An ex-OpenAI researcher just deleted language from the LLM...
FFireship@Fireship · 4.3M subscribers
- Views
- 3M
- 2,999,437
- Outlier
- ×3.0
- Median 1M
- Length
- 5:27
- Published
- Sep 21
- Monday
- Comments
- 1.3K
- 31.7K likes
Script
How it's written
- Words
- 1,086
- Words a minute
- 199
- Reading time
- 5 min
- Segments
- 4
The hook
0:00–0:25Last week, everything about AI changed forever. I realize everything about AI changes forever almost every week, but this time AI changed forever more than usual because one of the OpenAI researchers behind the instruction-following work that eventually became ChatGPT just released the next big frontier model after 2 years in stealth development. A model that can't talk, a model that can't write code, a model that can't write your college essays, and a model that will never tell you you're absolutely
Structure
Last week, everything about AI changed forever. I realize everything about AI changes forever almost every week, but this time AI changed forever more than usual because one of the OpenAI researchers behind the instruction-following work that eventually became ChatGPT just released the next big frontier model after 2 years in stealth development. A model that can't talk, a model that can't write code, a model that can't write your college essays, and a model that will never tell you you're absolutely right. Its name is Jev. My name is Jev. And this is a huge deal because large language models have one fatal flaw. They won't shut the hell up. You give Fable or Astra a simple instruction like return true or false, and it'll discover a third option after thinking for 4,000 tokens and then charge your credit card 11 cents. Jev fixed this problem with a radical solution. It deleted language from the large language model, and the result is a new type of classifier that's 200 times faster, 400 times cheaper, with free output tokens and zero hallucinations. Sounds too good to be true. So, in today's video, we'll take a look at Jev's code, its trust-me-bro benchmarks, and the dude who says he built an open-source Jev over a year ago. It is September 21st, 2026, and you're watching The Code Report. The big AI duopoly is literally shaking right now because Jev is a cheaper, faster way to solve basically any AI problem that requires a quick gut-instinct decision.
Watch from 0:00It's freight. But, the first thing you need to know is that Jev was created by an ex-OpenAI researcher, Diogo Almeida, and his company TypeSafe AI, which just raised $40 million. But, the company name is the first clue to what Jev really is. Like a regular large language model, you send it a question and some context, like a bunch of unstructured text. However, it differs because it behaves more like a type-safe programming language like TypeScript. The question you send to the model is a strongly typed question that must return a specific shape, one of three shapes actually, a choice, a score, and a null, which is basically just a yes or no. Its schema matching is guaranteed and a type error would be mathematically impossible to produce. They call Jev a system one model, which is a name that comes from Daniel Kahneman's Thinking Fast and Slow. A system one model is fast and goes from gut instinct, while a system two model is slow and deliberate, like these old antique reasoning models like GPT-6 and Claude Fable that burn 40,000 tokens to name a variable. But the difference is huge for app developers like myself who want to integrate fast cheap AI into their applications. Like on Horse Tinder, we recently had an issue of some donkeys trying to use the app, which is strictly forbidden in the terms of service. Thanks to Jev, we implemented an AI moderation step that will insta-ban any account that is not a horse, which is accomplished by returning a null response to is this a horse? Not only is it extremely fast, if we are to believe these TMBBs, but more importantly, it's off the charts cheap, like 440 times cheaper than one of the big brand models. In fact, it's so fast
Watch from 1:19and cheap that you can even use it for real-time applications. Like developers are already using it to implement NPC behavior in video games. And this guy even used it to build the world's first real-time AI calculator. But just because the output is type safe, that doesn't mean it's always correct. And it's not even deterministic. Like you could send it the exact same question in the exact same context and get different results, just like any regular large language model. But to get an idea of the response quality, it returns something called the calibrated confidence number. The chat models are trained to please human readers, and humans love confidence, which is how we got models that are wrong with the confidence of Kanye. Jev gained its confidence through a technique called RLCD, or reinforcement learning for calibrated decisions. This means every response provides a confidence value, like say 60%, which means 60% of the time it's right every time. But the big question is how does Jev actually work? Well, nobody knows for sure because the CEO says the architecture is staying close to the chest with a paper possibly coming in the future, maybe. But Jev also has some doubters. Some people say it's no different than zero-shot classifiers of the past, but the company gives no credit to the original pioneers of this technique like Jin Yang who were building zero-shot classifiers over a decade ago. In addition, this guy claims his paper he released a year ago is the exact same thing as Jeb. And another developer already built OpenJeb which reproduces the entire interface by reading option probabilities off a frozen Qwen-4B model in a single forward pass. It requires no new training and can run on a 3090. And there's even a web GPU demo you can run in your browser
Watch from 2:51right now. It's an awesome time to be a developer, which is why you need to check out Mux, the sponsor of today's video. Their highly customizable API is by far the easiest way to add video features to your application without getting jump-scared by FFmpeg. We've used it for years to handle all the hosting and streaming for our courses, but it does a lot more than just infrastructure. When you upload a video to Mux, you automatically get transcripts, storyboards, thumbnails, and clips along with structured data about what's actually in the video. That powers Mux robots, which is their AI hosted workflows that can translate your audio into other languages, moderate content, and lots more without you needing to host a model or maintain a pipeline. You can automate all this with directives where you define a workflow once and it runs on every new upload. And you only pay for the jobs that actually run. Perplexity, Patreon, and many other prestigious companies all trust Mux and their free plan includes 10 videos and 100,000 delivery minutes per month with no credit card required. And you can get an extra $50 credit at the link below. This has been The Code Report. Thanks for watching and I will see you in the next one.
Watch from 4:21Full transcript
- 0:00
Last week, everything about AI changed forever. I realize everything about AI changes forever almost every week, but this time AI changed forever more than usual because one of the OpenAI researchers behind the instruction-following work that eventually became ChatGPT just released the next big frontier model after 2 years in stealth development. A model that can't talk, a model that can't
- 0:20
write code, a model that can't write your college essays, and a model that will never tell you you're absolutely right. Its name is Jev. My name is Jev. And this is a huge deal because large language models have one fatal flaw. They won't shut the hell up. You give Fable or Astra a simple instruction like return true or false, and it'll discover a third option after thinking for 4,000
- 0:40
tokens and then charge your credit card 11 cents. Jev fixed this problem with a radical solution. It deleted language from the large language model, and the result is a new type of classifier that's 200 times faster, 400 times cheaper, with free output tokens and zero hallucinations. Sounds too good to be true. So, in today's video, we'll take a look at Jev's code, its trust-me-bro benchmarks, and the dude
- 1:01
who says he built an open-source Jev over a year ago. It is September 21st, 2026, and you're watching The Code Report. The big AI duopoly is literally shaking right now because Jev is a cheaper, faster way to solve basically any AI problem that requires a quick gut-instinct decision. It's freight. But, the first thing you need to know is
- 1:22
that Jev was created by an ex-OpenAI researcher, Diogo Almeida, and his company TypeSafe AI, which just raised $40 million. But, the company name is the first clue to what Jev really is. Like a regular large language model, you send it a question and some context, like a bunch of unstructured text. However, it differs because it behaves more like a type-safe programming
- 1:43
language like TypeScript. The question you send to the model is a strongly typed question that must return a specific shape, one of three shapes actually, a choice, a score, and a null, which is basically just a yes or no. Its schema matching is guaranteed and a type error would be mathematically impossible to produce. They call Jev a system one model, which is a name that comes from
- 2:04
Daniel Kahneman's Thinking Fast and Slow. A system one model is fast and goes from gut instinct, while a system two model is slow and deliberate, like these old antique reasoning models like GPT-6 and Claude Fable that burn 40,000 tokens to name a variable. But the difference is huge for app developers like myself who want to integrate fast cheap AI into their applications. Like
- 2:24
on Horse Tinder, we recently had an issue of some donkeys trying to use the app, which is strictly forbidden in the terms of service. Thanks to Jev, we implemented an AI moderation step that will insta-ban any account that is not a horse, which is accomplished by returning a null response to is this a horse? Not only is it extremely fast, if we are to believe these TMBBs, but more importantly, it's off the charts cheap,
- 2:46
like 440 times cheaper than one of the big brand models. In fact, it's so fast and cheap that you can even use it for real-time applications. Like developers are already using it to implement NPC behavior in video games. And this guy even used it to build the world's first real-time AI calculator. But just because the output is type safe, that doesn't mean it's always correct. And it's not even deterministic. Like you
- 3:07
could send it the exact same question in the exact same context and get different results, just like any regular large language model. But to get an idea of the response quality, it returns something called the calibrated confidence number. The chat models are trained to please human readers, and humans love confidence, which is how we got models that are wrong with the confidence of Kanye. Jev gained its confidence through a technique called
- 3:28
RLCD, or reinforcement learning for calibrated decisions. This means every response provides a confidence value, like say 60%, which means 60% of the time it's right every time. But the big question is how does Jev actually work? Well, nobody knows for sure because the CEO says the architecture is staying close to the chest with a paper possibly coming in the future, maybe. But Jev
- 3:50
also has some doubters. Some people say it's no different than zero-shot classifiers of the past, but the company gives no credit to the original pioneers of this technique like Jin Yang who were building zero-shot classifiers over a decade ago. In addition, this guy claims his paper he released a year ago is the exact same thing as Jeb. And another developer already built OpenJeb which reproduces the entire interface by
- 4:11
reading option probabilities off a frozen Qwen-4B model in a single forward pass. It requires no new training and can run on a 3090. And there's even a web GPU demo you can run in your browser right now. It's an awesome time to be a developer, which is why you need to check out Mux, the sponsor of today's video. Their highly customizable API is by far the easiest way to add video features to your application without
- 4:32
getting jump-scared by FFmpeg. We've used it for years to handle all the hosting and streaming for our courses, but it does a lot more than just infrastructure. When you upload a video to Mux, you automatically get transcripts, storyboards, thumbnails, and clips along with structured data about what's actually in the video. That powers Mux robots, which is their AI hosted workflows that can translate your
- 4:54
audio into other languages, moderate content, and lots more without you needing to host a model or maintain a pipeline. You can automate all this with directives where you define a workflow once and it runs on every new upload. And you only pay for the jobs that actually run. Perplexity, Patreon, and many other prestigious companies all trust Mux and their free plan includes 10 videos and 100,000 delivery minutes
- 5:16
per month with no credit card required. And you can get an extra $50 credit at the link below. This has been The Code Report. Thanks for watching and I will see you in the next one.
Description
SponsorMux is the video API that translates audio and answers questions about your video. Get $50 in free credits: https://mux.com/fireship
Ex-OpenAI researcher Diogo Almeida spent two years in stealth building Jev, a "System 1" AI model that can't talk or write code but claims to be 200x faster, 400x cheaper, and hallucination-free. Let's dive in.
Want more Fireship?
🗞️ Newsletter: https://bytes.dev
🧠 Courses: https://fireship.dev
More from Fireship
×4.27:27Did OpenAI actually build AGI? GPT-6 Astra first look
Fireship4.2M views
×4.15:56Every app on Earth is now open source... whether it likes it or not
Fireship4.1M views
×3.45:45OpenAI's biggest math breakthrough is getting ugly...
Fireship3.4M views
×3.37:22Tragic mistake... Anthropic leaks Claude’s source code
Fireship3.3M views
