Twitxy Saved

Yoav Levine

Verified account This account is verified. Learn more

@YoavLevine

co-founder @doubleai; assistant professor @ Tel Aviv University
Joined January 2018
94 Following 680 Followers 102 Posts

Latest posts 5

Yoav Levine Verified account @YoavLevine
Our Maxwell found staggering improvements across the throughput/ interactivity plot for serving DeepSeek-V4.1-Flash.

This model is out for several weeks now, and everyone has taken a stab at optimizing it for inference 💸 Maxwell topped all published results, and even surpassed GB300 serving numbers with B200 💪🏻
doubleAI Verified account @doubleAI Nov 11
1/
We set Maxwell to optimize DeepSeek-V4.1-Flash serving on B200s.

Maxwell’s engine outputs 3x more tokens per GPU than vLLM at the same speed per user, or serves 2x speed per user for typical loads.

It beats every published B200 setup, and is even faster than all comparable GB300 results.

Blog: doubleai.s.gy/gb300
View quoted post on X
1 replies 1 reposts 23 likes 1.3K views View on X
Yoav Levine Verified account @YoavLevine
The Transformer architecture was designed by scientists. Our agents evolve it during training, injecting code layers and modifying the architecture as the network trains.

Huge kudos to @AlonVin for leading this!
doubleAI Verified account @doubleAI Nov 6
1/
What if an LLM could not just edit a model's weights, but re-architect it as it trains?

We tried just that. The resulting architectures exit the known territory of transformers: the coding agents spliced code directly into the network.

On nanochat, the loss fell below ordinary training after each new code layer was introduced. On algorithmic reasoning, accuracy rose from 55% to 84%. We even let DeepSeek-V4.1-Flash act as the agent modifying itself, raising its AIME and HMMT 2026 score from 64% to 71%.

To do all this, the agents built a set of tools we’ve never seen before, which we call mechanistic intervention!

Blog: doubleai.s.gy/llm-redesign
View quoted post on X
3 replies 2 reposts 56 likes 2.9K views View on X
Yoav Levine Verified account @YoavLevine
@doubleai  introduces Code Execution Models.

Coding agents help humans write code. Code Execution Models help the hardware run it, unlocking a world of 100% hardware utilization.

This was impossible without formally verifying code, because translating from programs to binary must be exact. We did that. Meet Maxwell.

heymaxwell.ai
heymaxwell.ai Maxwell: The first Code Execution Model, by doubleAI Maxwell improves code while it runs, adapting to hardware and data. Every improvement is formally proven equivalent to the original code.
doubleAI Verified account @doubleAI Oct 19
What if code could think while it runs?

Meet Maxwell: the world’s first Code Execution Model.

A new species of AI that lives inside the machine.

We believe it’s the key to unlocking 100% of compute.

Requests for early access are open, link below.
View quoted post on X
1 replies 0 reposts 24 likes 409 views View on X
Yoav Levine Verified account @YoavLevine
Our Double Agent keeps innovating. This time on one of the toughest domains humanity has faced — chess riddles!
doubleAI Verified account @doubleAI Oct 13
1/ What is the maximum number of legal moves in a chess position where every one of them gives check? The record on that question had stood since 1968. Double Agent, our AI research system, just broke it.

How many checks can you count here?
View quoted post on X
Yoav Levine Verified account @YoavLevine
Amnon’s lecture on the breakthrough results by our double agent system and our novel verification techniques. A fascinating 45 minute watch on the cutting-edge of AI capabilities and limitations.
Amnon Shashua Verified account @AmnonShashua Oct 9
Our breakthrough on the Moore girth problem raises a larger question: when will AI become a great scientist?
In this lecture, I discuss what made this result possible—and why verification remains a fundamental barrier across science.
Watch the lecture youtube.com/watch?v=AIMnYE…
View quoted post on X