must make in relation to AI-enabled tools have less to do with technique in any physical, procedural sense, and much more to do with cultivating a clearer and more comprehensive mental map of what we’re seeking to achieve, and develop a greater command of where and how AI tools might provide specific help.

One of the drums that ZeroWidth consistently bangs on is the idea that AI is an INGREDIENT TECHNOLOGY - it provides at most a COGNITIVE ASSIST for human actors, it does not perform meaningful work on its own, does not think independently.

Next week, in Part 2, we’ll examine Stratified Attention in practice and contemplate the changing nature of work.

The To-Do List of Sisyphus

2026-04-22T00:00:00Z

NB: Typically, I create these posts as the (omniscient, trusted, wry-but-sexy) voice of ZeroWidth, a kind of collectively wrangled, collaboratively generated result of much conversation and sometimes contention between myself and ZW co-founders John Cain and Pete Binggeser to arrive at a shared sense of the positions we wish to adopt/express for this newsletter. For THIS edition, however, I am writing as myself, as Ian, an (spotty-knowledged, unreliable narrator-ish) individual, since I’ll be drawing upon my own lived experience to make some of the points we wish to hit on.

Inciting Incident:

So this AI company spent a mess of money to hire Matthew Broderick to do a down-tempo version of his decades-ago Ferris Bueller character, spectacled and in a selected-by-committee-seeming unbuttoned denim shirt angled at Appearing Maximally Unthreatening. The company, Genspark, really seems to have taken a lean-into-it approach where Fears That AI Will Replace Us are concerned - rather than the more intuitive-seeming Adopt Reassuring Tone about the long timeline for such an outcome or the incremental nature of such things, they go full I’m-Rubber-You’re-Glue on the issue and seem to adopt a stance of “But really? Secretly? You WANT to get replaced, though.”

Resulting Reflections*:

On the one hand, Genspark’s counter-intuitive-seeming messaging approach would appear to cede conceptual territory to the Techno-Alarmists among us who’re convinced we’re like a month away from our machine enslavement that’ll be like a way less cool-looking Blade Runner dystopia where instead of neon and cinematic rain, everything’s just a drab, joy-suck strip mall, or it’s some dollar-store Matrix with zero martial arts but all of the ponderous philoso-jabber. On the other hand, though, they pick at one of the scabs covering one of the largely unspoken elements of the “the machines will replace us” narrative - namely that lots of the tasks that fill our workdays are onerous and boring and stultifying in their repetition, and every dang one of us would be thrilled to dang death to be rid of them. All the eye-roll parts of our jobs, all the meetings-that-coulda-been-an-emails and items we keep avoidance-pushing downward on our to-do lists, the things that really makes us feel like we’re by-god earning every dang nickel of that paycheck - surely these are the aspects of working life that AI was sent here to dispense with for us, no?

Personal History:

I was a disastrously poor student. If I had a nickel for every time I received the “so much potential” speech, I’d have like thirty bucks, easy. My trouble was not lack of intellectual chops or deficient curiosity, it was that I was born three quarters of the way to bored and if American schooling was a nation, its main export would be boredom. So in spite of some native smarts, I was a refusal machine calibrated to blockade my own academic success - my brain and temperament conspired to keep my participation minimal and my commentary venomous. So if I am nothing else, the cocktail** of my neuro-situation and my foul-ish mood state have always combined to make me a Canary in the Novelty-Seeking Coal Mine.

If you think I’m exaggerating, the quote in my senior yearbook*** was: Blow it off. On the face of it, this might appear to indicate a person averse to all effort of any kind. Which was never the case.**** What this (timeless, lofty) sentiment was directed at was not activity, per se, but activity that smacked of the purposeless, activity that felt like it did not matter in any substantive sense. So what all my refusal and resistance and insolence was reserved for was the Like most of us, I’m always grateful to feel fully, meaningfully engaged in my work, and where this is the case, I can and do put forth prodigious effort. But where I see my labor to be of the wheel-spinning variety, where it consists of repetitive look-busy pointlessness, where we’re promised a pair of curled and fading Post-It notes at the entrance of the Cubicle of Sisyphus, one reading “Push rock,” the other “Get liver pecked out,” and we just toggle back and forth between these till we’re made obsolete by robots, it is this level of activity at which the primal Blow it off in each of us must be directed.

So who knows, maybe the brain trust over in Genspark marketing has unwittingly unearthed a vital vein that runs through the AI Will Replace Us All narrative, a vein that inverts the underlying suppositions it rests on, namely that we must dread the approach of AI since it’s allegedly here to supplant and surpass all our capacities, to sub in for us in all arenas of (historically) human activity. And while, yes, the prospect of AI poetry displacing our handwrought verse, of AI auteurs crafting films that crowd out our own, ought to send a chill down our spines, the prospect of AI stepping in to fill the cells of our spreadsheets and schedule our meetings ought, it seems to me, to cause us to rise to our damn feet and greet it as our liberator. It’s this vein of the story that Genspark digs up - this I’m-not-lazy-but-ONE-more-bit-of-makework-nonsense-will-sap-my-will-to-live vein that exists in each of us to varying degrees.

Concloosh

Which is why these zeitgeist-y stories we’re fed on virtually any topic warrant lots more scrutiny than we might be inclined to give them. And listen, I get it - the AI Will Replace Us story is doused in exactly the type of outrage-kerosene that makes such stories flammable inside my mind. The vestiges of punk rock me, urging us all to be Blow it off, whatever it may turn out to be, coupled with the aging me, fed up with encroachments upon my life and mind and mood, seizures from my pocketbook and prospects and planet from all directions that have me Mad as Hell and Not Going to Take it Anymore, conspire to keep me riled up when such a story pierces the bubble of my fractured attention. But as is almost always the case, the kneejerk state of riled-upness is seldom the optimal one for assessing much of anything, and responses arising from such a state are likely to be pretty nuance-free.

So when you come at me with the AI Will Replace Us story, rather than letting my lizard brain take the wheel and cause me to be all umbrage-taking and doing a mess of pre-fight neck rolls to duke it out with imaginary robots (who like the Iron Giant would just crush me without breaking stride,) wouldn’t it make better sense to greet this story critically and with curiosity. Instead of running in circles Kermit-yelling my anxieties, wouldn’t it make better sense to ask follow-up questions? Like “How do you mean that?” and “Can it just swipe the stuff I’m fed up with already? Cuz I have a list handy.” Because it doesn’t take an Aging Ferris Bueller to recognize the utility and practicality of the well-placed Blow it off.

*I recognize, given the source and subject, that “Reflections” is a bit elevated for what’s going on here. But my weakness for alliteration is well-established, and “Stray Thoughts of Poor Quality” impedes the flow of the line, and runs the risk that you’d just skim over it entirely.

** More of a slurry, really.

*** The Goldbug, Amherst Regional, whose name is a mystery, since our team mascot was the Hurricanes. Which also made no sense, as we were a hundred miles from the Atlantic.

****In my 20s, I was co-founder of a storefront theatre company in Chicago. In all of human experience, there’s no endeavor more effortful than this. But hey, silver lining: it comes with an ironclad guarantee of zero material gain.

Semantic Wayfinding

2026-03-20T00:00:00Z

As noted elsewhere - in other editions of this newsletter, other publications covering the field of AI, etc. - a common pitfall of Large Language Model (LLM) use is the fallacy that it’s a place to offload the need for human thinking - that we can just set a platform - ChatGPT, Claude, etc. - a task and it’ll not only perform that task, but it will carry out all the cognition that necessarily underlies/precedes the task itself. This does not yield results that are fully wrong every time, but it does lead to vagueness, approximation, inaccuracy, etc. I can pull up Claude, give it a topic and word count, feed it my own previous writing in the general subject area, and request it to produce an edition of this newsletter while I go visit a day spa or hit the M&M Store again or whatever.

The resulting text would almost certainly be a lackluster one - not because mine is such a singular genius or because what I might have to say is rare and wondrous - but because the motive that underlies such an approach is not to seek the most penetrating insight or greatest novelty of expression, but the avoidance of effort. This, we at ZeroWidth have come to conclude, is perhaps the greatest peril implicit in LLM use (our Mindset Myth series of posts explore this topic in greater depth) - this fallacy that it’s a substitute for thinking, that if we plug in some parameters, the platform will make new ideas for us. This will not take place, no matter how conscientious our inputs to any such system, no matter how fervent our wish for it to be so - the resulting text will necessarily be a guesswork pastiche, an approximation of the ideas we’re seeking to express. If we think of cognition as a deck of cards, any LLM, no matter how fast or sophisticated, will demonstrate a near-infinite ability to shuffle that deck, to rearrange its constituent parts in seemingly limitless ways; the LLM cannot generate anything, though - it isn’t inventing new suits or face cards, is not adding to body of card knowledge.

Whether we’re writing with a quill on parchment or via a virtual technological platform, it is the brain activity of a person that renders a text clear or obscure, persuasive or not, lively or dull. We all have access to the same set of words, so it is isn’t these that make writing vital; we all have societal and cultural frames in and from which we operate, so it isn’t these cause writing to be singular - it is the sensitivities and esthetic judgments, the expressive audacity and acuity, etc. of the individual creator that sets one piece of writing apart from the rest. So there is literally NO SUBSTITUTE FOR THE LIVELY AND RESTLESS HUMAN MIND, no product that results from any attempt to match the generative capacities of people can succeed.

So, if the LLM is NOT a form of alternate or substitute cognition, of original generative thought, then what is it for? What utility or value does it provide? Well, at ZeroWidth, the way we’ve come to understand the LLM as a tool for refining, shaping, and strengthening human thought, not as a surrogate or replacement for it. So all the advisory and consulting work we do, all the tools we design and build adopt this concept as central, as indispensable in the adoption and deployment of any LLM solution.

At the recent Berkeley Master of Design (Berkeley MDes) Forum on Emerging Technologies conference, ZeroWidth co-founder John Cain delivered a talk that provided what we hope you’ll find is a useful framework for navigating this complex topic. The subject of the conference was “Interface to Agency” (lineup and overview HERE) and John’s talk centered on a process by which we can formalize/codify the process for design ideation that INCORPORATES but does not rely exclusively upon AI. A process that hinges upon a foundational understanding that it is the brain activity of people that is and must be the animating energy underlying these forms of complex solution-seeking, that any process that’s wholly AI-reliant will effectively be a series of stabs in the dark nested inside a mess of impressive technology.

And while it’s obviously possible to make the argument that human-driven design iteration is also a series of stabs in the dark, the reality (or the hope, anyway) is that the experience and judgment, the discernment and discipline of the trained human mind (whether individual or in collaboration with others) will result in stabs that grow less dark over time, less wrong with each adjustment. The productivity in such a process of iteration dwells in a human capacity to measure, assess, amend, and reconceive, whereas the LLM can merely rearrange. So even though the speed at which it can do so is impressive, we must keep sight of the reality that extracting the value from the output from the LLM hinges upon a human ability to evaluate, rank, reconfigure, and most crucially reconceive.

In his talk, John likens the LLM to advances made in the Renaissance era in our ability to render images using linear perspective - techniques that did not mean we were suddenly able to see with greater depth or precision, but that the tools of representation at our disposal were expanded in a transformative way. And subsequently that gains we made in the act of representation caused an expanded cognitive palette for our forms of visual expression. The tools and techniques of better draftsmanship - of creating representations that were more faithful to the directly observed world we around us - were not in themselves the basis for the resulting artistic revolution but rather the human minds that developed an ability to incorporate these into the way they conceived of things.

In the same way that a map is not a territory, or a painting is not a bowl of fruit, a plan is not time travel - it is a representation of an imagined future.

A similar principle of representative power holds where LLM use is concerned. In addition to speed, one of the virtues of the LLM is scale, meaning it’s better able than we are to hold conceptual sprawl. No matter how nimble or capable, the human brain has a finite set of contingencies, consequences, causations, etc. that it can keep sight of at any one time. The LLM, by contrast, is able to hold as many as its data set might contain. You or I might be able to commit to memory and recite a number of sonnets or soliloquies, whereas the LLM can retain the whole of Shakespeare ready to refer to. The LLM excels at drawing upon information at a scale beyond our abilities. The role of the LLM, therefore, is not to think on our behalf, it is to assist us in doing so by helping us navigate the many branching avenues of potentiality while still in the conceptual stage, there’s no reason to fully execute functioning real-world prototypes of every prospective path.

To utilize these powerful new tools to good effect requires firstly that we treat them as such - tools, not colleagues, not our replacements, not some auxiliary brain. Further, if we’re to unlock the full potential of these tools, it will be in tandem with our own minds and capacities - let the LLM do what it’s good at, namely summoning numerous pieces of information in any combination, while we retain the decision-making about what matters.

A fog descending upon us, upon our minds, is a common trope for characterizing the difficulties inherent in clear thinking. Typically, we conceive of the fog as something we must cut our way through in order to arrive at some sensible destination. But what if this is a less interesting way to think about the problem? What if we reconceive of the fog - as not being an impediment to progress, but as a necessary aspect of it? What if we regard the fog of emerging-but-not-yet-resolved thought not as a hindrance to getting where we wish to go, but as an integral and even indispensable part of the territory we’re exploring? What if, in other words, the fog IS the territory?

In order to extract meaning from the swirl and opacity of the fog, we must be able to map our pathways through it - even (or perhaps especially) if these are branching and indirect; to summon what’s apt and useful in among the seeming obscurity, we must capture the insights, questions, and contingencies generated along the way. For the past year, ZeroWidth has been at work to devise a flexible, adaptable, collaborative, customizable suite of tools for precisely this kind of guided cognition - a means of both generating and representing the pathways through and connections between thoughts, a means of giving shape to inspiration, of systematizing innovation.

Before the full launch of it, we invite you to play around with it at no cost to you. This link permits you to learn about and noodle around in and refine a process of your own creation. Give it a whirl and let us know what you think.

~~https://semfield.ai/~~

~~This free access is temporary for the purposes of beta-testing, so hop in now.~~

As of August 1st, 2026 - semfield is fully integrated and expanded within the ZeroWidth platform. Make a free account to try it today.

Consulting the Generalist

2026-02-03T00:00:00Z

Query: When faced with the dizzying pace of AI’s advance - and the mounting anxiety that it will supplant us all one day soon that accompanies it - how is it possible not to be freaking out about it more or less around the clock? When the evidence seems to suggest that Large Language Model (LLM) AI is as knowledgeable and quick-thinking and capable as any one of us, how can we refrain from unplugging ourselves and walking off into the mossy hush of a forest never to return? How, after having accrued years or decades of expertise that now seems extraneous, can we summon the wherewithal to carry on? How can we refrain from collapsing into a gelatinous sweatpants-ed heap of self-pity monster fit only for stress eating and cyclical napping?

The answer: Quit Gist-ing.

You of course know the noun form of the word “gist” - the main point or aspect of an idea. The gist - importantly - is the overall thrust of something, not its totality, the core of it, maybe, but lacking detail. For our purposes, the fresh-minted verb form to gist means to base one’s decisions/responses to/comprehension of a situation on an approximate estimation of its meaning/impacts, a generalized sense of its content/form/shape. In other words, succumbing to precisely the kind of flawed and rudimentary “thinking” that AI systems are perpetrating all around us. It’s worth pausing here with a reminder that LLMs, for all their impressiveness, are NOT generative, they are predictive - an important distinction, where cognition is concerned.

Where you or I, where we’re being thorough, will seek to arrive at the best, the most comprehensive, the fullest and sharpest version of an idea, the LLM will autofill to adequate, meaning its “goal” (a human term/framing we ascribe to a non-volitional technology) is not right, or even least wrong, but something closer to less wrong enough. The Sieve of Credulity through which LLM-generated content must pass is necessarily devised by us - the thresholds of detail, of veracity, of completeness, etc., in other words, the ingredients of expertise - so the quality and character of what the LLM produces is ultimately determined by the relative finesse of the parameters we set for it.

Which is why the most effective means of managing your own AI anxieties is to consult LLM results in some arena of your own expertise, some knowledge area where you’re adept at spotting any tendency to gist, a richly detailed, measurably accurate part of your Mind Map, not the scary less known part labeled “here be dragons.”

If, as for many of us lately, your faith in your own value is shaken, try the following: request output from your LLM of choice concerning a topic you know a great deal about, not a topic of simple interest to you, but one that you know deeply. For instance, I have a keen interest in mid-century crime fiction, I have an abiding interest in cinema, I have interests of varying depth and intensity in history and psychology and nature and British comedy and a dozen other subjects, but none of these have become sufficiently elevated in the Rankings of My Attention that I’m prepared to devote myself to them.

By contrast, I (mis)spent over two decades of my life as a performer - I was an actor, comedian, and solo performer. Over that time, I appeared in and produced scores of shows, hundreds, maybe. So it’s fair to say that now, as a result of that devotion, I probably watch movies and shows in a different way than you do - that the calibration of my assessing apparatus is dialed into a frequency you may not pick up, that the eye for detail I’ve acquired as the result of tens of thousands of hours’ worth of participatory and contemplative engagement with these disciplines has rendered me something beyond a passive consumer of them. The sweat equity poured into these activities has imbued me with a reflexive capacity to take in a hundred micro cues and other data points to form an instantaneous-yet-robust appraisal. These nano-impressions coalesce into a global evaluation of quality that, though quick, is not knee-jerk. My experiential capital, sunk into the pursuit of a discipline, amounts to a kind of lasik surgery, an ability to see greater detail with more clarity than the casual fan or enthusiast. This electron-microscopic view is compounded by the fact that I’ve been a writer for longer now than I was a performer, so I’m able to parse the textual and performative to an absurd (often as not pretty joy-killing) degree - if a joke doesn’t land, I can tell you exactly why and identify the parties responsible; if an actor’s portrayal is lackluster, I can tell you where that flatness originates and assign blame.

Similarly, my eldest is obsessive about all things automotive, so can identify at a glance every car by make, model, year, trim level, aftermarket mods, etc. - a dizzying, confounding level of detail entirely lost on me, but for him, a level that his natural affinity and sustained enthusiasm have elevated to the plain as day. Or the way my grandpa, a lifelong birder, would cock his head and pause when birdsong reached us, and could declare with the confidence arising from decades of close observation “Cedar Waxwing,” or “Eastern Bluebird,” or whatever. You and just about everybody you know holds comparable forms of expertise - some version of you-know-so-much-it’s-hard-to-even-notice-how-much-you-know. We all have our subject areas where we could begin every sentence with “Well, actually…”

When watching a stand-up comedy special (or a play or a movie or a show,) I have no expectation that every person I might watch with will possess the same ruthless level of evaluative apparatus that I do. Over time, I’ve attempted, sometimes successfully but usually not, to remember this and to remain respectful and tolerant of their (incorrect) enjoyment of things I know to be lacking. I know that opinions, no matter how hard-won the expertise informing them might be, are the result of preferences as much as anything else, so I try (and fail) not to go berserk when a show or movie gallops through the zeitgeist and gobbles up (in my view) undeserved praise.That’s my curse. Your own expertise is the curse under which you must toil.

Back to LLMs - the best, most effective way to dial down our freak-out about them is to futz around with them in one of your areas of expertise, one of your Well, actuallys. You’ll find right away that the results you get are riddled with approximations, wrongheaded assertions, naive assumptions, etc. The LLM has as much information as we can stuff into it, yes, but it lacks your storehouse of experiential knowhow and the judgment that comes of deep interest, lots of trial and error, the accrual of tips-n-tricks we gather over time, the pitfalls and missteps to avoid, etc.

Boom. Your anxiety is addressed. The number of times you go “that’s not quite right,” or “it doesn’t work that way” are likely to be numerous enough that you can see AI essentially gist-ing like crazy in ways that somebody without your deep knowledge of your subject would not be as likely to catch. In your arena of expertise, AI is a wading pool - a wide one, surely, drawing as it does on a vastness of data, one that might be said to stretch over the horizon - but diving headlong into it would get you a broken neck; your knowledge base is a lake - its waters so deep you can’t see the bottom, currents of philosophical insight and specialized literacy, swirls of memory and wisdom and object lessons and time devoted. Seeing this distinction - between your lake and AI’s acres-wide kiddie pool, permits you to see AI as it actually is: generally impressive, but still requiring considerable human intervention and guidance in order to be a reliable source of knowledge or advice.

For instance, I would trust AI to provide an overview of canine vascular surgery - its principles, procedural description, history, etc. - but would never rely upon it to walk me through performing surgery on my dog. Nor would I regard it as a viable substitute for an experienced veterinary surgeon. If Dog Surgery is a subject at a (macabre-seeming) pub trivia night, AI might prove a solid teammate - whereas entrusting the health of an actual animal to it would be misguided and reckless. The same holds true for a thousand different forms of specialized knowledge - information alone is no substitute for what is deeply and fully known, and mistaking it as such can only arise from the gist-ing that comes as the result of having generalized impressions of a field or subject we ourselves do not know well. This is what the LLM is - a verbose and voluble companion that has an acquaintance with any topic we can name, but that does not have the relationship with that subject that leads to real knowhow or the recognition of real meaning.

LLM Mindset Myth # 6 - AI Can Talk to You

2025-09-04T00:00:00Z

There is perhaps no technology so prone to sparking beliefs about itself than Artificial Intelligence. If the beliefs springing up in its wake were consistent, if humans could concur on the nature and implications of AI, we would have little to say about it, and would certainly never have launched this newsletter. But the beliefs that cluster around AI are as varied, as contradictory, as unreconciled as it’s possible to be while still ostensibly talking about the same thing.

The beliefs we hold about AI are a strangling bramble patch of dense, often thorny stems, some taking root in anxieties and aspirations already present in us, some in credible information about the technology and its capabilities, some in our misconstruing of the realities surrounding it. The first category, those beliefs shaped by existing worldviews, is not the arena where any of us can expect to move the needle much - if your sense is that this technology will usher in an unprecedented age of human misery and subjugation, there is nothing much anybody can say that will dislodge you from this view; if you conclude that this technology represents the dawn of an unheard-of new flowering of human potential, one where, liberated from busywork that commands our attention and fills our days, our creativity will be unleashed and we will know only contentment and purpose, there is a low likelihood that any of us could dissuade you from holding fast to this conviction.

At either extreme, since these visions of our AI future are rooted as much in individual predilections as anything - rooted, in other words, in a person’s past - there can be little hope of extracting anybody at these extremes from what amounts to the dogma underlying these visions. So these folks, whether at the rose-colored extreme, or the storm clouds-gathering one, are doomed to be disappointed when their -topia of whichever flavor does not pan out.

Which leaves the rest of us - those whose conclusions are less foregone. If we remain open to incoming information and analysis, then our opinions will obviously evolve as the uses and misuses of the technology continue to unfold. Where our beliefs arise from fundamental misunderstandings about the way the technology operates, then if confronted with definitions that are more current and precise, we recognize that we’ve been mistaken and adjust our views accordingly.

One of the most commonly held misconceptions about Large Language Model (LLM) AI is that it can talk to you. It cannot. Any platform you can name, with any plug-in, or any set of training data, or any other means of altering or augmenting the LLM in question, it can NOT talk to you. Period.

It may APPEAR to. It may mimic aspects of doing so very convincingly. But it is not engaging with you in a conversational exchange the way a person can.

We at ZeroWidth cannot avoid concluding, however, that this fallacy is both harmless and probably inevitable to some degree. Let us explain. First, let’s consider some of the features common to human conversation.

Regardless of topic or duration, what are some of the elements shared by just about any conversation between people? Any human conversation is necessarily an exchange that is all or most of the following:

It is reciprocal. This notion must be understood to be more than simple turn-taking, which the LLM can undeniably do. For a back-and-forth to be fully reciprocal, there is a dynamic whereby each participant seeks to contribute as much as to gain, where there exists an impulse to share, commiserate, understand, etc.

It is improvisational. In an improv scene, each participant must surrender some measure of their “sovereignty,” meaning that each subordinates their own individual preferences and objectives somewhat to the collective act of creation taking place. There is an ongoing, constantly updating process of assessing and adjusting, of interpreting and aligning or opposing, etc.The same principles apply to conversation.

It is cumulative. An exchange between people results in more - more opinion expressed, more information imparted, more insight gained, etc.

It is generative. In a movie scene, the actors deliver lines of dialogue provided to them in advance. In a conversation, we devise and compose as we go, on every scale - from the broad shape of our ideas to each word choice - it is not predetermined.

It is dynamic. A conversation follows its own path, a path communally set by its participants, adjusting in real time to new information, fresh revelations, etc.

It is subtextual and tonal. Much of what gets passed between us goes unsaid. And in any conversation, there is a layer, perhaps multiple layers, of communication that go unspoken - shadings of intent, conflicts left unaddressed, factors like irony or sarcasm, etc.

The act of talking to each other, therefore, whether simply shooting the breeze/exchanging pleasantries or exploring deep philosophical questions, is never an uncomplicated endeavor. Even if the substance of a conversation is not deep or consequential, the characteristics common to just about every such exchange are numerous and complex. If we contend with the real density and intricacy of human communication, it becomes unrealistic to conclude that any technology is carrying out anything like it.

Given this list of the characteristics of human conversation - a list that’s only partial - it would not be reasonable to conclude that that a technology still in its relative infancy could be expected to master, much less replicate, such a complex and variable form of communication.

All that having been said, however - it is also true that the only folks consistently needing to have a deep, accurate understanding of what’s going on under the hood of LLMs are the engineers who structure and deploy it. Within ZeroWidth, for instance, our own Pete Binggeser could give you a clear and comprehensive rundown of the technical processes underlying the created appearance of LLMs conducting conversations with us.

For the rest of us, it isn’t simply comprehensible that we talk and think in terms of AI talking to us, it is a natural and fitting response to the emergence of a technology that so closely approximates the kinds of communication intrinsic to us. There is perhaps even a necessary coping dimension to applying such innately human terminology to processes taking place within complex technological systems. Such language represents a cognitive leap we make in order to incorporate this new aspect of reality into our understanding of things. In big and small ways, the arrival of any new technology alters our experience of the world; as a consequence, our understanding shifts and adapts to accommodate this new experience. Lastly, new (as well as newly repurposed or expanded) language will naturally arise from our altered experience and understanding,rather than being imposed upon them.

So, while it is unnecessary for us to be dropping lots of “Well, actually”s on our in-laws or colleagues or neighbors who, yes, technically misspeak when referring to ChatGPT or whichever platform “talking to them.” We can bore and probably confuse them with a mess of jargon about tokens and predictive text and training data, but as we all know being correct and being right are not the same thing. It is also helpful, we think, to regard such characterizations largely as placeholder language as we make our collective way through these dawning days of this technology. Because really, there’s no advantage or point to being pedantic about a still-emerging field prey to unresolved conceptions and little in the way of agreed-upon language.

What is most helpful, probably, in shaping our understanding of these new models of thinking and modes of working, along with the beliefs we hold and the language we use to describe them, is to remember that at present we are in a provisional place, an ambiguous place, on a trail yet blazed, as it were, that no matter how much we might wish to have all this done and decided, it is anything but. We would all do well to recall how new LLMs are, how recent a development AI is, and to adjust our sense of how finalized we take it to be, up to and including the way we talk about it.

The Primacy of People - AI’s Virtues and Flaws Belong to Us

2025-06-19T00:00:00Z

The prime asset, the actual brains behind, the true decision maker, the only free will and agency underlying ANY AI system or platform is a HUMAN ACTOR. Whether individual or in a group, it is ONLY humans that can shape, propel, focus, enliven, smarten up, etc. any application of AI. And conversely is ONLY the individual or group of humans that can render an AI application or system or platform exploitative, deceptive, larcenous, evasive, duplicitous, etc. ANY of the excesses of which AI may be accused, whether in the vein of “too smart/capable for its own good” or “self-aware/sentient” OR in the “it’s a big fat lying liar who’s trying to steal from us” the ONLY means by which it warrants any such accusation is DIRECTLY attributable to the influence of the humans designing, developing, and deploying it.

It is WE who make AI helpful and useful; it is WE who make AI destructive and distorted.

Period, full stop.

So all the bot-blaming that folks are doing in response to the headline-grabbing, clickbatory (in the present age, a sadly necessary portmanteau of “clickbait” and “masturbatory,” or, perhaps somewhat less crudely clickbasistic, which naturally is a combo of “clickbait” and “solipsistic”) items about AI. Items that are not aimed at any kind of legitimate analysis or examination of the properties of AI, but ultimately are intended simply to bend your attention in the direction of the company, tech guru, platform, etc. that’s the subject of the article/post/video clip, etc. - invariably an entity/person/product with a (usually unstated) economic stake in instilling some particular form of anxiety in you as you read/view it. If you took a sampling of any dozen such pieces and strip away the details, you’re likely to find a structure something like this:

Alarming revelation about something AI has “done on its own,” or that represents some “AI-generated alteration of its intended purpose or authorized arena of activity.”

Interviewer lobs leading questions to the subject of the piece, permitting them to sketch out some Nightmarish Dawn of Our Overthrow by the Now-Sentient Bots scenario. A scenario which relies upon your lack of technical knowledge about the way these systems work, usually through a combination of dense jargon and volitional-sounding language when referring to the “errant” platform/tool, like “it wanted,” or “it decided.”

Now that you as reader/viewer are sufficiently rattled, the subject of the interview may step in as savior - since you’ve just witnessed their command of the jargon, it follows that they and they alone are equal to the now-urgent task of lopping off the proliferation of hydra heads on this (apparently) marauding beast you weren’t (made to be) worried about till you came upon this piece. Or do you not even CARE about the extinction-level sky-is-falling event that’s been described for you by a concerned citizen who just HAPPENS to the CEO of a startup in the emergent meteorite defense sector?

We’ll call this form of story The Huckster Stepping in to Save You From a Nonexistent Threat.

This kind of phenomenon is obviously nothing new. In any newly forming marketplace, especially one that carries with it credible causes for concern, there will be those who are keen to latch onto the anxiety generated by the new category of product or technology, whose focus will be upon codifying and intensifying this anxiety so as to monetize its exploitation.

All of which is a province of exclusively human activity - the product or technology at the supposed heart of the freakout is a kind of combination of neutral bystander and scapegoat. A famously gut-wrenching example of this phenomenon is how, in 1903, Thomas Edison publicly electrocuted Topsy the elephant as a demonstration of the supposed hazards of alternating current (AC,) which the inventor/ruthless capitalist asserted were greater than the direct current (DC) that he championed. If possible, the fact that Edison and his confederates (who included circus impresario P.T. Barnum) did not actually electrocute Topsy to death, they jolted and cooked her corpse, which had been publicly executed by poisoning before a crowd of 1,500 on Coney Island. Topsy, abused all through her life in captivity, had earned a reputation for aggression (what today we might be inclined to call “self defense,”) so was put down.

The obverse form of the kind of Huckster story outlined above is perhaps equally prevalent. That version goes something like this:

Mindblowing description of some miraculous-seeming, too-good-to-be-true faculty or capability AI has recently demonstrated that will LITERALLY TRANSFORM EVERY ASPECT OF THE WAY WE LIVE AND WORK FOR THE BETTER.

Interviewer lobs leading softball questions to the subject of the piece, permitting them to sketch out the dawn of some Dewy New Utopia, an Unprecedented Golden Age of Peace and Plenty made possible, as luck would have it, by the platform or product they stopped by to promote. This tale is also told using a mess of jargon, peppered with lots of friendly-sounding helpful phrases like “it’s trying to understand you,” “assist you,” and “support you” however it can. A scenario that also relies upon your lack of technical knowledge about how these systems work - the shift in rhetoric being from an emphasis on “volition” or “agency,” to one of “understanding and assisting.”

Now that you as reader/viewer are sufficiently lulled, the subject of the interview may step in as a sort of avuncular concierge whose aim is your eternal contentment and freedom from worry and prosperity. And as you can only conclude, given their command of jargon and soothing tone, it is they and they alone who are capable of flapping their cherub wings to pluck you up and set you down in this paradise (about which you only just came to learn about over the course of this article/video.) Or do you not even WANT the happiness and untold riches offered up to you by a good neighbor who just HAPPENS to be the CEO of a recently launched LLC in the digital wellness space?!?

We might call this narrative The Fairy Godmanager Waving Their Wand to Grant Wishes You Didn’t Know You Had.

At either extreme of the stories springing up to form the Foundational Lore of this technology, and all the narrative gradients between, it is PEOPLE devising and shaping these stories. Furthermore, it isn’t simply that humans are deciding the course of the stories getting told about this technology, people are responsible for ALL the decisions that determine the shape, flow, operation, scope, and nature of the technology.

Make no mistake: it IS TRUE that LLMs and other AI systems are unlike any previous tool or technology - the degree to which they are adaptive is without precedent (the “machine learning” we hear so much about,) and staggering speed at which they adapt is like nothing that’s come before it. This automated ability to self-refine over time renders this technology distinct from any preceding it, sets it apart from any tool we’ve used THROUGHOUT ALL OF OUR HISTORY. But AI remains precisely that - a tool, which means that like any tool it is subject to the manipulations and underlying intentions of the people building and wielding it.

While discussing the essence of AI during a ZeroWidth brainstorm, John Cain hit upon (or maybe he was quoting something - I can’t now recall) the notion that AI is the widest mirror in human history. Meaning that, like a mirror, AI is flattened, it is a representation, and it is a reflection.

AI is flattened. As discussed in previous editions of this newsletter, AI in its present form cannot approach the depth and complexity of human thought. Where you or I are able to make associative leaps, combine reason and intuition to arrive at conclusions, etc., AI does none of this. Despite the vastness of the data set upon which it may draw, it does not know or understand anything in the way we do.

AI is a representation. If you paint somebody’s portrait or take a photo, no matter how faithful and accurate that likeness is, it remains a likeness - it is lifelike, it is not life. A similar principle pertains to AI - it approximates or depicts the patterns and sense of human communication without generating them the way we do. For emphasis, a sculpture, having three dimensions, remains a representation; a film or video, having motion, remains a representation - any means of image-making or -capture, whatever its characteristics, remains a representation. Likewise AI, for all its seeming ingenuity, remains a representation.

AI is a reflection. The volition or agency that gets mis-ascribed to AI with seeming greater frequency is merely a reflection of the autonomy of its creators and users - what appears to be AI’s “self-determination” is neither: it has no “self” and it is we who determine what it is and becomes.

We would all do well to remember that the stories we tell about AI - whether of the Huckster variety, the Fairy Godmanager, or the stories we tell ourselves - they are exactly that, stories. And as we know, the truth of a thing is almost never as neat or symmetrical as a story. If we are to attain any kind of clarity in our thinking about AI, we must keep sight of the ways such stories alter our conceptions and beliefs. As often as not, a helpful filter to apply to such stories is to ask: “Who or what stands to gain if I permit my thinking to align with this story?”

LLM Mindset Myth #5: Implementing AI is PURELY TECHNICAL

2025-06-02T00:00:00Z

The cordyceps is a type of fungal organism, a weird relative to the mushroom. Like all fungal organisms, the cordyceps releases spores to reproduce. Its adaptive strategy is unusual, however - the spores, ingested by an ant, will grow inside the ant and eventually take over the body of the insect, effectively “driving” the ant as a means of locomotion it would not otherwise have. The cordyceps will steer the now-zombified ant to a high tree branch where it will release its spores, the altitude maximizing the area over which to scatter them, increasing the likelihood of ingestion by another ant, perpetuating the cycle.

This life cycle is the inspiration for the creatures in the popular game/HBO series The Last of Us, where much larger and more complex cordyceps variants take over human bodies. For this viewer, it is this element of the creatures that render them especially creepy - this volition/autonomy-robbing capacity - that inspires greater dread than other entrants in zombie lore. The essence of these fungally inhabited creatures, more than other zombies (the rage virus of 28 Days Later, say, or the toxic waste of Return of the Living Dead) all of which involve their subject getting overtaken by some pathogen that causes the self to be erased, then driven by impulses not natural to the once-human host. But with a cordyceps-style creature, there seems to be some implication of a self displaced, or perhaps subsumed by the will of the organism inhabiting us - there is the sense, in other words, of a self still being trapped inside there someplace, as mute and helpless witness as the creature you’ve unwittingly become lumbers around causing mayhem. To my mind, it’s more unnerving to contemplate the prospect of an imprisoned consciousness, robbed of its agency, than a consciousness that, having been replaced, is simply gone.

As with zombies, so it is with technology. Many of our anxieties surrounding technological advances have to do with our sense of becoming reduced to bystander status by them, of our status, worth, and purpose - our fundamental selves, essentially - getting diminished by the expansion of their capabilities. We experience the growing sophistication of technology NOT AS AN ENHANCEMENT BUT AS AN ENCROACHMENT.

With Large Language Model AI (LLM,) which we’re correct to surmise has entered virtually every area of our media and communications, can for many of us inspire a similar kind of zombie-style unease. We may be forgiven for succumbing to the notion that more and more of what we hear and read and see is not authored by humans, but is the product of what we imagine to be a technological cordyceps, an ongoing campaign of capacity-theft, of some unholy hybrid entity that has hijacked our human channels of expression. If, as noted previously throughout this LLM Mindset Myths series, the use of language is an ongoing endeavor of attempting to impart intent and meaning, then in this cordyceps conception, there is a usurping of this of meaning-making, the piloting of intent gets shanghaied by a meddling technological intercessor. Given the present state of AI, however, this is NOT what’s taking place. This sense of stolen volition, of a helm wrested from our hands arises where our instructions, overall objectives, or underlying reasoning are less than thorough or clear. The LLM is not making off with meaning, it is doing its best with what it’s been given.

Which reintroduces the power of tinkering. Rather than issuing a single set vague or contradictory directives to the LLM, better to tweak in drafts, in conjunction with the LLM - map out objectives and approach, rely upon the speed and power of the LLM to provide a series of refining assists along the way. This form of interplay is what the LLM is actually for - iterative improvement to OUR (not its) work product, subject to human oversight and course correction. In other words, AI is not a cordyceps that steers the husk of us around, it is a passenger on the organism of our thoughts. If the intention underlying our communication is the shark, the LLM is the remora hitching a ride on the belly of our ideas - a symbiotically beneficial stowaway that isn’t steering anything. It’s the difference between A) ordering “LLM: execute second quarter marketing strategy” without providing any context or guidance - a sure recipe for useless gibberish; and B) requesting a series of drafts on particular documents, scripts, workflows, processes, timelines, etc. for achieving your broader objective and, in conjunction with the LLM, revising each of these for consistency and clarity - an expedited pathway certain to yield better, more useful results.

The introduction of any powerful technology always involves operational, ethical, design, idiomatic, and cultural considerations in order for it to function as intended, and there will still be unforeseen consequences and ramifications no matter how deliberate and thoughtful its introduction has been. Human actors of different disciplines must remain attentive to AI’s proposed uses, and remain so as it grows more commonplace.

LLM Mindset Myth #4: The BIGGER the AI, the SMARTER IT IS

2025-05-23T00:00:00Z

An ongoing dimension of the ZeroWidth project is for us to try and serve as a weed whacker to the fast-proliferating thicket of hype and falsehood that surrounds AI. One of these kill-your-darlings fallacies is the prevailing fiction (not idiocy, nobody said “idiocy”) that where AI is concerned, bigger is automatically better; that hugeness is somehow a virtue in its own right. The field seems to have fallen into lockstep beneath the Banner of Giganticness For Its Own Sake with regard to AI, a tendency that’s no doubt driven and intensified by the headline-hungry profiteers of the venture capital world, who seem daily to summon the tech media with an arms race of click-baity claims about immensity of scale.

But for you - the individual person or organization attempting to complete your day’s tasks - the vastness of a system’s data scale is at best conceptually impressive, but is not inherently impactful on the work you’re seeking to complete. But while an aircraft carrier is an indisputably impressive piece of hardware, it is not well suited to every task and cannot be wedged into every setting. To say nothing of the turning radius on one of those guys.

Because in order to determine if an AI platform or system is “better,” we need to arrive at some means of assessing how we mean that. Left unexamined, the notion of “smartness” is an obvious good; but for any AI-enabled agent to qualify as “smart,” it must meet some standard of applicableness, of aptness, of pertinence, of precision. And experiential, logistical considerations must figure in, as well: considerations like ease of use, how readily it is adapted to your purposes, how fully it lends itself to collaboration among teams whose members may have different objectives and approaches, different idiomatic vocabularies, differing levels of technical savvy. In short, if your team is trying to use a platform that’s frustrating to use, or that demands some esoteric level of expertise, or that yields inaccurate or unsuitable results, then the mindblowing size of the data set underlying it immediately ceases to matter to you.

We all have an intuitive grasp of the reality that “big” does not equate with “smart” - it can, certainly, but it’s not a given. It’s possible for example that Hulk Hogan has many fine qualities, but I don’t think any of us are inclined to list “prodigious intellect” among them - we’d probably have to push pretty far past “well-muscled” and “dauntingly tanned” and, I don’t know, “growly-voiced” to do so. But nobody’s bursting into the room going “Hulkster! We need you to recalibrate the particle accelerator! Stat!” And if for some reason we did, we’d rightly feel immediate remorse and mounting anxiety about it. And no amount of “You can count on me, brother!” from him could lessen our unease.

It’s the same with Large Language Model (LLM) AI systems - it’s like asking auto-fill to compose the Magna Carta. Even if the drop-down menu is effectively infinite, the menu has no volition, no contextual grasp, no ability to ascribe weight and priority. These functions fall to us - we must imbue such systems with discernment, with sound judgment, we must instruct them in the making of meaning.

Where AI is concerned, “smart” must be understood to mean how specifically applicable its interventions are - so vastness of data scale is no measure of effectiveness. Accurate, specific AI outputs hinge upon the training it receives - so applicable, well-curated data is of greater importance than quantity. AI is actually only as “smart” as the quality of teaching it receives from us. And if you’re saying “Well, but, I don’t know how to do that, though” - fear not. ZeroWidth has built a suite of variable, customizable, intuitive tools that can assist you in navigating this stuff.

LLM Mindset Myth # 3, Part 2 - AI is a REPLACEMENT

2025-05-02T00:00:00Z

Continued from yesterday…

Because as noted previously, “writer” is not simply a vocational identifier for me - it’s a part of how I understand who and what I AM. So instead of hearing “Hey - there’s a robot that can do some of what you do,” and going “Huh. Interesting. I’d be interested to see how.” I get immediately so defensive that what I HEAR is “Hey, loser, this machine is here to DISPLACE and ERADICATE you.” The former is something rational-ish “Welp, my line of work just got a little tougher, I guess,” whereas the latter is existential, is closer to the triggered adolescent insecurities that writhe inside all our brains: “I am futile and dumb and gross and everyone is right to hate me.” So I slam the door of my mind and crank my old punk records and rock back and forth panic-eating a tube of raw cookie dough.

So isn't it possible, maybe even likely, that at least SOME portion of the alarmist rhetoric around LLMs is rooted in this kind of existential panic?

The chorus of dystopian voices breathlessly warning us all that Skynet is seconds away from sentience and will be launching the nukes to eradicate us later today may mean well (or they may, some of them, stand to profit from stoking our fear.) In either case, it’s worth pausing to consider the motivations of any such dystopian voice (“Are they weaponizing my AI fears to sell me something?”) and to “check the math” of their thinking (“Are they overestimating, intentionally or not, the capabilities of AI in terms of what it can do and deliver?”) Pushing any such rhetoric and thinking through the twin sieves of What’s In It For Them? and Are They Just Plain Wrong? Can spare us all a lot of needless Chicken Little-ing.

It’s also instructive and helpful to note where doomsayers are right, which they sometimes are.

Alarmists are right, for instance, in cautioning about the real perils of permitting artificial intelligence into too many areas of our command and control systems, our communications, etc.

But the broader notion that Large Language Models are some kind of invasive species with A) some inborn ability to replicate or improve upon human expressive capacities, which they do not, and B) that this - again, nonexistent - replication capacity will grow more sophisticated at a galloping, exponential rate, and will take over entirely. At the core of such alarmist rhetoric are a pair of great whopping fallacies - A) the Fallacy of Grade Inflation, and B) the Fallacy of Misdiagnosis.

The Fallacy of Grade Inflation holds that if something is newish and seems pretty great, we’ll wrongly imbue with a swelling list of capabilities it does not have; the Fallacy of Misdiagnosis is rooted in our tendency, where our emotions run high, to misunderstand the nature and implications of a given problem.

Succumbing to these fallacies represents a fundamental failure to comprehend what AI is, in terms of its capability, and its purpose, which is to augment and accelerate human activity. LLMs are not generative in any way that we would recognize; simply put, they do not create. Broadly speaking, AI is here to help, not to overtake or displace - it is not competent enough to do so, and that’s not what it was born to do.

LLM Mindset Myth # 3, Part 1: AI is a REPLACEMENT -

2025-05-01T00:00:00Z

There are many voices alleging that Large Language Models (LLMs) and other AI-based systems are moments away from performing much of the work presently carried out by humans. This usurpation, once it happens (any day now, we’re assured by doomsayers,) will be permanent and irreversible - work till now done by us will forever forward be done by generative AI. The idea is that AI has gotten so adept, so capable, so quickly, that its work product is effectively indistinguishable from that of actual people. Which stirs up lots of feelings.

In most cases, the LLM is quicker, surely, than you or I at “composing” a sentence. But that’s not what it’s doing, not in the way you or I do. Where you or I must deliberate a bit, to ensure that the sense of what we’re attempting to say is conveyed in a sentence, LLMs are freed of any such burden; where you or I will furthermore consider the tone and style of the sentence we’re building, LLMs lack the discernment necessary to make such determinations. An LLM is a gifted mimic, certainly, and can often imitate your style or mine pretty convincingly, but it isn’t arriving at sense-making and tone-setting decisions the way you or I do.

Which isn’t to say that LLMs will not one day - perhaps one day soon, given the pace of their development - be capable of convincingly mimicking these meaning-making functions. Therein lies a crucial distinction, however: the approximation carried out by LLMs of devising text, of constructing sense and style will be exactly that, an approximation.

My name is Ian. I’m a human person. Like all of us, I have certain characteristics that define and individuate me: blue eyes, seasonal allergies, caustically funny. Perhaps my most bedrock trait, the one I’d reckon is most intrinsic and foundational to my sense of self, is that I’m a writer. My (fragile, most days) sense of worth and my (flimsy-feeling, usually) sense of purpose are bound up in this peculiar pursuit. Any belief I might have about the contribution I can make in this world is built atop whatever audacity and ingenuity I can summon with language. For me, to write is to venture something. As with any sustained endeavor, there is a failure rate that necessarily attends its pursuit - faltering and falling are an unavoidable consequence of attempting.

Being a writer is integral to what I imagine myself to be, at least as much as any of the relational things I am - husband, father, citizen, neighbor, etc. - writing is something I am, as much as something I do, as much a livelihood as a life. The voiceover for the nature documentary about me might go: “The North American Great Lumpy Ian is an expressive mammal who barters with other creatures, exchanging bits of language for nuts and berries and the occasional bit of carrion…”

So when voices in the culture, in the tech press and elsewhere, keep hollering “Stand aside, Scrivener - these bots can awreddy do your job!” and then push you to the ground and kick you in the face while cackling like they’ve long been itching for this Day of Retribution, I’m not gonna lie, it can feel a little hurtful. And personal. Which clouds my beliefs about AI.

For emphasis: factually speaking, I know that no one is pushing me to the ground and kicking my face. And while there may be cackling from some quarters, only a small percentage of it is directed at me. But the FEELINGS I experience in response to headlines and think pieces and social media ruminations that I have ALREADY BEEN replaced, that the functions I’ve grown adept at carrying out CAN ALREADY be executed AS WELL OR BETTER by this technology - these feelings can be intense. My rational self might still be going “Is that even true, though?” while my emotional self has snatched his go bag and is off in the deep woods, setting man traps along the perimeter of his prepper cabin, haunted as an imagined chorus echoes inside my skull “The gleaming killbots write at LEAST as well as you ever have, AND they don’t have all that nose hair.”

Hearing you might get REPLACED means getting confronted by the reality that you are replaceable. Which is discomfiting, and even scary.

Next up, in LLM Mindset Myth # 3, Part 2: AI is a REPLACEMENT - while the FEELINGS are real, is the THREAT?

LLM Mindset Myth #2: AI is GOOD TO GO

2025-04-04T00:00:00Z

Because Large Language Models (LLMs) are so impressive - even now, when still in their infancy, there is an already powerful and growing temptation to regard them as fully manifested, as having attained the full measure of their potential. This is simply not so. The reality that this technology remains quite new - a reality that necessitates keeping sight of a couple key ideas:

As with every tool or platform following its adoption and wide use, there will be a sometimes extensive Period of the Unresolved, during which its unseen problems and oddities make themselves gradually known to its users - problems that require fixing, oddities we must adapt to. Which brings up:

The still-emergent nature of our Relationship to the Technology. The introduction of any new tool into the way we do stuff necessarily involves becoming acquainted with its capabilities, gaining confidence in its use, and developing a (mostly ad hoc) “code of conduct” about it - beliefs about it, acknowledged or not, a set of rules we adopt, consciously or not, regarding it.

This last part is more consequential than you might think - as with any relationship, the one between people and LLMs will be shaped and even governed by thinking of which we’re not fully conscious, a kind of sub-cognition where beliefs and rules form. Now, before any of these become too entrenched, it is worth examining them and ensuring they’re as clear-eyed and intentional as we can make them.

In the literature of design ideation, there is a bedrock principle that applies to any project that demands answering two broad questions: 1) Should we choose NOT to make/implement this thing at all? And, 2) If the thing DOES exist, are we working to shape it in the smartest, most optimal way? In other words, once past the point where considering whether a thing should exist, what are we doing to make it better and more effective? This is where we are with LLMs - they undeniably exist, so instead of wringing our hands about whether or not they should BE, it is contingent upon us to understand the nature and scope of them, to develop methods and means for shaping them in productive ways, and noticing the ways they are shaping us.

Given the speed and complexity of LLMs, we succumb too readily to the idea that they are complete and fully robust, that the definitive versions of them are already upon us. This tendency is intensified by the pace at which LLMs have already winnowed their way into so many different aspects of so many of our processes. This misapprehension of their “completeness” is founded on the perhaps overly hopeful notion not only that LLMs are fully ready to assist us, and, more crucially, that we are fully ready to contend with the implications of LLMs upon our lives and work. Neither half of this notion is true - as in any relationship, there are kinks that need working out on both sides; we need to learn about each other for this to go well.

Assimilating any new product or platform into human work is almost never a plug-and-play, ready-from-the-jump proposition - doing so effectively entails careful planning, clear goals, iterative testing, and ongoing management. LLMs must remain subject to the same care and consideration as any other tool or platform.

It’s perhaps tempting to think “The technology must adapt to us.” This is true, certainly, but incomplete. Because equally true is “We must adapt to the technology.” Every innovation, even as it relieves us of the need to perform certain menial tasks or diminishes some of the low-level attention we must devote to something, carries with it newly minted obligations for us, the adoption not only of new procedures, but of new customs, the rise of a new idiom. Early in the life of any widely adopted innovation, people must collectively arrive at an agreed-to understanding of it, a shared sense of the thing - what’s “proper” and “fitting” in how we’re meant to use and relate to it. At the dawn of adoption of email, for instance, there was a prevalence of greater formality in language and tone (and length!) that carried over from earlier forms of correspondence. As we grew habituated to the form, we gained a clearer sense of how we were supposed to be using it - briefer, more direct, less genteel.

A consequence of any new technology, and dependable feature of the Period of the Unresolved mentioned above, is that there will always be some among us inclined to be unconvinced about its impacts, or even outright dismissive of it. “Internet? Psssh. That’s just for nerds.” At the dawn of the internet, this sentiment emerged as a trope right away - and, though it turns out obviously not to have been a defensible position, it was at the time a comprehensible one, in that we could be forgiven at the time for failing to see the seismic power and coming ubiquity of the technology. Likewise with LLMs, there are those who denigrate it as some parlor trick, some extension of our own laziness, some extraneous plug-in that performs no core function.

Conversely, there will be those during the Period of the Unresolved who will claim, just as wrongly, that we are witnessing the Final Grand Unifying Technology That Will Heal Our Wounded Hearts and Resolve All Conflict Ushering in Deathless Peace and Prosperity. Ridiculous. A tack hammer is ideal for certain tasks, but using it to play harpsichord would be reckless and dumb.

As with any new tool or technology whose legitimate uses are still making themselves known to us, its misuses will spring up in tandem. Along with naysayers, there will be those who will recognize illicit potential and will employ it to lie and steal. Our lack of familiarity will constitute grounds for exploitation until a pervasive sense of these potentials for misuse becomes widely known - in their earliest days, the various “Nigerian prince” scams were obviously far more effective than it’s possible for them to be today; because now we’re all on the lookout for such stuff, so they pose little credible threat.

The discernment we need - whether a technology is used for good or ill - can only develop over time and through use; familiarity, comfort, and savvy only result from experience and exposure.

The present moment we’re in with regard to LLMs remains deeply unreconciled - we have yet to hit upon a collective clearly defined sense of what these things are for, what they’re good at, what we need they meet. By the same token, we have yet to understand and develop our relationship to this technology.

LLM Mindset Myth #1: AI is an ANSWER MACHINE

2025-02-20T00:00:00Z

We’ve grown accustomed to “search engine delivery” - we request information and expect to have it provided instantaneously. With the rise of digital assistants like Alexa and Siri, we verge on a Star Trek-ish assumption that we can simply call out “Computer!” then bark out any question and receive the immediate (correct, we assume) answer in an affectless robotic voice. Over time, a pitfall of such an assumptive approach arises: sapped of our capacity to engage critically, we lapse into unexamined, often baseless faith in the results. With AI, though, an approach like this represents a fundamental misunderstanding of how AI operates and what it is meant to do - AI is not intended to function in such a “transactional completion” model, there is no one-to-one ratio of input-to-output, of query to answer. AI, and in particular Large Language Model (LLM) AI, should be thought of as iterative, as almost Socratic in its pathway to good results - it nudges along toward the precise and applicable.

It is important to keep sight of the fact that AI on its own is not a source of information - it draws upon information. It is therefore limited in important ways: the biases and blindspots of the people selecting and inputting the data, the borders - of topic, of detail, etc. - on a given subset of knowledge (because no matter how large, it is always a subset.) As with any information source, the quality, comprehensiveness, and precision of the outputs provided by any AI system are the direct result of how complete and correct the data going into it is. To deploy a Large Language Model well requires a reorientation away from this Magic 8-Ball mindset in favor of a more reciprocal one. When provided with appropriate conceptual and rhetorical parameters, LLM AI is extraordinarily adept at helping us to parse problems and draft (not “declare”) solutions.

It is true that LLMs are mind boggling in their speed, and can draw upon vast troves of public (and, regrettably, private) information sources for material, but it lacks the inborn and sophisticated meaning-making so natural to humans. It is better to think of LLM AI as an intern with access (not actual comprehension in any functional sense) to all of human knowledge, but who is naive and lacks sound instincts. AI is not some opaque fog out of which emerges apt responses to our questions - it doesn’t know everything; it knows what we instruct it to know. An LLM is not a librarian, it’s an interface - or, more precisely, if the LLM is a librarian, it is rookie capable of a basic set of tasks, not the sophisticated mastery of the seasoned polymath: the LLM has no sense of what it does not know; an expert human librarian will know that a title is not in the collection, and can direct you to another institution that’s likely to have it - AI can’t do anything resembling this*.*

In a general sense, it is fitting to think of any Large Language Model not as some know-it-all colleague, as some of us seem to, but as as a machine. And like any machine - a drill press, a drawbridge, a cash register - it is well suited to performing the finite set of operations of its intended use. It’s perhaps better to conceive of an LLM as a QUESTION Machine, NOT an Answer Machine - the LLM excels at providing us with short-span bridges between bodies of knowledge, or in compiling an aggregated sense of what many fields or disciplines have to say about a subject. If your project is to design a tugboat, the LLM obviously cannot do that for you, but it can draw upon pertinent bodies of knowledge - fluid dynamics and engine requirements and maritime law and so forth - none of which will generate renderings of a tugboat. All of which can, however, provide a series of evaluative filters through which we may assess our evolving plan.

The LLM, in short, can be tremendously capable at the iterative interrogation of a project, where we provide it with applicable information. The LLM does not and cannot complete any complicated project for us, but is excellent at helping us ask about our own thinking about a complex objective: What conditions and factors might I have missed in devising this project? How might I refine and improve upon the current state of this project? It is this that separates the LLM from all previous machines - instead of firing rivets or spot welding, it can help us think. A Large Language Model, where we are clear on its intended use and capabilities, can introduce a kind of distributed cognition, a kind of multi-brain, into our processes - adding a kind of research assistant that does not generate novel ideas, but provides a useful, perhaps one day even indispensable, sounding board for tinkering with our own models and methods, for nudging our practices toward “best.”

LLM MINDSET MYTH #1A: “WHY CAN’T WE JUST RUN IT THROUGH CHATGPT?”

Well, you can, obviously. But will doing so provide you with anything resembling what you’re looking to attain? In terms of sharpening THE means of expressing ideas or refining your understanding of the problems you’re looking to solve? Not likely.

LLM platforms like ChatGPT are better understood as a series of switches than as a vast, sentient content generator. The LLM is NOT sentient, it does not GENERATE ANYTHING, not in the At the granular level, LLM systems function a lot more like auto-fill than human intelligence - it’s possible to have ChatGPT write you a sonnet, yes, in the sense that it will follow the rules of meter and rhyme and so forth, but it will have nothing like the depth of feeling or insight that a human writer can achieve. A person is able to compose thoughts and express them, in real time; the LLM guesses at what the next word or phrase is likeliest to be - with, it’s true, a speed that’s near incomprehensible, which no doubt contributes to our inflated sense about its capabilities. But this difference is crucially important to note - people naturally have a holistic approach to grasping and conveying concepts, the impressions we form and the language for making these known emerges more or less simultaneously; whereas even the most sophisticated LLM is inherently, unavoidably sequential and probabilistic.

The peril of this unmediated “run it through” approach is not that the results of doing so will be useless (though left to their own devices, the tendency of such platforms to produce gibberish is well documented,) it’s that they will be flat. If in the human brain, intelligence may be said to reside in complexities of structure, in the folds and wrinkles, unaided LLM’s “thinking” might fairly be called smooth-brained. It will tell us variations of what we already know - it will produce phrases adjacent to the expected. Or arrive at conclusions that are naive - not incorrect, exactly, but obvious, lacking nuance. The knowledge upon which LLMs will draw is vast but shallow, effectively the limitless inexpertise of the generalist. What LLM is exceptionally good at, however, is not in what we think of as cognition, but in providing us with cognitive assists - we should regard these systems as an enrichment tool, not as a generative one. Even for humans of the most arresting and singular intelligence, those transformative jolts of inspired, visionary thought are comparatively rare - but the prospect of them remains within reach. For ChatGPT and other LLM AI platforms, such leaps are not possible.

Say you are an architect hired by a client to build a home outside of Santa Fe. You will bring your knowledge to the project, certainly - all the years of information and insight you’ve gathered about materials and processes, esthetics and function. So far, LLM - provided access to facts about materials and building codes and architectural technique - will more or less be able to keep pace. What such platforms cannot do, however, is draw upon the flashes of inspiration that come to people, or the facility we have for interpreting the mind and desires of a client, or the intuitions that arise across the life of such a project that we heed or ignore. Consequently, if we submit a request to ChatGPT or any comparable LLM platform to “help design a house outside Santa Fe,” anything it tells us is likely to have a “received wisdom” feel, that might be broadly apt in principle, but is not germane or applicable to this project.

Reorienting ourselves to what LLMs actually ARE competent to provide such a project would be to rightly regard it as a component-level tool, an interlocutor adept at helping to navigate the dozens or hundreds of branches on the decision tree of conceiving of such a project. For instance, an LLM could harvest climate data for the exact GPS coordinates of the proposed site, and provide aspects to consider in the selection of building materials, or could match green building standards and water conservation practices to the proposed structure and landscaping. In other words, AI does not, cannot, design in the way that we can, in the sense of synthesizing information and inspiration to arrive at wholly novel conceptions of how to grapple with a particular set of problems. A Large Language Model, where it has been provided the right data sets, is very good at helping us bridge the gaps (leap synapses, find new relationships) between various bodies (points, areas) of abstract knowledge - in other words, it is a computational means of arriving at the unexpected, one where each step in attaining a “thought” bears some clear relational connection to each one before and after it. AI is not well suited to shaping an overall vision, but is highly useful in helping to render the many this-vs.-that distinctions that constitute the ingredients of vision.

Where a Large Language Model like ChatGPT can prove truly transformational, even revolutionary help to us, is where it serves not as our motorcycle, but as a sidecar - it does not generate power, it does not steer, nor decide on destination, but it can expand our capacity. We - as the drivers - must keep clear on our own cognitive capabilities and on the augmentations that the LLM component can provide. It is only when we do so, where we retain a solid “who’s good at what” grasp of the nature of the dynamic between us, that these platforms may realize their fullest potential as the fastest, smartest assistant we’ve ever had.

LLMs and Mindset Myths - an Introduction

2025-02-13T00:00:00Z

Over the next several issues of this newsletter, we’ll be examining some of the beliefs that have sprung up about Large Language Model platforms (LLMs) - what they can and cannot do, what the nature and purpose of them is, what their relation to us is. We’ve identified ten or so types of what we’ve come to call Mindset Myths surrounding AI generally and LLMs in particular that have come to define and drive how we conceive of, comprehend, and relate to this technology. While this might seem like a purely theoretical or even trivial concern, we at ZeroWidth do not believe this to be the case - the ways we understand things shape the decisions we make about them, and the subsequent actions we take. With a technology as impactful and far-ranging as LLMs, a technology - perhaps the first, really - that can be said to improve and expand upon itself, the manner in which we engage with it and the beliefs we hold about it come to be very consequential. Taken together, these tens or hundreds of thousands of decisions, propelled as they by underlying belief, aggregate to form the basis of broad practice and policy.

The truest subject of this newsletter is the relationship between people and this technology. We’re not here to not to wonk out about the tech specs of LLMs (Pete is the only member of the ZeroWidth squad capable of doing so - John is Generation Can Read Cursive) - there are plenty of places offering that; we’re not here to conjecture about the state of the field in a decade’s time. In fact, given the pace of change, the “role” of “futurist” (a “field” that might as well be called The Professionally Vague, or Dependably Wrong) - lord knows there are plenty of places to find stuff; we’re not here to crypto-bro proselytize about some new branch of (Godless) Prosperity Theology where LLMs are performing some (imagined) form of automated and (impossibly) lucrative content generation while you get daiquiri-hammered on the deck of your yacht - sweet heaven knows there’s MORE than enough of that stuff available to you.

If we examine any among the (growing, God help us) number of pronouncement-makers and prognosticators ostensibly analyzing the field of AI, we find as much attempted persuasion as we do anything resembling a serious examination of the present state of the technology and its prospects - people, in other words, attempting to shape your beliefs about this subject. Here at Distributed Cognition, we aren’t shilling for any one position, we’re not advocates or evangelists - the aim of our ongoing inquiry is to encourage the reader - you, individual human - to identify and assess your beliefs about LLMs and to determine for yourself if they square with what you conclude to be true.

We’re attempting to contend with the nature and outcomes of the interplay between us and LLMs. The once general-use phrase “human centered” has grown narrow over time and through specified use - “human centered design” is a by now familiar (even thumb-worn) phrase intended to speak to a specific philosophical approach to design, a set of considerations and precepts in a particular form of problem-solving. What this newsletter posits is the need of some equivalent set of reflections and recommendations - an approach that is human-centered in the broadest sense. What we resolve to offer is an arena of ongoing inquiry that explores the intersection of people with these platforms, and to demonstrate ways each partner in this relationship - human and LLM - can unlock the fullest potential of the other.

In order for this to be possible, though, requires a deep understanding of the nature of the relationship and the direction it’s taking. Understanding this demands that we observe our own thinking about it. So perhaps “human directed” is a more apt and applicable phrase for expressing what we’re after, here. Over the next several editions of this newsletter, we’ll discuss each of the Mindset Myths we’ve identified and the ways in which these can limit and even distort our comprehension of LLMs and their capabilities, and to start devising a clearer sense of our own obligations and expectations in this relationship, a kind of Code of Cognitive Conduct, if you will.

As you read through these next few issues, our hope is that you’ll begin to notice the Myths we’ve identified at play in your own interactions with LLMs, and that - since we don’t presume to have the final word on any of this stuff, that you’ll let us know about comparable Myths that you’ve witnessed or fallen prey to. This newsletter strives to serve as one voice among many within the field, to number among those that join us in (we hope) considering deeply the many questions that arise at what is still the dawn of this technology and our relationship with it.

The Mechanical Turk, Part 3: Lying Liar Filled With Power

2025-01-31T00:00:00Z

If we accept the Mechanical Turk as a fitting metaphor for an aspect or potentiality of LLMs, namely of their ability to aid in deception, then we’re obviously confronted with the question of what to do about it. If we aim to avoid such deception, our impulse might simply be to refrain from engaging with material generated with ChatGPT or any of the others, a position that grows less and less possible with every passing day; the near-ubiquity of LLM content renders it essentially unavoidable. If, in our efforts to avoid deception, we resolve to do a better job at detecting the presence of LLM intervention, we miss the point. The point is not simply to affix some kind of lie detector plug-in.

The recent proliferation of “AI detection” platforms and software, ostensibly designed to detect novelty of language, as means of inferring originality of composition - a function, it’s worth nothing, they cannot actually carry out in their present form - does nothing to interrogate the intent underlying or animating a piece of text. These detection systems are essentially a binary answer to the question “Written by robots? Y/N?” This is a valid question, surely. But is it the most salient one? Does such a question help unearth the objectives of a text’s creator? Their motivations? Their tactical approach? No. Does this binary robot question even contain any element of “Is this a lie?” It does not. This kind of low level Turing test is not what this series of posts has been about, however - it’s a call for the elevation of our own critical faculties, a leaning into our own capacities of discernment and distinction-making. It is a call, above everything, to remain un-dazzled. At the dawn of any new technology, there is a natural tendency to become so impressed by novel capabilities that we lose some of the skepticism required to make sense of the world. After this dip in our dubiousness, we return to it once we’ve grown accustomed to the once-new technology, we return to ourselves, once again becoming alert to the presence of bad actors. Who are always human.

It’s like seeing somebody on stilts for the first time - after a lizard-brain scream of “AIGHH! A GIANT!” we calm down, figure out what’s going on, and scoff at the circus nerd.

The antidote to bedazzlement is familiarity - use over time engenders comfort and clarity. The advent of indoor plumbing or gas lanterns or automobiles were all likely pretty mind blowing for folks first encountering them - over the course of intervening generations, these have all been folded into our prevailing sense of what life is like, and we take in stride the improvements made to them over time.

We return to the caution Lector cave (“reader beware”) - because rightly or wrongly, the onus of determining the motivations underlying any communication we receive from whatever source falls to us as the recipients of it. If there are bad intentions at the heart of any message, whether these are intended, or it they arise from inattention or ignorance on the part of the sender, the consequences will likely be ours. Whatever the case, we are always better served by a close rather than a cursory reading.

A close reading demands an active form of engagement just about antithetical to the ethos that springs up around any supposedly miraculous new form of communication. The more “seamless” and “automated” it becomes to generate content, the more tempting it becomes to treat the other end of communication, the deciphering end, the same way, to skim, to glean no more than the gist. The perils and downsides of this kind of generalizing and approximation are manifest - erosion of shared understanding, diminished clarity, time lost to avoidable repetition. And ever-widening gaps, in other words, through which lies may pass.

Think of written communication over the course of our history - in broad strokes - the earliest clay tablets contain essentially mercantile transfers of information, inventories, lists, and receipts, essentially, where there could be said to be a one-to-one ratio of representation: a phrase like “50 casks of wine” was not meant to convey anything beyond an amount of a readily understood noun. As the sophistication of language and our means of expression have snowballed over millennia, the linkage between our representations and the substance they’re meant to convey grows less and less concrete.

The means by which a written communique comes to us have grown similarly more remote and abstracted. Getting a handwritten letter - even one concerning some comparatively straightforward business transaction, say - carries with it tangible evidence of the effort put forth, a physical artifact that invites and insists upon scrutiny. Whereas a telegram might feel to its recipient less directly attributable to a single intelligence - the effort to compose it is there, but it arrives in a standardized fashion, delivered by intermediaries. An email, then, dispenses with the physical artifact entirely, in an insubstantial form identical to all other emails. Each subsequent digital format and platform has firmed up this tendency toward standardization - even the little “personal” touches have been codified (emojis, digital signatures, etc.) The pace and volume at which communication arrives grows with every advance.

Today, we must add to this puzzle the fast emerging question of whether something was even written by a person. But again, this outside of certain narrow contexts - academia, say, where notions of intellectual ownership are more consequential than elsewhere - this is not the central question we’ve grown convinced that it is. Fixating on the “Written by robots? Y/N?” question funnels us into a false “Real/Not Real” binary and encourages shallow-pass conclusions. If, in our dulling faculties, we stick to the deciphering side of communications, becoming uncritical consumers of the present Glorified Autofill/Rudimentary Chat Bot path that LLM use mostly follows, we succumb to the Fallacy of the Turk - mindlessly accepting checkmate at the hands of a herky-jerky turban-wearing opponent without ever knocking on the side of the cabinet and calling “Well played” to the human player we SHOULD know to be inside. By exploring the generative side of communication and the many ways LLMs can assist us in framing, testing, and developing ideas, we become true early adopters of the technology, not simply more adept and acclimated than most, but helping to shape how it evolves.

The ZeroWidth Workbench is intended to provide its users with just such a laboratory. Over time, we’ll detail the ways in which use of the Workbench doesn’t simply streamline processes, it can provide an array of cognitive jolts that improve and clarify the user’s thinking on a subject, expanding and deepening our capacities for prototyping and problem solving.

The Mechanical Turk, Part 2: Lying Liar Filled With Promise

2025-01-30T00:00:00Z

Returning to the Mechanical Turk (as agent of deception). As the use of LLMs spreads and winnows its way into more and more forms of the ways we talk to each other, the series of questions we must develop the habit of asking is less centered in the “Is this a human?” arena.

Every technological advance brings an EXPANSION of complexity.

This process is an accretion - existing modes of complexity are not displaced by new ones, they form the base layer of a new structure. The telegraph compelled us to become able to carry out rapid exchanges with unseen people far away, the telephone required that such exchanges be conducted verbally in real time, etc. It’s by now well understood that we’ve moved, in the present age of data and information, into a largely symbolic economy—one where meaning and value are increasingly constructed through representations, abstractions, and interpretations rather than tangible goods or direct transactions - the condition of being a “buyer” has come to more fully encompass being a “reader,” our role is as much “interpreter” as it is “recipient.”

The real world still exists (for the moment, anyway,) but more and more we must occupy a kind of liminal space between actual and abstraction. This ascendent place of symbols in the exchanges between us - in the form of brands, data visualizations, algorithms, etc. - means they carry great (and growing) weight in our decision making. These symbols - abstractions - can mask the underlying uncertainties and biases present in their creators - actual people in the actual world; with the result that the abstraction can eclipse the actual. The reason, therefore, that Lector cave (“READER beware,” and it’s Latin, so “cah-vay,” not not long “a” “cave”) comes to have such importance is that each interpretive act is necessarily the result of multiple forms of deciphering comprised of the meaning and relevance of abstracted symbols, the intentions and competence of the sender.

This Accretion of Complexity principle demands, therefore, that the lens through which we view everything grows more critical. An Attention Economy does not mean simply that our attention has been commodified - though obviously it has - it means that we as participants in an Attention Economy must necessarily grow more discerning in where we permit that attention to be directed.

Imagine the Mechanical Turk but with a cabinet made of glass - we’d be able to see the human operator wedged inside. The nature of the Turk would be unalterably changed, the purpose driving it (i.e. to deceive the viewer) eliminated entirely. Or imagine that the components of the Turk are disarticulated - if his pieces are unbundled from one another, we can immediately separate those carrying out some function (what we’re here calling the actual) from those intended to misdirect or misrepresent (the veneer of abstraction that performs no work/serves no purpose).

Applying a similar thought experiment to LLMs today means parsing A) the intent of human actors who provide the inciting inputs (actual), B) the resulting textual output (abstracted/symbolic) and squaring all that with C) what we’ve determined to be true/accurate/pertinent (a hybrid of the actual and abstracted). Lector cave thereby becomes not simply a caution about checking the veracity and validity of all the textual stuff that comes our way, but a kind of Prime Directive for how we must navigate the whole of this hybridized actual-plus-abstracted world in which we are all operating.

If we think of the Turk in relation to LLMs, the opacity problem is comparable but is not identical - it’s important to keep sight of the fact that for LLMs, the nature, character, and source of the opacity differs from that of the Turk: where the Turk was born to deceive, an LLM is not. LLMs can be wrong, they can be mistaken, but left to perform the tasks we set for them, they will not provide outputs with any intent to deceive. Deceit by design is (so far, anyway,) the province of people. In its “natural” state, an LLM system is devoid of any capacity for moral lapse or ill intent - it’s only the people steering it that can introduce such elements.

The LLM can be wrongheaded, it can be shortsighted or misguided, but it is not a liar.

We are, though, or can be - we alone have that capacity. So Lector cave comes to occupy a crucial and central role in how we engage with the information bombarding us each day. At the core of a concept like “media literacy” in fact, there must be this ongoing, updating, constantly revised estimation of both content and intent. Proficiency with the tools in question can go a long way to sharpening this capacity for estimation - a solid understanding of what LLMs can and cannot do, how they function, and the precise manners in which they may be deployed in service of deceptive intent.

A core aspect of the ZeroWidth approach, the heart and backbone of the ZeroWidth Workbench, is to give the LLM-Turk a cabinet of glass, so to speak, or - even further, to dismantle the Turk, to remake our relationship with LLMs. To unpack the monolith into comprehensible constituent parts, modules that will perform the tasks we set it in ways that are transparent, ways that render the product of LLMs a bit more human. By pulling the veil away from the workings of Turk, we make the structures inside and underlying the Turk to direct his efforts in truthful fashion.

In the coming weeks and months, ZeroWidth will demonstrate how its products and services seek to be a principled and conscientious presence in the field of LLM use. We’ll do that in part by outlining ways that the prevailing thinking about LLMs gets it wrong or misses the mark. We wish to become a resource in helping to plot the course this technology takes, and a leader in developing sound thinking around it. Here’s hoping you’ll tag along.

Next, our final part of this series: The Mechanical Turk, Part 3: Lying Liar Filled With Power

The Mechanical Turk, Part 1: Lying Liar Filled With Lies

2025-01-29T00:00:00Z

At one end of the insanity range, we hear stuff like: “Inside a year, AI’s will go rogue and swipe your medical records and start selling your organs on the Dark Web. And funneling the proceeds to ISIS. hashtag-hellscape!!!”

At the other end of the lunacy spectrum, we hear stuff like: “By this time next week, AI will be performing life saving surgery unassisted. While doing kickass skateboard tricks and shredding on guitar. hashtag-rad!!!”

At this point, where AI is concerned, it’s become difficult even to adequately exaggerate the claims being made about it. And Large Language Models (LLMs) are by no stretch exempted from this frothing madness.

There is a period, early in the life of any technology - especially a technology with transformative potential that can be broadly applied - where just about anything we might have to say about it is wrong. The truth - as with most truths - is shaded with aspects of both extremes: LLMs have already proven to have great utility and to have transformed many aspects of living; they also represent a risk when misused.

Just about all the pronouncements and prognostications, often as not delivered by those attempting to monetize their own pronouncement-making, are wrong. The landscape of analysis seems crowded with as many utopians as doomsayers, and much of what they have to say is rooted in exaggerating either the promise or the perils of the technology. When really, as with just about every tool or platform, the peril or promise hinge upon the intent of those using it.

The Mechanical Turk was (claimed to be) a chess-playing automaton that toured Europe in the 1770s. The machine would take on and defeat any comers in a live game. And while the notion of buying a ticket to go see some variation on Zoltar (1988’s Big) play chess against some series of randos lands in the mind of the present-day entertainment seeker as extraordinarily lame and boring, in its day, the Turk was a sensation. This turban-wearing wonder, wrought of wood, wire, and metal, toured the Capitals of the Continent, where crowds were hungry, apparently, to watch a crudely painted (and frankly quite racist, by today’s standards) simulated fella from the Mysterious East move very slightly in a herky-jerky way. The Turk would face any opponent, and remained undefeated in his public contests.

It goes without saying, of course, that the entire thing was a fraud. An artful fraud, yes, an ingenious fraud, without question. But a fraud nonetheless.

Style note: though the Mechanical Turk was without question an object, because that object was crafted in the form of a person, and because that person-shaped thing was constructed in a way that it appeared to be performing tasks, for clarity/simplicity, we’ll be referring to the Turk as “he/him”, as we would a person, and not “it.” Cheers, eds.

And, when thinking about LLMs, it's this element of the Turk that’s most salient and instructive. I’ve come across a couple bloggity/newslettery posts that draw parallels between the Turk and LLMs where he serves to illustrate the “Black Box Problem.” The Black Box Problem describes the persistent reality that as an LLM engages in its blistering-fast predictive work to arrive at its best guess for the next right word/phrase/sentence, we are not able to know/see, given the vastness of its inputs, what source material it’s drawing from, or to reproduce the “chain of cognition” pathway that leads it from our request to the output it provides. The Black Box Problem is a hugely complex and fascinating element of the LLM puzzle, but the Mechanical Turk is NOT a fitting image for describing it. Yes, there were unseen operations taking place inside the cabinet of the Turk; yes, there was an (intended) opacity to the means by which he arrived at his results. While this interpretation of the Turk as relates to LLMs is comprehensible (his workings were not visible, his functions not readily understandable, etc.,) it is not a defensible one - the entire reason for the Turk coming into being was to deceive his audience for the gain of those running the scam.

The DEFINING characteristic of the Turk was DECEPTION.

For all the cleverness and engineering sophistication contained within him, the Turk’s raison d’etre (Hey. Whoa. Whoa, whoa, whoa. Let’s keep the French to a minimum, shall we? - eds.) was to instill in his spectators the false belief that he (vs. some unseen operator) was a chess master. Which he very much was not. He was not simply unskilled in chess, he was unskilled in every way imaginable. He couldn’t move on his own. His sightless eyes were painted on. He did not strategize, did not respond to an opponent’s moves, could not see the chessboard in front of him. The intelligence underlying the Turk, the only agency at play, was that of his human operators - the host/presenter who lied about the miraculous clockworks inside the Turk’s cabinet that made him “move” and allowed him to “interpret the board and state of play,” and the (necessarily short and slender) human chess master wedged inside the cabinet who actually carried out the gameplay.

The Turk’s cabinet was fitted with two sets of hatch doors with a false wall inside, so that the Host Charlatan could open them up, magician-style, to let audience members inspect the inside and establish that there was no human operator in there. But, like a magician, these hatches would reveal one compartment at a time, permitting the Concealed Gameplay Charlatan to contort himself out of view when each was opened. So it’s THIS component of the Turk’s construction and operation that is MOST salient when relating him to the Large Language Models of the present day. The demonstrable intent to deceive, the misrepresentation of the Turk’s (nonexistent) capabilities that renders it an instructive episode of history. It is a useful caution to us today, especially those among us who wish to maintain a healthy skepticism about LLMs.

Most of us are familiar with the Latin phrase caveat emptor, or “buyer beware.” Which is sound advice that applies to a wide range of interactions. It is advice, however, that’s applicable mainly to material goods and services, advice that reminds us to measure the item we’re provided against the promises made about it.

We posit that Lector cave, which (according to the near consensus of several no doubt AI-powered online English-to-Latin translators,) is “reader beware” is a more germane and helpful caution to keep in our minds.

Next up: The Mechanical Turk, Part 2: Lying Liar Filled With Promise

Who Even Are You People? Part 3: Pete Binggeser

2024-12-05T00:00:00Z

As noted in Part 2, Pete Binggeser was still an architecture student when John Cain saw his mad scientist potential and had Pete come work with him at his company Iota Partners. In the first part of Distributed Cognition (see please Cuttlefish.) we described Pete as a tinkerer - a word than can imply something of the dotty amateur, futzing around in a garage (think of Belle’s dad in Beauty and the Beast,) but can also convey - and this is very much true in Pete’s case, somebody who’s resolved to crack the code of things, to both arrive at a deep understanding of how a thing functions, and, having decoded it, to devise ways to improve it.

And here’s where the question of deep understanding becomes a critical distinction, because “improvement” to Pete does not simply mean little tweaks that yield some tiny incremental gain, it’s a willingness to radically reimagine the thing itself and to remake it in ways that push it closer to the essence of its purpose, that make it more quintessentially itself. He’ll be the first to tell you that sometimes it’s incremental improvement that’s the only way currently possible to do this, so is an approach always worth considering and pursuing, but his is a mind willing to take great leaps, as well. As Cain notes about him, Pete has a rare capacity for holding the details of a thing in his mind, and the little ways of improvising those, while simultaneously reconceiving the thing altogether.

When Cain interviewed Pete for the gig with Iota (which readers will recall was a company wrestling with the many applications and questions at the dawn of Big Data, mostly by affixing various sensors and readers on products and in environments to harvest information for later analysis,) Pete brought in a visualization project he’d done that incorporated a bit of the “sensors in buildings” stuff he’d been doing in professor T.J. McLeish’s class at IIT, but in a way that reduced it to its essentials, portraying the capabilities of the technology at least as much as the behavior being analyzed.

The image Pete brought to John was a hectic-looking many-pointed asterisk type of deal, a picture that looked to have captured lots of ricocheting movement, which is exactly what it was: with a MacGuyver-ed combo of Arduinos (open-source single-board micro-controllers - this is the kind of word Pete routinely drops into conversation with the expectation that everybody would know what these are - I did not,) reflective tape for clear motion capture, and the repurposed remote for a Nintendo Wii, Pete created a visualization of his Boston Terrier playing fetch in Pete’s tiny apartment. In some mix of being impressed by this level of overkill-to-solve-what-was-never-a-problem, and the elegance of the improvised spliced-together components that yielded a clear and compelling visual representation of this under-studied phenomenon. Because WHEN will federal funders FINALLY devote adequate research dollars to the seminal question “But where exactly does the doggie go when playing fetch within a small space, though?”

After a couple years, Iota Partners got bought by Sapient, so the scrappy start-up team got folded into a much bigger organization. There was a foosball table in Sapient’s Chicago office which was a central fixture that provided frequent decompress-while-letting-your-less-conscious-brain-turn-some-problem-over-while-you’re-doing-something-else activity that, in a deadline-intensive, high-pressure environment, was a welcome aspect of office culture. Pete being Pete, he quickly concluded that “Hey - we should probably be ‘sampling the wares,’ here.” Meaning that the work for clients involved “instrumenting the world” in a bunch of cool ways, and “we’ve got hundreds of switches and microcontrollers and wiring lying around,” so why not build a custom thing to measure an existing component of our workplace conduct AND up the stakes of gameplay.

So he adds a couple goal-detecting switches and LEDs to automate score-keeping, then mounts a card-reader next to the table. Now every player can scan their company ID badge at the start of each game. He builds a database with everybody’s badge number to start, so you swipe your ID, enter the final score on a tablet, and an informal, ongoing tournament starts taking place. But then Pete’s like “Manually entering? What’re we, cavemen?” So he mounts a switch on each goal, boom, the score gets tallied automatically, winners and losers of each game updated in the database without a keystroke. As the body of foosball data grew, Pete finagled with his system some more, till gameplay got to the point where each player’s assigned a global ranking, like in the world of competitive chess, and the table itself is live-tracking, compiling, and updating the data to the point where the system could provide visualizations on the stats of any player.

This hobby project - which Pete started on more or less a lark, and which he added to and augmented on his own time - came to be a prominent stop on client office tours. This peculiar artifact demonstrated a great deal about the approach the company wished to portray: taking seriously problems that might seem minor, even trivial; bending and reshaping existing technology to uncover new uses for it; embedding the information capture required to analyze a problem into the environment where it takes place; synthesizing a wealth of information from multiple sources and presenting it in clear narrative ways. Over time, the foosball table came to be regarded as so emblematic of so many dimensions of the work that Pete got sent to other offices to install versions of it there, building a serious multi-site live-updating data set for a frivolous activity. In Toronto, though, where they are squares, they were a bit taken aback, like “You guys play foosball during work hours!?!” Yeah, Moose Knuckle. We do.

Here’s a snapshot of the foosball capture-and-analyze solutions Pete devised:

So, now that Pete is a grownup professional, co-founder of a cool company, guest lecturer at places like Berkley, etc., you might think he’s slowed his nerd roll a little bit. THINK AGAIN, Buster! When he’s not devising elegant work-related solutions in the material and virtual worlds, he's building his own synthesizers! That’s right. Mr. Man is deep into configuring and reconfiguring modular components of analog synthesizers. He calls forth music from NOTHING. And he can still tie a monkey’s fist, I bet.

Who Even Are You People? Part 2: John Cain

2024-11-21T00:00:00Z

Ian Belknap (IB): So you’ve founded a number of different companies - tell me about those.

John Cain (JC): I co-founded (with Rick Robinson) E-Lab LLC (an at the time revolutionary research and design consulting firm that relied on the ethnographic tools and frameworks of the social sciences, an approach that’s come to be codified today under the banner of UX design) in the 90s (full disclosure - this is where John and I met - I worked for him at E-Lab for a few years, as a kind of uncredentialed court jester) which got bought out by Sapient. I stayed on a Sapient for a few years - on the one hand it was gratifying to see projects incorporating the approach developed at E-Lab carried out on a larger scale, on the other, I chaffed a little bit feeling like a cog in a machine. I left there, did some consulting. Then my friend Des from Sapient called me to help turn around this data informatics company in Portland, Maine, of all places - It was overpayment detection. It used data and analytics to look for patterns of fraud, waste, and abuse in Medicaid claims. After a few years, that company got bought by United Health Group, where I stayed till the day of my earn-out. Not a day longer. No love lost, there.

But the thing that was kind of cool was I learned a lot about Big Data - this was before it was even called Big Data. And I learned this word “algorithm,” and I hired people that wrote code, And I kind of geek on that stuff because I went to engineering school as an undergrad.

This was around 2010. And there were all kinds of new f***-ed up sensors out there. Things on your wrist, things in your hand on your phone, the iPhone came out, 2007. It was all kinds of accelerometers and crap on your phone and car companies are reading accelerometer data off your phone and they're saying, “Hey you're a fast driver, we got to up your rates,” and that's a really interesting shift in the way the world works. And I became fascinated with this change.

And I wondered if you could do an E-Lab-style company from the basis of capturing and analyzing streams of data? So, I started this thing called Iota Partners with Rick (Robinson), and a couple of other E-Lab veterans. I found this guy named T.J. McLeish, and lo and behold a freakin’ miracle happened. I'm in Chicago - not Cambridge or Silicon Valley, or any of those places. I need some kind of Thomas Edison-meets-wizard type of person that can build sensor boards and we can deploy them on Windex bottles and put them on dog food, bowls and s*** like that because everything's miniature and Wi-Fi… we put them on packs of gum and we put them on Skittles bags. We put them in Target stores and we put them in homes and on boxes and bags. And we start generating all this data, like a huge amount. And we start playing around with the data like, “Hey, T.J., we need some people that can monkey with the data.” And he goes “Oh, I'm teaching a class at IIT (Illinois Institute of Technology) in the Architecture Department about using sensor data, I got a couple of students I could bring them in.” And I'm like, “Great! " because we have very little money. And one of these students is a guy named Peter Binggeser. And I pay him some piddly wage and feed him all my weird ideas, and Pete's always like “Yeah, I could make a visualization like that.”

Then eventually Sapient steps back in to acquire Iota Partners. And then we do even more s*** because we now have huge budgets to waste on cabinets full of lithium-ion batteries, just frickin’ Looney Tunes s*** - it looked like the MIT Media Lab in there. And we got all kinds of kids soldering up circuit boards, and putting Google glass on robots, and hollowing out Mr. Potato Heads and putting cameras in them to observe the way people wash dishes, and stuff. And I’m like “This Pete kid is an absolute wizard.”

Pete has got this speed and pragmatism. But he's also got this open-ended creative component, like his former instructor T.J. does, but T.J. gets lost in it, whereas Pete can direct his creative problem-solving towards the solution to a thing and make it overnight, and that's astonishing quite frankly.

IB: Why’s he like this, do you think?

JC: Pete tells a story about this one Christmas his grandfather gave him this footlong piece of rope along with this little Xeroxed paper, like three inches square, with instructions for tying various knots. This was obviously an assembly line/”same gift for every grandkid” kind of deal. So, where most kids would be like “thanks for nothing, Gramps” and toss it aside, Pete studied the little square and took the rope and mastered that group of knots. His grandfather's German, so there's this ribbon of ascetic pragmatism that’s like “ you have to learn to be curious about the simple thing because it's really complex, and you have to learn how to tie a knot, and you have to learn how to get something done completely.” He likes the act of creation but he's got this, just remarkable ability to turn the creative act into a finished thing. Usually in a super elegant way. It's quite rare.

So, anyway, after a while of doing Iota-within-Sapient, we kind of conclude that there’s way bigger players and we probably can’t make the kind of impact we hope to, so I go off to work with Bill Kendrick (who’d engineered the Iota buyout,) who’s gone on to start Material, and Pete goes off and works at Twitter. But we stayed in touch and did little projects here and there.

And then like any observer, I see AI bubbling up everywhere. And what I see in this shift towards AI is both amazing upside potential, but the possibility for horrifying dystopia, as well, if we don't get after shaping it differently. So, there’s this seismic societal and technological shift taking place, akin to the one that led to the blending of disciplines at E-Lab. A kind of shift that one really doesn't deserve to see twice in one's lifetime, but here I am watching another one happen. I want to get attached to it…

So after I got bored enough by consulting gigs that were essentially very similar, and Pete quit Twitter, we got together and started thinking through some of the implications and applications of this stuff. And we immediately began seeing all this potential for this impossibly potent tool that was not getting widely utilized, and starting noodling on how to unlock some of that. The more we delved into, though - the more we concluded that AI’s real power would get unleashed if we built this ongoing, constantly expanding set of tools where people could monkey around with it themselves - not just feeding scripts to chat-bots, which seemed mostly what LLMs were getting used for - but in real-time prototyping, in interrogating and modifying concepts on the fly, in building and testing processes for almost no cost. So, while we’re obviously grateful for the consulting work we get, the real legacy we’re trying to put in place is a nimble, adaptable, easy-to-use set of modular components that put the power of this frankly astonishing technology in the hands of just about anybody.

That’s what the ZeroWidth Workbench is.

Next Up:

A Story of Foosball Overkill with Pete Binggeser

Who Even ARE You People? Part 1

2024-11-14T00:00:00Z

I’m Ian Belknap, human writer.

John Cain and Peter Binggeser are the founding principals of ZeroWidth, a company that resolves to change the way we think about and use AI. They have hired me to help figure out what to say about all this.

“All right, then, PROVE it,” you might find yourself saying, a little too loudly. To which we say: we will. We will prove it. In upcoming Distributed Cognition posts. So maybe dial it down a little, bud?

In THIS post, we’ll tell you about John and Pete and why they work so well together. Here are some elements they share in common:

Both began their professional lives in one discipline - Industrial Design for Cain and Architecture for Binggeser - and have since greatly expanded their palette of inquiry to include many other arenas of thought.

Both wear corrective lenses.

Neither of them keeps weird pets - your iguanas, your skinks, your tarantulas - these are mammal guys. And not ferrets or chinchillas or cats - we’re talking about primo mammals, here - which means we’re talking about dogs.

Each is “holy crap” smart in his own way. And it’s the way these “holy crap” smarts braid together, in such arresting ways, that sets ZeroWidth apart in this now-crowded field.

Both have little patience and zero reverence for precedent - both believe that just because something has been done some way forever is no reason to keep doing it that way. This extends to beliefs and stories around the way we do things - if these are not effective, hit the eject button.

Both also have little patience and zero inclination for trend-hopping, the “stampede after latest glittering object” impulse that drives too much of LLM activity (we’ll have a series of posts about some common fallacies that contribute to this phenomenon.) If you’re chasing something simply because it’s new without some compelling justification for wanting it, hit the eject button.

Each carries an eject button with them at all times.

Both demonstrate a restlessness of mind, an abiding curiosity that drives them to master new skills and grapple with new modes of thinking. ‘Member how during lockdown everybody seemed to be learning crochet and baking bread and studying Mandarin? These two are like that all the time. Which, for a bucket of sloth meat such as myself, is exhausting.

Neither of them was instrumental in getting Breakdance included as an event in the 2024 Paris Summer Olympics. So take your backlash elsewhere.

Both are semi-respectable family men. Meaning their wives and children half-respect them. Zing, boys. Zing.

Both have wide-ranging professional experience, alternating between large (air quotes) prestigious corporations, scrappy startups, and educational institutions.

In addition to these separated-at-birth similarities between Pete and John, there are also several significant-yet-complementary differences between them. Distilled to its essence, I’d say it like this: Cain is a Puzzler, Binggeser is a Tinkerer. And, sure, in the Venn diagram between Puzzler and Tinkerer, there is substantial overlap - each having tendencies and qualities in common. But what I’d call their primary affinity, the predominant, inciting orientation that defines them as makers and thinkers, would be this: John wants to figure out how humans function, their rituals and routines, beliefs and convictions, processes and procedures, as relates to the material world and our place/role within it. Pete wants to figure out how everything works - mechanical objects, workflow systems, virtual environments - all of it. If Cain wants to understand how we think things, Binggeser wants to understand how we do stuff and make things.

John has said of Pete: “He’s the guy who, after an afternoon’s worth of brainstorming, can go off and figure out what to make that will address the problem.” John, in my observation, is the guy who helps deepen and sharpen the brainstorming session - asking the probing questions that help identify the nature of the problem.

These prevailing sensibilities complement each other phenomenally well - Pete and John riff off each other in conversation like twins - that almost uncanny shorthand of communication, each with a deep grasp of where the other is headed and how to accelerate and amplify any idea they’re collaborating on. Each has a sophisticated spatial sense, a deep appreciation of materials, a deft capacity for observing thought, method, pace, flow, and rhythm in the way people perform tasks, reason through things, and design processes. Both are able to take in and synthesize lots of information and impressions to arrive at elegant interpretive snapshots of things, identify bottlenecks, and devise solutions. Watching John and Pete collaborate is like witnessing two guys in high-level discussion of a complex topic - building an energy-efficient house in a desert, say, or inefficiencies of process within an insurance conglomerate - while, without looking, they solve a Rubik’s Cube they’re tossing back and forth to each other.

Each has a way of seizing upon the other’s insights and building upon, modifying, expanding, intensifying, and tweaking them - each improves and broadens the thinking of the other.

They do NOT, however, finish each other’s sentences. That’s not rapport. That’s just interrupting, which is rude. Plus, it’s one of the biggest cliches there is in describing compatibility. So it’s beneath us all to resort to it. Watching them is like a bromance, without any of the maudlin, forced-feeling sentiment.

I interviewed each of them to learn more about how they came to converge on Large Language Models as the arena to tussle with, about the circuitous-seeming paths each has taken to get here, and what they hope ZeroWidth will be able to contribute to the field. The results of those interviews will appear in our next two posts.

Of Cuttlefish, Steam Power, and Killbots

2024-11-07T00:00:00Z

The human brain - three or so glorious pounds of electrified jelly - is the centralized source of everything we call thought, from the tiny and localized (thrusting our hand out, feeling droplets, concluding that it is raining) to the vast and conceptual (things we cannot directly observe like the movement of galaxies.) This morphology - a single, central interpretive hub reading the signals traveling up and down the branches of a nervous system - makes intuitive, visceral sense to us. Our only apparatus for conceiving of anything is trapped inside one, after all.

It’s likewise sensible and “natural-feeling” to ascribe such a system to other creatures. When we gaze into the face of our dog (or, to lesser degree, our cat,) behind those warm eyes we infer a kindred and similarly mammalian means of forming thought - a globe of bone protecting an oblong of electrified jelly that sets her tail wagging. Your dog sucks at understanding particle physics, probably, but the cognition system inside her more or less resembles our own. It is easy to anthropomorphize her - to ascribe to her inner life attributes and capacities that resemble your own, and even though her little thoughts differ from our own, we can make sense of the pathways they travel.

This tendency comes less naturally when we look upon an octopus. Everything about the cephalopod is alien to us, the way they are put together and the ways they interact with the world defy our attempts to overlay the sensical map of our own consciousness upon them. It’s possible, certainly, to discern intelligence when looking into the weird, weird eyes of a squid or a nautilus or a cuttlefish, but it’s an opaque form of intelligence, one that does not make the same kind of intuitive sense to us as our dog’s. Study after study demonstrates that cephalopods are extraordinarily smart - able to solve novel problems with astonishing speed, freeing themselves from inside sealed jars, for instance - a famous recent example: Inky, the Common New Zealand octopus who escaped from his aquarium in 2016, busting out of his tank, crawling along the floor in the open air, and slopping down into a drainpipe into the harbor outside the building.

One of the deepest ways the cephalopod is foreign to us is that they’re multi-brained. The centrality of cognition that seems so natural to us does not pertain to an octopus or a squid - they have brains in each of their limbs, as well as the one in their head - each limb has autonomy, has volition, can make decisions independent of the animal’s other constituent parts - its brains may coordinate and collaborate, or each may function independently. Which is a profoundly weird thing to try to imagine - it’d be like if you had a brain in each hand and foot that could choose whether or not to listen to the brain in your head, like your head brain might tell your hand brain: “Grab that outcropping - let’s finish climbing this rock wall,” and your hand brain might be like “No can do - I’m dealing blackjack,” and your foot brain might be like “Do you MIND? I’m TRYING to read down here,” or whatever. But then, when all your brains feel your body starting to fall, and everybody puts their own projects aside and pitches in and you climb the rock wall.

A Large Language Model (LLM), when utilized well and fully, is like a many-brained cephalopod—a collective of intelligences capable of independent decision-making, alternating their focus between individual side quests and shared objectives. Just as each arm of an octopus can explore and interact with its environment independently yet still serve the needs of the whole organism, an LLM processes information in ways that are BOTH distributed AND cohesive.

Alarmists and catastrophizers believe we are at the dawn of Skynet. They argue that the explosion of AI capabilities will inevitably lead to an adversary smarter and swifter than we are, one which will conclude that humanity is its only real competition and seek to destroy us. However, they fail to grasp the nature of the brain in question—it's an arm-brain; it does arm stuff. In other words, AI operates within a specialized scope, executing the tasks it's designed for (tasks, in other words, that we assign to it) rather than devising grand schemes against humanity.

An LLM processes information in distributed yet cohesive ways, utilizing its neural network architecture to generate responses based on learned patterns. However, it's essential to recognize that while the model can produce a wide range of outputs, it operates within the boundaries of its training data and does not possess an infinite repository of knowledge or infinite capacity for creativity.

The fallacy of centrality, the mistaken if organic-seeming belief that a centralized system is inherently superior for organizing and processing information or tasks, has a couple of readily foreseeable pitfalls as a consequence of cleaving to it too closely. Since this “natural” inclination readily congeals into a preference, then crystalizes into a*[n imagined]* requirement, we impose a similar “body plan” onto larger systems - a central brain does the thinking/reasoning/planning, subordinate components carry out the tasks. As the Industrial Age displaced the agrarian model humans had lived under for millennia, there was a practical advantage to centralizing - bring the labor to the machines in order to do the work; a textile mill is a gathering of looms that need people to run them, so it made sense to have those looms in a fixed location and bring the human operators to them, not vice versa. 20th century corporations were bound by similar consolidating impulses: when the flow of information must be input by hand onto physical paper, it made sense to have a typing pool in a single office space, or when the thousands of pieces that combine to make an automobile need assembling, it was logistically sensible for the hundreds of people required to do so on a factory floor under one roof.

This centralized manner of doing things contrasts sharply with much of today’s work - the multiple forms of cognition required to build a mobile app, say, or an operating system, an interface or a videogame, are not bound by the same rules of geographical centrality. Even work that produces a physical artifact - a smartphone, for example - has many layers of thought and decision making that are not involved in the form and materials of the ultimate product. We have already begun our transition into distributed cognition - if a team in Taipei is writing code for a tablet that’s to be assembled in Shenzhen, say, or the semiconductors made in a plant in Dresden that are to power components designed in Rotterdam for an electric bike to be built in Glasgow - there is already precedent for the multi-brained cephalopod approach to solving complex problems. It’s quite common, now, for each of us to enjoy the fruits of this kind of far flung thought, and likely as not we’re daily participants in some form of it in our own work - a graphic designer tweaks the arrangement of text written by someone else far away, a developer writes code to prototype a concept devised by a team in a different division of a company, etc. We collaborate, co-thinking about the solutions to problems, asynchronously and across distances all the time.

BIG HONKIN’ ASTERISK (the only punctuation so aptly many-limbed) - we are NOT suggesting that the rise of AI means there will be autonomous brains everywhere. The LLM is NOT its own brain. The LLM permits ready ACCESS to bodies of knowledge, it does not GENERATE novel thought.

The effective embrace of distributed cognition allows us to harness the collective intelligence of diverse minds, tools, and approaches, paving the way for innovative solutions to complex problems. AI tools like LLMs facilitate this distributed cognition by providing access to information of every description and aiding collaboration, but they are tools that augment human thought, they do not and cannot replace it.

As a company, ZeroWidth endeavors to both sharpen the tools in the LLM shed, and to smarten our own thinking about how these tools can help us. These intertwined, interdependent threads of thought will animate and inform the ongoing inquiry of this newsletter. We believe that the way forward must be propelled by a clear-eyed optimism - optimism for the soon-to-be astonishing ways LLMs can help us, clear-eyed about the capabilities and parameters of the technology, and about the skill and intent of the humans deploying it. By embracing this collaborative approach, ZeroWidth seeks to remain at the forefront of innovation, ready to tackle the complex challenges of today and tomorrow.