WhoKnows.
← All briefings
TECHACTION1 story

Daily Briefing — September 23, 2026

WhoKnows · Daily intelligence briefing

01

Who is Jev? Is this another DeepSeek moment?

TypeSafe →
What to do

Last week I saw a model called Jev come out. My first thought was "here's another faster, cheaper GPT substitute" — it arrived claiming to be roughly 200x faster and 400x cheaper than comparable large models on classification tasks, with free output. I looked into it, and that's not what it is. It doesn't even talk to you.

I've pulled together what I've read over the past few days, and there's a lot in here worth looking at. It comes from a company called TypeSafe AI, which opened early access on 15 September.

It doesn't produce text, it produces judgements

The AI we all understand right now works like this: you say something, it writes something back. ChatGPT, Claude, Gemini are all the same shape.

Jev generates no text at all.

You give it two things: a piece of program state (context, as I read it — the full text of a support ticket, say), and a set of questions defined in advance. It answers all of them in a single call and returns JSON with typed values, not prose you then have to parse. Response time is 70 to 500 milliseconds.

TypeSafe calls this kind of model a System One Model, borrowing Kahneman's System 1 / System 2 split. System 2 is slow, deliberate, effortful. System 1 is fast and intuitive, the answer you have the moment you look. The first time WhoKnows mentioned System 1 was in the interview with Professor Liu Jia, who said we don't need to worry about being replaced by AI just yet, because System 1 problems haven't been solved. So are we a step closer now? My take: still a long way off. But worth writing down so we can watch it.

Their argument is that most of what software calls an "AI decision" is really "which category is this," "how urgent is this," "should this be allowed through." All System 1 work. And for three years we've been paying System 2 prices for System 1 service. Jev is their fix for that.

It only answers three kinds of question

Jev takes exactly three question types:

  • Choice — pick one from up to 255 options.
  • Score — a position on a scale with 2 to 10 levels.
  • Noul — a yes/no question, except what comes back isn't "yes" or "no", it's a probability between 0 and 1.

Input is text only. No images, audio or video. State plus all questions caps out at 64k tokens.

The training method TypeSafe calls Reinforcement Learning for Calibrated Decisions. What it optimises for isn't "do humans like this answer" but "is this probability honest." That's the deepest fork between Jev and mainstream LLMs, a difference in motive.

Today's large models have a well-known flaw: ask how confident they are and they'll say "95% sure," and then be wrong. RLHF trains for being liked, and confidence sounds more likeable than hesitation. Jev calibrates its decisions whether you like the number or not.

Pricing: 4.2 cents per million input tokens. Output is free.

What can it actually do

While everyone was marvelling at Jev's arrival and struggling to name a use for it, the community site started listing products people claim to have built with it, everything from real-time fitting rooms to simple Q&A. Worth a browse if you want ideas.

A few scenarios:

Support tickets. A ticket comes in, and you ask it a dozen questions at once: which team should this go to, how upset is the customer, are they asking for a refund, does policy cover it. TypeSafe's own cookbook says batching 13 questions into one call was 12.2x cheaper and 10x faster than asking them one at a time. Across a million tickets, a Jev-first approach comes to about US$6,480, versus roughly US$30,400 for LLM-only.

Classification and filing. 1kpapers.com used it to sort 1,018 papers into 24 topics. The summarising part, which has to use a model that can actually write, cost US$3.99. The classification cost 8 cents, at a median 256 ms per paper. In the past, a job like this was simply too expensive for anyone to bother with unless the payoff was unusually large.

Guardrails for agents. Every time an agent wants to call a tool, reviewing it with a second large model doubles both latency and cost. Jev scores each proposed action in under half a second, and because the probabilities are calibrated, you can set a different threshold per action depending on how dangerous it is. That's exactly how LangChain's AutoModeMiddleware uses it.

So who is Jev

The name is a nod to the 19th-century economist William Stanley Jevons.

Straight from the launch post: we named it after Jevons because we expect machine intelligence to follow a path similar to coal, where steam engines got more efficient and demand went up rather than down.

What Jevons observed was this. Steam engines kept getting more coal-efficient, everyone assumed coal consumption would fall, and total coal consumption rose sharply instead. Saving coal meant that uses which previously couldn't afford coal suddenly could. The efficiency gain didn't save the resource, it created new demand.

I wrote about this in Big Ideas 2026, Part 2 — have a look.

A company naming its product to tell you upfront that using it will not lower your bill. That's the most honest piece of marketing I've seen this year, and also the smartest.

Why it blew up overnight

Several things stacked up.

The pedigree. Founder Diogo Almeida worked on RLHF and InstructGPT at OpenAI. The company opened with a US$40M seed round led by DCVC.

The price. Claimed ~200x faster and ~400x cheaper than comparable large models on classification tasks, with free output.

The demos. Playing Doom in real time, playing Minecraft, running a driving sim, flying a simulated drone.

The ecosystem. LangChain, Vercel AI Gateway and Browser Use all shipped integrations within 48 hours, and open-source clones of the interface like openjev showed up within days.

And one more layer. "Fast structured output" pushes AI's useful range well beyond the chat box, and the framework itself is easier to program against than a conventional LLM. We do already get structured output from conventional LLMs via function calling, but generation is the LLM's main job and function calling is a layer on top of it, so you save no tokens and it runs slower. Jev meets that need more directly.

Reasons to hold off

All the benchmarks are self-run. Speed, cost, accuracy, all of it comes from TypeSafe, with no third-party reproduction so far. They also acknowledge the evaluation workflow was built by their own team.

The "accuracy" column has no ground truth. What they measure is how close Jev's output is to the average output of GPT-6 Astra and Claude Fable 5.1, not how close it is to the facts. TypeSafe notes in its own docs that this biases towards OpenAI and Anthropic. Put differently: if the large models are collectively wrong, Jev is wrong with them.

The moat may be shallow. Sean Goedecke's independent analysis argues most of the speed advantage can be approached with existing LLMs by prefilling and then constraining generation to about one token per question. Other labs wouldn't find that hard to copy.

There's a ceiling on intelligence. No test-time compute means its capability is roughly capped at the level of non-reasoning models. On TypeSafe's own eval, Jev scores 67.8%, level with GPT-5.6 Terra and Claude Sonnet 5, behind Opus 5 at 73.1%.

"Can't hallucinate" is the wrong phrasing. The accurate version: it cannot return a value outside the schema. It can absolutely return a value that's correctly formatted and wrong.

It's early access. Waitlisted, US-hosted, rate limits shifting whenever GPU capacity lands.

Jev isn't the only answer

A few days after Jev's launch, a company called Convai Innovations released Laya under Apache 2.0. The same three decision primitives, the same API shape, plus open weights.

It takes the opposite route. Jev is a closed API you send your data to. Laya is an open model of 322M to 421M parameters that you deploy on your own machines.

On several dimensions it genuinely wins. About 33 ms per query on a single GPU, faster than Jev's 70–500 ms. Free to run, since you host it. Over 100 languages supported. And it can live entirely inside your own network, which for teams that can't send customer data to US servers makes every other comparison irrelevant.

But there's a precondition, and it changes the nature of the whole thing. Laya's zero-shot performance is close to guessing. You have to fine-tune and calibrate it for every single task before it's usable.

That's the real dividing line. Jev works out of the box: define your questions and you're running. Laya gives you speed and ownership, and the price is a round of training for every new task — labelling, fine-tuning, calibration, validation. You're training it yourself.

Two more limits. Context is only 512 to 1024 tokens, which is brutal, against Jev's 64k. And accuracy starts dropping once you go past roughly 20 options, where Jev advertises 255.

What does this mean for me?

  1. Separate "I need it to think" from "I need it to sort." Open whichever AI tool you use most, and find one judgement you make every day. Does this email need a reply, is this request high priority, is this feedback a complaint or a suggestion. Then ask: does this need reasoning, or just categorising? If it's the second, the tool in your hand is almost certainly overspecced. Overspecced isn't only expensive, it's slow.

  2. When you see a multiple, ask who ran the test. With "X times faster and Y times cheaper than the competition," the first question is never how big the multiple is. It's who ran it, against what, and whether there was a right answer to check against. Jev and Laya are a good sample right now: both claim to be faster and cheaper, both sets of numbers are self-run, and the two of them are still arguing about who thought of it first. Self-run benchmarks deserve a lot of suspicion.

  3. Promises of saving money rarely end in saved money. Jevons already called this for energy prices and consumption. Once unit cost falls past a certain point, suppressed demand is released all at once and total spending goes up. Cheap enough, and we humans will simply invent new things to want.

  4. There's still a gap between "free" and "usable." Laya is open source and free, but near-guessing zero-shot, so you can't just pick it up and run. What it gives startup projects is an opening and a good idea to build on. Give it time. Something new built on Laya will be open-sourced.

Sources