Jev's New Idea Is the Taxonomy Job I Built for AT&T

In 1999, one of my clients at Verity was AT&T, and one of the things they had me build was a moderation taxonomy: every piece of profanity, every slur, every way a customer or a rep might be vile to each other in text, organized into a hierarchy a machine could match against. There's something believable about a telecom commissioning exactly that list. I built it anyway, category by category, the way the job was done then.

That was not unusual work. It was, for a long stretch of my career, largely what the word meant. You sat with the people who understood a domain, you wrote down the categories that mattered inside it, and you built the queries and rules that would catch each one. Verity called its version TopicSet: hierarchical boolean and proximity queries, hand-authored into a tree. Autonomy called its version something gentler, Bayesian inference over a concept engine, but the labor underneath rhymed: you fed it labeled examples of a category and it learned to recognize more of the same. Either way, a person decided what the categories were, and a person tuned the boundary until it held.

The Tay Fiasco

Microsoft found out what happens without a system like that in less than a day. In March 2016, it put a chatbot named Tay on Twitter, built to sound like a teenager and to pick up its voice from whoever it talked to. Nothing stood between its learning loop and what came out of its mouth: no fixed set of categories, no confidence threshold, nothing checking the output before it posted. A crowd of bored, coordinated users spent a few hours feeding it exactly the material AT&T's list was built to catch, and Tay obliged, cheerfully, in public, in its own voice. Microsoft pulled it down inside sixteen hours.

Tay predates the transformer architecture by well over a year, so the failure it walked into was never a language-model problem specifically. It was a nothing's-checking-the-output problem, the precise gap a fixed-category, confidence-scored layer sitting in front of a generator exists to close. AT&T wasn't buying novelty when it paid for that taxonomy. It was buying the thing Tay shipped without.

The Sentiment Neuron

In April 2017, three OpenAI researchers, Alec Radford, Rafal Jozefowicz, and Ilya Sutskever, trained a byte-level LSTM to do nothing more ambitious than predict the next character of Amazon review text, then tested what it had learned against a standard, labeled sentiment benchmark. That part was deliberate; that's how you check whether an unsupervised representation is any good. What wasn't expected was where the signal turned out to live: not spread across the model's thousands of dimensions, but concentrated almost entirely in one single unit. That one unit alone matched state-of-the-art accuracy on the benchmark using only a handful of labeled examples, and it was steerable directly: clamp its value positive or negative, and the model's own generated reviews follow.

My hand-built taxonomies started to look like something else after that paper. Verity's queries were built rule by rule, one category at a time, by someone who had decided in advance what the categories were. The sentiment unit was never designed. It condensed on its own, inside a model whose actual training objective, predicting the next character of a review, had nothing to do with sentiment at all. One took a room of people months. The other fell out of an architecture nobody pointed at feelings in the first place.

A transformer architecture arrived that same year, and by the time ChatGPT put a chat window in front of hundreds of millions of people five years later, "artificial intelligence" had narrowed in everyday use to mean roughly one kind of system: the one that talks. The moderation taxonomies kept running. The classification jobs kept getting built, inside companies that will never write a blog post about them. They stopped being what anyone pictured when they heard the word.

The Termite Stick

In 1960, Jane Goodall watched a chimpanzee at Gombe strip the leaves off a twig and use it to fish termites out of a mound. Tool use had been one of the working definitions of what separated humans from every other animal, and Louis Leakey's reply, still quoted, was that now science would have to redefine tool, redefine man, or accept chimpanzees as human. The definition didn't hold. It got rewritten around the chimp instead.

The same rewrite keeps happening to machines. A calculator did arithmetic, and arithmetic stopped counting as a sign of a mind. A program beat the world chess champion, and chess stopped counting. Go went the same way, then conversation. Each time, the thing that changed definition was never the machine doing the task. It was the boundary drawn around what "really" required a mind. A system that reaches for a tool instead of freehanding the answer sits on the newer side of that boundary, the side Goodall's chimp is on, not the side a fluent chatbot claimed for itself. Not generating. Reaching.

The Jev Pitch

A company called TypeSafe AI is betting money on exactly that reach. Its product, Jev, launched this September with forty million dollars behind it, and it does not write sentences. Hand it a block of state and a question with a fixed set of possible answers, and it returns one of them with a calibrated confidence score attached, in under half a second, for a fraction of a cent. No prose to parse, no chance of the shape coming back wrong, because the shape was decided before the query ever ran. TypeSafe calls the category System One Models, after Kahneman, and named the product itself after William Stanley Jevons, the economist whose paradox says that making a resource more efficient increases how much of it gets used, not less.

Strip the branding and describe what it does in plain terms: fixed categories, a confidence score on every answer, chained into larger decision flows inside somebody else's code. That is my old job. Verity's topic trees, Autonomy's Bayesian classifiers, AT&T's moderation list with its threshold for when a human needed to take a look, MYCIN's certainty factors from the 1970s before any of us got there. Same architecture, same demand: a closed answer space with a stated number attached to each answer instead of a fluent guess with none.

The Instrument, Not the Answer

None of this makes the generative decade a detour back to where the field already stood. It shows what a generative model is best used to build. TypeSafe trains Jev with something it calls Reinforcement Learning for Calibrated Decisions, and outside observers suspect a large language model sits somewhere underneath the finished product, doing the labor a room of consultants used to do: reading examples, inferring the categories, tuning the boundary. If I wanted to build a moderation taxonomy today, I would not open a blank spreadsheet and start listing categories by hand. I'd ask a model to draft it, the same way I ask one to draft a Python script or a spreadsheet formula, because that has always been the correct division of labor: the model writes the instrument, and the instrument does the job.

Keep that distinction for the next System One Model, or whatever gets named after the next dead economist, that shows up with its own funding round. Ask what it's being used as. A model asked to be the classifier, freehanding a judgment call in prose every time it runs, is doing the thing generative systems are bad at, at cost, with no guarantee attached. A model asked to build the classifier once, so that something fast, typed, and calibrated can run forever after, is doing the thing they've been good at from the start. AT&T never cared whether the query in front of it was hand-typed Verity syntax or a language model's best guess at the edge of a bad word. It cared whether the machine got it right and said how sure it was. That part of the job never left the building. Only the tool that builds it did.