Friends I got early access to jev and started testing it

Set four Jev instances to play the Turkish card game 101 against each other, evaluating hand, table state, and opponents' moves.

The original post, translated

Friends

I got early access to jev and started testing it, and I have to say I'm really impressed

Please watch the demo I recorded.

This is definitely not a sped-up video. 4 jevs are playing the 101 game with each other, look at the speed of the game.

Those who follow me know.

For years, every time a new model comes out, I play our famous 101 OK game.

One of my personal benchmarks is AI models playing 101.

"Ha, there's no better benchmark to enter our home.

Because 101 is much harder than it looks

The model first needs to understand the rules of the game, then fully follow these rules during the game. But this is not enough. It needs to evaluate the stones in its hand, the state of the table, that is, the last state, and the opponent's moves, and make the right decision in time. Each model needs to know the instantaneous state of the table and be able to make decisions quickly

Especially the last part is by far the hardest part for models

Even models like openai, anthropic, etc. can make good decisions, sometimes they think for 30 seconds or longer. In real-time games, making the right decision is meaningless, because you made the right decision, but you missed your turn, which is no good. So, speed + correct decision-making is the most important metric here.

That's why I started using Early Access before the announcement. I remember my 101 Benchmark

Actually, as a concept, jev is not something unfamiliar.

Simply put, we can describe it as a generalized zero-shot classifier. It doesn't produce long text, but instead makes a decision between defined options based on a given state.

But the most beautiful thing is how fast he does it. I have tried different models for three to four years, but I did not get full efficiency, and it seems the problem of AI playing 101 has been solved

Making decisions in milliseconds, and performing very accurately in the game

Chapter 101 is not the most important thing

Chapter 101 this is a tiny little thing

In areas like computer use, browser use, games that need speed, that is, in every scenario where the model is constantly thinking about what I should do, jev is speeding things up. For millions of small decisions, it seems very ideal

A huge generative model thinking for 30 seconds about millions of micro-decisions is actually not workable. Instead, a model that can make good enough decisions within 100 milliseconds makes more sense.

That's what Jev does.

I think it meets and will meet a huge need.

Remember, when there is no high-speed internet infrastructure like 4G, real-time streaming like tiktok vs is impossible. TikTok needs 4G or high-speed internet. Now, jev = 4g. Solving the speed problem. So, let's see what kind of innovation Jev will bring through Jev's Toku in the coming semesters.

If you have a concept that needs millions of micro-decisions made immediately, this is a very ideal thing.

I think companies like openai and anthropic will also gradually separate "generation" and "decision-making"

The post above is a machine translation from tr; the untranslated text is in the fold-out.

Quoted post

C

@CompleteSkeptic34.6M views

source

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

Engagement when collected

Numbers are a snapshot taken from X when the case was added to the library (schema v1, collected 2026-09-19); they will not match today.