Why Is Jev, a Model That Only Answers Multiple Choice, Everywhere? Three Questions from TechWave EP155

Contents

1. What this episode is about
2. The main points
3. 1. Jev does one thing: it spreads probability across the options you give it
4. 2. It is 200x faster because it runs once
5. 3. Any language model can be turned into Jev; the moat is the training recipe
6. 4. “Zero hallucinations” rules out one kind of error
7. 5. Computer use: over 100x cheaper, only about 3x faster
8. 6. Harry’s own test: a payment reminder one day past due
9. 7. Jev speeds up an old road: model routing
10. Going further
11. When a launch says “200x faster, 400x cheaper,” how much should I believe?
12. Every week something is “going to change the world.” How do I tell a new road from an old road getting faster?
13. Worth a look
14. One thing to take away

Things gather by kind and divide by group, and from this, fortune and misfortune arise.

—— The Book of Changes, Great Appendix, Part I (pre-Qin; translated by the author)

What this episode is about

For the past week, one name has been all over AI feeds: Jev. In TechWave EP155 (2026-09-21), host Harry sits down to take it apart. Jev comes from TypeSafe AI, a company founded in 2024 that didn’t ship a single product in its first two years, and when it finally did, the launch post led with three claims: zero hallucinations, 200x faster, and 400x cheaper.

The CEO, Diogo Almeida, has quite a résumé. Back when people were turning models that could only complete text into chatbots like ChatGPT, he was one of the lead authors on that paper. He also likes to turn up at events in a bright pink puffer coat, which Harry says makes him look like Doflamingo from One Piece. But Harry puts his cards on the table right at the start: there’s no free lunch, so to pull off those three things, the model has to give something up. The whole episode is a hunt for what it gave up and what it got in return.

The main points

1. Jev does one thing: it spreads probability across the options you give it

Under the hood Jev is a large language model too, but it doesn’t write text, draw pictures, or write code. You hand it a situation, a question, and a few options, and all it gives back is how likely each option is. Harry’s example is a customer support message: someone writes in, “I’d like to know where my package is,” the options are billing, technical, and shipping, and Jev comes back with shipping at 92%.

TypeSafe calls it a “System 1 model,” a name borrowed from the fast, intuitive brain in Thinking, Fast and Slow, the part that sees a question and just reacts without mulling it over.

Harry thinks a more accurate label is “a general-purpose classification model built on an LLM,” and to me “general-purpose” is the part that matters most. It used to be that a bank wanting to predict whether customers would default had to train a separate model on its own historical data, and that model could do exactly one thing. Jev can take a different question without being retrained. But that’s also where its limit is: a question like default risk needs the bank’s internal data before any intuition can form, and that’s something Jev can’t pick up from the web.

2. It is 200x faster because it runs once

When a normal model writes a 2,000-word answer, it spits it out one word at a time, and every word means running the whole model again, so that’s 2,000 passes in total. Jev answers a question in a single pass, which works out to about the same compute a normal model spends on its very first word.

The CEO laid out the core trick on Hacker News, the tech forum: a model already calculates the odds of “which word comes next,” and Jev simply switches that to “which option does this question get.”

Harry also works backward from the API limits. Each question can have at most 255 options, which happens to be 2 to the 8th power minus 1. A situation plus one question can’t go over 32K tokens (tokens are the units a model uses to measure text length), and he guesses that’s how much the model can read in one go. The company also says multiple questions get split up and run at the same time, so piling on more questions doesn’t make you wait much longer.

3. Any language model can be turned into Jev; the moat is the training recipe

For me this was the most counterintuitive part of the episode. Within a few days, people in the community had built an open-source version on Qwen3.5-4B, a model small enough to run on a laptop. The method is simple: label the options A and B in the prompt, and when the model is about to produce its first word, look only at how much it scores “A” and “B,” then run those through softmax (a formula that turns scores into probabilities) so they add up to 100%.

It needs no extra training, and on easy and medium questions it lands close to Jev, without losing on speed or cost either. Where it falls behind is on hard questions, and on how accurate its probabilities are.

That gap comes from TypeSafe’s own post-training method, RLCD, short for Reinforcement Learning for Calibrated Decisions, which means the model has to pick the right answer and also give probabilities that match reality. Harry’s example makes it easy to picture: a Google review says “this restaurant is insane.” Most people who write that are complaining, but a few mean it as praise, as in “insane, how did they make a burger this good.” So the ideal answer is 80% negative and 20% positive.

Jev’s architecture does have a few tweaks, but the CEO himself says those aren’t what matters. The valuable part, the recipe and the training data, is private, so outsiders only get to use it through the API.

4. “Zero hallucinations” rules out one kind of error

Jev only spreads probability across the options you give it, so if you give it five, it won’t make up a sixth, and it won’t misspell one of them either. That kind of mistake is called a type error, and Harry guesses that’s where the name TypeSafe comes from. But it can still pick the wrong answer with total confidence, and if you count wrong answers as hallucinations, then Jev hallucinates too.

Put plainly, the marketing took a very narrow definition and sold it as a very broad promise.

5. Computer use: over 100x cheaper, only about 3x faster

Someone online used Jev to put together a system that operates a computer on its own, and pitted it against Claude Opus. Both got the task done, and Jev’s cost was more than 100 times lower, yet it was only about three times faster.

Why only three times? Because Jev can’t see the screen. It has to read the page’s raw structure and the accessibility tags written for blind users’ screen readers, then bolt on a separate OCR model (software that reads text from images), and that’s where most of the time goes. On sites with poorly built tags (Harry names Notion), it’s left leaning on OCR alone. If a button is just an icon, it can’t click it, and fine movements like dragging a slider are out of reach too.

Even if the next Jev gets vision, that won’t solve it. A normal model can just name the screen coordinates it wants to click, so any of millions of pixels is fair game, while Jev tops out at 255 options per question, which means the page has to be boiled down to a list of options first.

Harry holds this up against his own experience. He had GPT-6 Astra book a restaurant for him, and the site’s CAPTCHA showed three cartoon characters bouncing around, asking you to pick the one that jumped highest. Astra took a single screenshot the first time and got it wrong, so it thought for a moment, switched to taking several screenshots in a row, and got through. A task like that is one Jev can’t even get its hands on.

6. Harry’s own test: a payment reminder one day past due

Harry also gave Jev real work emails from his own inbox to sort, and it got most of them right. The one that stuck with him was a payment reminder for an event whose deadline had passed a day earlier. Some people would figure it’s already missed, so there’s no rush, but Harry would put it at the top, because it was less than 24 hours late and the organizer might still be willing to bend. In the end GPT-6 answered “cannot determine,” while Jev picked the most urgent level, at 47%.

Harry loved that 47%. It’s a question reasonable people could argue about, and the model didn’t put on a show of being sure.

7. Jev speeds up an old road: model routing

Harry’s closing take is that Jev speeds up a trend that was already happening, called model routing: you put several models of different strength in one system and hand out work based on how hard each task is. That road has been slow going, because workflows change too fast. A process you set up at the start of the year gets swapped out by midyear once the models improve, so the routing rules never get a chance to settle. OpenAI also tried letting its system pick the model for users automatically, got a ton of flak, and pulled it back. Harry thinks what people were really mad about was how bad that router was, and that with a good enough router, nobody would want to pick by hand.

Jev lowers the bar for all of this. You can use it as the router directly, or let it take over simple judgments itself, like narrowing 5,000 résumés down to 200 before handing them to a large model for a close read.

So most people will never see Jev directly. They’ll just notice the AI assistant in their hands getting faster and cheaper.

Going further

When a launch says “200x faster, 400x cheaper,” how much should I believe?

Every product launch and every new company brings a bigger number, and each one reads like a revolution. I often don’t know how much to believe. After this episode I wrote down three questions.

First: what did it give up? Jev’s 200x has a physical reason behind it, since it only runs once, and the price is easy to name too: it can only answer multiple choice. When a big number comes with no stated price, I mark it down.

Second: how much of the number survives inside the full system? Computer use was only about three times faster, because the time went into reading page structure and running OCR. The slowest step sets the speed of the whole line. It is the same question as finding the bottleneck layer in investing: if one part gets 100 times faster, how much faster the finished product gets depends on which step is now the slowest.

Third: how much can others reach with free tools? Qwen 4B already comes close to Jev on easy questions, so raw speed will not keep competitors out. What keeps them out is the gap on hard questions and on how accurate its probabilities are. When I read about a company, I ask a similar thing: how much would it cost a rival to copy its most impressive number? Whatever can be copied will not hold a premium for long. I used to get excited the moment I saw “10x.” Now I keep these three questions next to it.

Every week something is “going to change the world.” How do I tell a new road from an old road getting faster?

I am afraid of missing out, and just as afraid of chasing the wrong thing. Harry’s approach is to put Jev back on a road already being traveled. The nice thing about doing that is you get a known sticking point to measure the impact against. Model routing has spread slowly because workflows are unstable, and Jev doesn’t fix that; what it lowers is the cost of building the router.

Following that logic, the first to benefit are people whose workflows are already fixed and high-volume, like résumé screening or support ticket triage. Harry uses it to sort emails and saves a few seconds each time; he says it doesn’t change much for him. Three days of hype is noise. Cheaper simple judgments are structure.

When I hear about something new, I ask three things. Which existing road does it speed up? Where is that road stuck? Did it move the obstacle? I also write down in advance what would prove me wrong. Harry guesses the next step is large models calling Jev on their own in the middle of a task. If that happens, routing moves from something engineers set up by hand to something the model decides for itself. That would be a new road, and I would have to change my read.

Worth a look

• TechWave EP155 (2026-09-21), “What’s All the Hype About Jev? Are AI Agents About to Change?”
• Daniel Kahneman, Thinking, Fast and Slow: where System 1 and System 2 come from
• TypeSafe AI’s official description of Jev and its API documentation
• Hacker News threads where TypeSafe’s CEO replied

One thing to take away

A judgment needs a confidence level attached before you can check it later. Jev’s 47% reminded me that when I make a call, I usually just say “yes” or “no,” so looking back later, I can’t tell whether I got lucky or read it right.

One thing I tried: each day, pick three small things that will be settled soon, and write a percentage for each. For example, “this meeting ends on time, 60%,” “it rains tonight, 30%,” “they reply to that email today, 80%.” Check the answers the next day. After two weeks, pull out every entry you marked 80% and see whether eight out of ten came true. The first time I counted, my 80% calls had come true only half the time. Now my hand pauses every time I’m about to write a number.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.