Why Modern Creator?

Matthew Berman · YouTube

Jev is a decision engine, not a thinker, and it just lost a chess match to prove it

ChatGPT co-inventor Diogo Almeida ships a model that's 200x faster and effectively free, but Matthew Berman's own chess test shows it losing on skill, only surviving the clock.

Posted

September 18th

2 days ago

Duration

12:31

Format

Talking Head

hype

Views

301K

5.1K likes

Part of the collectionJev, explainedEvery Jev breakdown, synthesized into one page.

Read the playbook

Big Idea

The argument in one line.

Jev trades general reasoning for raw speed and cost, so it wins at narrow, high-volume decisions but loses outright when an actual game of skill like chess requires it to think rather than just answer fast.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You run a product that needs thousands of cheap, low-latency classification or routing decisions per second, like support ticket triage or ad-slot blocking.
You're evaluating whether a non-chat 'decision model' architecture belongs in an existing automation stack like Zapier.
You want to understand what a zero-hallucination claim actually means in practice, not just as marketing copy.

SKIP IF…

You're looking for a chat, coding, or general-reasoning model to replace ChatGPT or Claude, Jev explicitly isn't built for that.
You want a rigorous benchmark comparison, this video is a single creator's hands-on reaction to demos, not an independent evaluation.

TL;DR

The full version, fast.

Diogo Almeida, a ChatGPT co-inventor, spent two years building Jev at TypeSafe AI: a 'decision model' trained with a new method called RLCD instead of RLHF, up to 200x faster and 400x cheaper than chat models, with free output tokens. It isn't a chatbot. It takes a state or a prompt and returns one of a fixed set of decisions, thousands of times per second, and TypeSafe claims zero hallucinations. Demos show it running a Doom agent, winning wiki-race speed tests, sorting 150,000 Skittles, and controlling a simulated Tesla and a AI Town crowd in real time. But when tested head to head at chess, Jev only beat Fable 5.1 by running the opponent out of time, and was cleanly checkmated by GPT-6 Astra in 18 moves, undercutting the idea that it can substitute for real reasoning.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:52

01 · The reveal

Berman introduces Jev, a new model that is 'hundreds of times faster' than traditional LLMs, with free unlimited output tokens.

00:52 – 01:50

02 · Origin story: RLCD

Diogo Almeida's viral tweet explains he co-invented ChatGPT, then spent two years building Jev on a new training method, RLCD, instead of RLHF.

01:50 – 03:09

03 · Speed benchmarks: Doom and the wiki race

Jev plays Doom in real time and wins a wiki-race speed test against GPT-5.6 Terra, Claude Haiku 4.5, and Claude Sonnet 5 by several seconds.

03:09 – 04:12

04 · Decision engine demos: ticket routing

A support-ticket routing example shows Jev picking from fixed categories, urgency, and priority options in milliseconds instead of generating open-ended text.

04:12 – 05:09

05 · The zero-hallucination claim

Berman frames RLHF's human-feedback bias as the cause of hallucination and cites an on-screen benchmark showing Jev-1.0 at 0.0% tool-call error rate versus double-digit rates for GPT-5.6 variants.

05:09 – 06:32

06 · Sponsor: Zapier

Mid-roll ad pitching Zapier as the pipe that feeds emails and support requests into a fast, cheap model like Jev inside automated workflows.

06:32 – 07:28

07 · AI Town: 50 decisions in 0.6 seconds

A simulated town built in Astra/Codex has Jev-powered characters react to a broadcast prompt (a bakery fire sale, then a poisonous-snake warning) almost instantly.

07:28 – 08:46

08 · Sorting 150,000 Skittles

A chopstick-sorting demo has Jev decide, one piece at a time, which of five color buckets each of 150,000 Skittles belongs in, even at maximum speed.

08:46 – 09:45

09 · Where Jev falls short: setting up the chess test

Berman states plainly that Jev is not for chat, code, or interactive sessions, then introduces a third-party chess experiment as the real stress test of its limits.

09:45 – 10:37

10 · Chess results: won on the clock, lost on skill

Jev V13 beats Claude Fable 5.1 only because Fable ran out of time after 29 moves of being outplayed; Jev V13 is then cleanly checkmated by GPT-6 Astra in 18 moves.

10:37 – 12:31

11 · More builders: model routing, ad blocking, Tesla FSD, Melee

Riley Brown uses Jev as a model router, Kitsy ships an ad/slot blocker called Unclutter, Justin Schroeder rebuilds Tesla FSD decision-making in under an hour, and a teammate has Jev play Melee in real time.

Atomic Insights

Lines worth screenshotting.

Jev was built by Diogo Almeida, a co-inventor of ChatGPT, specifically because he believes superhuman chat models have not led to AGI.
Jev is trained with RLCD (reinforcement learning for calibrated decisions) instead of RLHF, which TypeSafe AI claims removes the human-feedback bias that causes hallucination.
Output tokens from Jev are free; input tokens cost fractions of a penny, making it dramatically cheaper than any chat-model API.
In a tool-call error rate benchmark shown on screen, Jev-1.0 scored 0.0% versus 22.1% for GPT-5.6 Sol and 14.3% for GPT-5.6 Luna.
Jev finished a 5-hop wiki race in about half a second, while Claude Haiku 4.5 and Claude Sonnet 5 each took four to five seconds.
TypeSafe's stated positioning, 'we're building prod, not god,' is a direct jab at Anthropic's AGI-focused messaging.
Jev is not a chat model: it can't hold a conversation, explain its reasoning, or write code from scratch.
Jev powered 50 independent AI Town character decisions in 0.6 seconds after a single broadcast prompt.
In a real-time Skittle-sorting demo, Jev made a fresh sorting decision for each of 150,000 pieces individually rather than batching them.
Jev beat Claude Fable 5.1 at 5+0 blitz chess only because Fable burned 6 to 15 seconds per move on analysis and ran out of time, not because Jev outplayed it.
Jev lost outright to GPT-6 Astra at chess, getting checkmated in 18 moves, which the video treats as proof Jev is not a substitute for genuine reasoning.
A separate builder, Riley Brown, is using Jev purely as a model router, deciding which downstream chat model a prompt should go to rather than answering it directly.
A Tesla Full Self-Driving simulation, built with Jev handling the driving decisions, was assembled by one developer in under an hour.

Takeaway

Speed and cost aren't the same thing as intelligence

WHAT TO LEARN

A model can be radically faster and cheaper than every chat LLM on the market and still lose outright the moment a task requires actual reasoning instead of narrow decisions.

01The reveal

A 'hundreds of times faster, free output tokens' claim is worth watching skeptically until you see what task it was actually measured on.

02Origin story: RLCD

The person building a chat-model alternative is often someone who helped build the original and became frustrated with its ceiling, which is worth weighing alongside the marketing.
A new training method name (RLCD vs RLHF) is a claim about what the model optimizes for, not proof that the optimization works outside the demos shown.

03Speed benchmarks: Doom and the wiki race

Raw latency wins, like a five-hop wiki race finishing in half a second, only matter for tasks where speed itself is the value, not tasks that need deliberation.

04Decision engine demos: ticket routing

Constraining a model's output to a fixed menu of choices, like support categories and priority levels, is itself what makes millisecond decisions possible and reliable.

05The zero-hallucination claim

A 0% tool-call error rate on a specific benchmark is a narrow, task-bound claim, not a general reliability score you can extrapolate to other work.

07AI Town: 50 decisions in 0.6 seconds

Making many independent, low-stakes decisions in parallel is the use case this architecture is actually built for, not single deep decisions.

08Sorting 150,000 Skittles

A demo that looks impressive at normal speed can look ordinary at high speed, which is itself informative about what's actually hard for the model versus what's just slow.

09Where Jev falls short: setting up the chess test

The creator's own disclosed limitations, like 'not for chat, not for coding,' are worth taking at face value before assuming a new model is a universal replacement.

10Chess results: won on the clock, lost on skill

A model winning on paper because of a technicality, like an opponent running out of time, is not the same as winning on skill, and that distinction matters before adopting a benchmark claim at face value.
Getting checkmated in 18 moves by a general chat model is direct evidence of a reasoning gap, not just a stylistic difference in how the two models play.

11More builders: model routing, ad blocking, Tesla FSD, Melee

When a builder pairs a fast decision model with an existing chat model as a router, they get the best of both: cheap triage at the front, real reasoning only where it is actually needed.
A working prototype built in under an hour, like a Tesla FSD simulation, demonstrates feasibility, not production-grade reliability.

Glossary

Terms worth knowing.

Decision model: A model architecture that takes a state or prompt plus a fixed menu of options and returns one choice, rather than generating open-ended chat or code.
RLCD (Reinforcement Learning for Calibrated Decisions): Jev's training method, positioned by its creator as an alternative to RLHF that optimizes for calibrated decisions instead of human-pleasing chat responses.
RLHF: Reinforcement learning from human feedback, the standard method used to align chat models like ChatGPT and Claude, which this video frames as a source of hallucination because humans themselves make mistakes.
Flagging: In chess, causing an opponent to lose on time rather than by being outplayed on the board, which is how Jev won its match against Fable 5.1.
Tool call error rate: A benchmark measuring how often a model incorrectly executes or misfires a function/tool call, used in the video to argue Jev is unusually reliable for narrow tasks.

Resources

Things they pointed at.

01:00productDiogo Almeida / TypeSafe AI (Jev)

05:09toolZapier

09:20linkAI/ML API chess benchmark thread

10:38productRiley Brown model router

11:02toolUnclutter (by Kitsy)

11:31productJustin Schroeder Tesla FSD rebuild

Quotables

Lines you could clip.

00:22

“That is not sped up. That is actually real time.”

clean disbelief hook, works as a cold open for any speed-demo clip→ TikTok hook↗ Tweet quote

04:05

“Their motto is we're building prod, not God, a direct shot at Anthropic.”

quotable rivalry line with a built-in pun→ newsletter pull-quote↗ Tweet quote

10:10

“Astra didn't bother winning on material, it mated Jev in 18 moves.”

the whole video's real verdict in one sentence→ IG reel cold open↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphoranalogystory

There's a new AI taking the internet by storm right now, and it is so fast. Watch this. That is not sped up.

That is actually real time. This is Jev. And this is the guy who just launched it.

He co -invented a little product called ChatGPT and now has released something that is completely different from the architecture of ChatGPT and made it hundreds of times faster. than traditional large language models. I mean, the speed is truly insane to watch, but there were trade offs.

And so I'm going to tell you about Jeff. I'm going to tell you what it's good for, what it's not as good for. And then I'm going to show you some incredible demos.

And here's the thing that I want you to keep in mind. Jeff is so efficient. They actually made it free.

You get unlimited output tokens. absolutely free. And the input tokens, what you actually prompt it with, they are fractions of a penny.

So this is substantially cheaper than anything else on the market. Let me tell you about it. Here's the tweet nearly at 30 million views.

After co -inventing ChatGPT, I kept asking myself, why have superhuman chat models not led to AGI? I've spent the last two years in stealth building a new way to train models, RLCD, that is different from RLHF, reinforcement learning. with human feedback.

This is reinforcement learning for calibrated decisions up to 200 times faster and up to 400 times cheaper when output tokens are so cheap they just made them free. This model is not a chat model, though. It is a decision model, but it is a generalized decision model.

You can give it any decision that you need to have made and it will make it. and it'll make thousands of them in seconds. And here's a benchmark.

This is TypeSafe over here, basically on par with Luna and Terra and Sonnet 5, above Opus 5, above Sol, but a fraction of a penny. And in fact, it is so fast, and I'm trying to convey how fast it is, it can play Doom in real time. It is making all of the decisions inside the loop of the actual Doom software.

It is looking at what's happening and making decisions of what to do in absolute real time. Here's another example of browser use.

And again, it is so fast. And what it's doing is clicking through the browser in something called a wiki race, basically trying to click from wiki page to wiki page and eventually finding something. Watch how fast it is.

So we have Jev in the top left, 5 .6 Tera top right, Haiku 4 .5 and Sonnet 5. in the bottom right so watch how fast it goes three two one and look at that it finished three hops in faster than you can even see it here's another race ready let's see if we can even see how fast it goes there it goes five hops in half of a second let's see how long the other ones took four seconds five seconds and five seconds so a fraction of the time.

Let me show you a few examples of how you can use Jev. And these are basic examples. And then I'm going to show you more complicated examples and incredible demos.

So you can kind of think of it as a decision engine. You can give it an input, whether it's a state or a bunch of information and have it make decisions. And not only one decision and not sequentially, but.

hundreds or thousands of decisions in parallel. Here's an example. So this is for support ticket routing.

I was charged twice and need this fixed today. They're on the pro plan and their account age is 420 days. Okay, so here are the questions.

What type of support request is this? Does this need urgent handling? and rate support priority, and we give it the options to choose from.

It looks at the input and decides the output. And all of this is happening again in like milliseconds. And they're saying or their motto is we're building prod, not God, a direct shot.

at Anthropic. There are a few other properties about Jev that make it incredibly special. So one of the problems with using reinforcement learning with human feedback is models are optimized for humans and humans make mistakes.

And so that leads to traditional large language models hallucinating, something we're all familiar with. Now, the rate of hallucination over the last three years has dropped significantly, but for some use cases, any hallucination is catastrophic.

Think about critical use cases where decisions are being made where lives are on the line. Healthcare, military targeting, even traffic. But with Jev, it is so reliable, they are claiming zero hallucinations.

And this type of speed is incredible, especially for businesses where they have to make thousands of decisions per second on a wide variety of things. And one of the ways to pipe all of those decisions to be made into a model like GEV is with the sponsor of today's video, Zapier.

Now, imagine this. With Zapier, you take all of your emails or you take all of your customer service requests and then you plug Jev into Zapier and now you're paying a fraction of the price and 100x, 200x the speed to make decisions within one of your Zapier workflows. Zapier allows you to connect over 9 ,000 different applications in different ways.

build entire automated workflows all with artificial intelligence and so you can easily plug in the obvious ones like cloud and cloud code and chat gpt you can do gmail and calendar and whatever your customer support software is and again now plug in jev to it pay a fraction of the price get speeds where your customers are going to absolutely love the response rate, and it's already used and trusted by the world's biggest companies like Nvidia and Shopify, Meta, Cursor, Samsung and so many more.

So go check it out. Huge fan. Click the link down below to let them know I sent you.

Now, let me keep telling you about Jeff. All right. So here's a little demo I made to show off Jeff's decision making.

So I actually used Astra in Codex to build this world because that is not what Jev is for. It's not necessarily for building code from scratch.

Now, there are aspects of the code building workflow that you can offload to Jev, but for this, I built it with Astra. And what you're seeing is a little town. And all of these little characters in the town are powered by artificial intelligence.

Specifically, they're powered by Jev. And so I can give them a prompt. they will all make a decision about how to react to the prompt in less than a second.

So let's watch. We have a fire sale at the bakery. Everything must go.

I'm going to click broadcast. So 0 .6 seconds, 50 different decisions about what each of these people are going to do. 39 of them decided to just keep doing what they're doing.

Six decided to investigate, four to join in, and one to warn others. Okay, that was a pretty benign prompt. But what if I did something more aggressive?

So everyone who doesn't go to the fountain will be bitten by a poisonous snake. Let's see what they decide. And here we go.

We can see almost all of them. I guess some of them are not afraid of poisonous snakes, but almost all of them are moving towards the fountain. I love that so many of them are just like, oh, I'm going to carry on doing what I'm doing.

Here's another one. There are 150 ,000 Skittles here. All the different colors.

And one by one, powered by Jev, the chopsticks are picking them out of the pile and sorting them into one of these five color buckets. You can actually see it zoomed in right here. You can see each skittle being picked by these chopsticks one by one.

And again, this is being done by Jev. But at this speed, it looks like a normal large language model. Now, what if we increase the speed to maximum?

Watch how fast it goes. Look at this, it is actually working. It is actually picking up each Skittle one by one.

It's making the decision in parallel. So potentially thousands at the same time, but it is still telling the chopstick one by one where to grab, which Skittle, which bucket. All right, so I have a few thoughts about this and then I'm gonna show you some incredible demos that I found on the internet.

Now, number one, this is a completely new architecture for artificial intelligence. It is much more structured. It is much more about decision -making.

It is not a chat model. So you're not going to be using it for coding from the ground up. You're not really gonna be using it for interactive chat sessions.

It is much more, here's a question I have, or here's a thousand questions I have, answer them as quickly as you can. So although it's not for every single use case, there are a lot of use cases that can benefit from the speed, the cost, and the reliability, having 0 % hallucination is a major value to many different industries.

Now, they specifically called out, for example, that Jev is not going to be nearly as good at playing chess than a chat GPT or a Claude model. And in fact, this guy right here put them head to head. So here's Jev playing Fable and here's Jev playing GPT -6 Astra.

And let's see. what happens so on the bottom astra one interestingly check this out look fable is about to run out of time so jev actually won against fable and that's important because it was actually the time constraint that made fable lose although the game looks like it was more or less over because there was only two pawns and two kings left so very interesting to see this so versus fable fable outplayed jev And by move 29, it was plus 16 immaterial and even promoted a second queen.

But it kept burning six to 15 seconds per move on analysis. And Jev answered it in 2 .6 seconds. So Jev could potentially win at bullet chess almost every time simply due to flagging, which is when you cause the other player to run out of time.

Here's Riley Brown, who built a model router. So take a prompt and route it to the best, most efficient, cheapest model possible. And Jev is the perfect model to have in between as the model router, deciding which model should this prompt go to.

It's not actually going to answer the question, but it'll route it to the appropriate model. Here's one by Kitsy introducing Unclutter, a smart ad plus slot blocker with that basically auto remove advertising and slot from web pages. And it does so in.

a fraction of a second if you want to try this out i'll drop a link down below i'm going to go install it right now i mean it's free you just have to bring your own key you're going to pay jev a few cents maybe per month it's like it's going to be nothing and it's open source and then possibly the coolest demo we have this guy justin schroeder who rebuilt tesla full self -driving in jev in less than an hour so you can see he created this world probably using codex or claude But the actual decisions being made of where to go based on all the information that the Tesla car is actually giving the model is being made by Jeff in real time.

And so you can see here it goes. It's going forward, straight, east, left, east, right. You know, it's a little wonky at times, but for essentially building this in an hour, it's actually quite impressive.

OK, so it sees the stop sign. OK, it's going to stop. And then it's going to continue.

Here's another example of it controlling a game in real time. This is from Alex from our team. And I'll drop his ex profile below if you want to follow him.

This is Jev controlling Melee. Look at this. Real time.

Unreal. So cool. So I think Jev looks incredible.

And I think we're just starting to understand how to use it. And the more demos we see, the more people get their hands on it. the more we're going to understand how valuable speed is going to be.

The Hook

The bait, then the rug-pull.

A ChatGPT co-inventor just shipped a model that plays Doom, sorts Skittles, and drives a simulated Tesla in real time, for free. Then it got checkmated in 18 moves.

Frameworks

Named ideas worth stealing.

01:12concept

RLCD (Reinforcement Learning for Calibrated Decisions)

TypeSafe AI's alternative to RLHF, which the creator argues trains a model for calibrated decision output rather than human-pleasing chat, cutting hallucination for narrow-scope tasks.

Steal forframing any 'not a chatbot, a decision layer' pitch for an automation product

CTA Breakdown

How they asked for the click.

VERBAL ASK

05:09product

“Go check it out. Huge fan. Click the link down below to let them know I sent you.”

Standard mid-roll sponsor read for Zapier, positioned as the pipe that connects real workflows to a fast/cheap model like Jev, framed as complementary rather than competing content.

Storyboard

Visual structure at a glance.

hook

hookhook00:00

origin/RLCD

promiseorigin/RLCD01:12

speed benchmark

valuespeed benchmark02:34

sponsor

ctasponsor05:09

chess verdict

valuechess verdict09:45

Frame Gallery

Visual moments.

hook

Frame at 00:12 from Jev is a decision engine, not a thinker, and it just lost a chess match to prove it

Frame at 00:23 from Jev is a decision engine, not a thinker, and it just lost a chess match to prove it

Frame at 00:32 from Jev is a decision engine, not a thinker, and it just lost a chess match to prove it

Frame at 00:42 from Jev is a decision engine, not a thinker, and it just lost a chess match to prove it

Frame at 00:51 from Jev is a decision engine, not a thinker, and it just lost a chess match to prove it

One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.

Watch next

More from this channel + related breakdowns.

08:54

Matthew Berman · Review

GPT-5.6 is FINALLY HERE (WOAH)

A 'dot' release plays out like a full generational leap: two five-to-seven-day unsupervised coding runs, a sponsor benchmark, and a live pricing and capability standoff against a rawer, higher-ceiling rival model.

July 9th

18:57

Matthew Berman · Review

Cancel Your Subscriptions, Ox-Alpha Is Here (GLM-5.3-Flash)

The mystery model that took over OpenRouter turns out to be a Chinese open-weights release that nearly matches frontier intelligence for a few cents a task.

August 29th

16:11

Matthew Berman · Tutorial

7 INSANE loops you need to try right now

A 16-minute walkthrough of 7 copy-paste AI coding agent loops — each one runs autonomously until a defined goal is met.

June 19th

26:49

Matthew Berman · Tutorial

You NEED to know these vibe coding secrets

A 27-minute systems playbook for turning AI coding tools into a self-managing development flywheel.

June 18th

33:44

Matthew Berman · Tutorial

21 INSANE Use Cases For OpenClaw

How one MacBook running Claude Opus 4.6 replaced a CRM, a security firm, a content team, and a personal chef -- with the exact prompts to copy every piece.

February 17th

14:17

Matthew Berman · Review

GPT-6 Astra: Full Review, Benchmarks, and Demos

An early-access hands-on with OpenAI's new GPT-6 Astra model — benchmark scorecard, alignment numbers, API pricing, and a run of 3D game and browser-automation demos.

September 3rd