TypeSafe’s Jev does not write sentences. Given a situation and a list of questions, it returns a probability for each answer — one pass, no generated text — and Vercel reports it was adopted faster than any other model in AI Gateway history. John Berryman, who worked on Copilot at GitHub, argues OpenAI is well positioned to fast-follow it, and better positioned to fold classification into its existing models than TypeSafe is to defend it.

The thesis

  • OpenAI has used large language models as classifiers for years — it just never trained them for general classification or sold that as a product.
  • Jev’s trick, as Berryman reads it, is that a model already assigns a probability to every possible next word. For a yes/no question you look only at the probabilities for true and false, normalize them against each other, and you have an answer with a confidence number attached.
  • Do that once, in one forward pass, and you skip the expensive part of asking a model to write out prose and reason its way to a verdict.
  • If that is all Jev is, and if TypeSafe’s advantage is only in how they trained the data, then replicating it is a matter of skill, hardware and funding — three things OpenAI has.

Why he thinks OpenAI already does this

Berryman’s evidence is old and concrete. Look at the internal markup (ChatML) that OpenAI uses to organize a conversation, and the first thing the model predicts after the assistant tag is either a newline or to=function. — a one-token decision about whether to call a tool at all. The next tokens pick which tool. The tag that ends a message is itself a classifier reading “this is finished.”

He also tells an anecdote from his Copilot days, working with a raw GPT-4 API. Responses would not stop: “Let me know if you have any other questions. Have a nice day. Have a great week. Have a good time. Have a wonderful life,” until the token limit. The cause was a missing header value, so the model was not allowed to use its own end-of-message token — it had no internal ability to shut itself up. Same conclusion: single tokens have been doing small classification jobs for years.

What a built-in classifier would buy

His sketch is a model that can ask itself a calibrated question mid-reasoning, through a tag like <prediction> — and since the answer is written back in the same pass, there is no handoff to outside tooling. It never leaves the GPU. Uses he lists:

  • Checking, mid-reasoning-trace, whether it is actually done and what to do next, instead of talking itself to a stop.
  • Screening its own tool call before running it — his example is a live API key sitting in plaintext arguments.
  • Scanning a tool’s response for prompt injection, collapsed over time into a simple safety score.
  • Routing work between a bigger, smaller or equal model, trained toward whatever is cheapest without losing accuracy.
  • Extending the same trick to images and speech, where a voice agent could decide in real time to interrupt, escalate or keep listening.

The moat question

Berryman is not that interested in the architecture. He assumes Jev is a conventional language model, notes that most early clones are, and says he is rooting for TypeSafe. The only defensible moat he sees is the training data and the process that makes it calibrated. TypeSafe cofounder Diogo Almeida says the company considers itself “a data research lab” and that “100% of our data is synthetic (but not the type of crap that is just spit out from an LLM obviously).”

The catch he raises himself: none of it matters unless Jev is actually accurate, and accuracy is the hardest property to check. He says he has already found domains where Jev’s probabilities do not hold up, and Jev’s own documentation says it works best for fast, instinctive judgments — not math or multi-step reasoning. If the moat is real, TypeSafe might get acquired. If it is thin, the window closes fast.

What the thread adds

The 176-comment thread on Hacker News

  • amluto — the most substantive correction. Yes, the model emits probabilities, “but there are huge caveats”: what it is estimating is the chance of the next word in its training text, which “is not at all the same thing as the probability that Paris is France under any distribution that you care about.” And the probabilities attached to tool calls come from training that made the model useful, not from a measure of truth. dgellow’s gloss: “I find it slightly more helpful to say they generate plausibility.”
  • orbital-decay — every major AI lab already runs classifiers: safeguards inside the inference pipeline, plus data preparation, analysis and research work. Offering them on a public API “doesn’t always make business sense.” Their read of the fuss: “looks like people that are new to all this are discovering that classifiers exist.”
  • andy12_ — doubts OpenAI bothers. They are all-in on reasoning models trained with reinforcement learning, “and Jev-like models are the total opposite.” They also challenge the article’s premise, saying there is no proof Jev is a conventional language model and that a different text encoder would explain its price.
  • pushpendraw — the practical reason people care: “the real win with jev isnt beating a trained classifier on accuracy, its that you can change what you’re classifying by editing a prompt instead of retraining and redeploying a model.”
  • nzoschke — hands-on counterweight to the theory: classifying a lot of email, they found OpenAI’s small embedding model helpful but “initial tests with Jev are very promising and much more ergonomic.” Berryman replies in the thread, linking his 2025 post on the same logprob trick.
  • HarHarVeryFunny — the demand-side answer to “why would OpenAI bother”: much business automation “doesn’t need AGI, doesn’t want to pay for AGI if they don’t have to,” and would be better served by cheap calibrated classifications with structured output.
  • jackb4040 — dissent from the “labs will just copy it” consensus: OpenAI’s “track record is not great when it comes to jumping on these micro paradigm shifts,” having already lost a first-mover advantage to Anthropic.
  • boshalfoshal — the strongest skeptical case, from inside the thread: the labs may have skipped this because the market is small, not because the capability is hard. “In reality, none of these really matter.”
  • prodigycorp — an attack on the piece itself, calling it hard to read and the moat framing hollow: “Moat talk merely projects the illusion of being engaged but, much more often than not, it’s hollow engagement.” jrochkind1 replies with an AI-detector guess of roughly 20% AI-written, and a good line about the direction of influence: “Eventually humans are going to start writing like AI if we read enough of it.”

The question the thread keeps asking

Does Jev actually work? Four commenters raise it from different angles without converging. deepsquirrelnet cannot see a validated market: “there’s no independent benchmarks that prove anything. To me that’s doesn’t pass the smell test.” EagnaIonat reports testing Jev and a competitor, and says “they both suffer from the same issues as LLMs that stop them being useful beyond limited classifications” — adding that they see no advantage over an ordinary machine-learning classifier. prometheus1992 points at free local alternatives and concludes “there is no sensible business proposition in Jev,” while floating the theory that the attention is paid for. And the article itself supplies the same doubt from the other direction: Berryman says he has personally found domains where Jev’s probabilities do not hold up.

A note on reading comments as evidence: HN handles are pseudonymous and the site publishes no per-comment scores, so the ordering here is HN’s own ranking, not a vote. This is a slice of the thread, and claims about who is funding what are quoted as theories, not findings.