Twitxy Saved

Alon Vinnikov

@AlonVin

doubleAI | ex-@MSFTResearch | Advancing LLMs, Speech, and Vision
Joined May 2014
251 Following 218 Followers 30 Posts

Latest posts 5

Alon Vinnikov @AlonVin
It's fascinating watching LLMs rewrite architectures during training.

Let's see how far mechanistic intervention takes us.
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
14 replies 7 reposts 39 likes 1.7K views View on X
Alon Vinnikov @AlonVin
Where does your AI live?

Until now: a chat box, a terminal, an IDE. Beside a human who plans, reviews, approves.

Maxwell lives inside your code while it runs, swapping in faster code. No human in the loop.

Like Lean for math: every rewrite is proven equivalent.
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
0 replies 0 reposts 12 likes 666 views View on X
Alon Vinnikov @AlonVin
This is the highest cost I've seen yet for buggy AI-generated code: your own research time.

Pokes a real hole in the RLVR premise. Is anything really a "verifiable" domain anymore? Models are strong enough now that the verifier is the weak link - we need to up our game.
doubleAI Verified account @doubleAI May 27
WarpSpeed, our autonomous optimization agent at @doubleAI, just took first place on @NVIDIA's new SOL-ExecBench: 235 of the hardest CUDA kernels in production.

But the more interesting story is what we found along the way.

Verifiers designed for human errors don't defend against AI reward hacks. We found four ways the same benchmark's verifiers can be silently fooled. The first one broke transformer training.

🧵
View quoted post on X
Alon Vinnikov @AlonVin
We built an AI system that optimizes GPU kernels and pointed it at cuGraph — NVIDIA's GPU-accelerated graph analytics library.

The outcome: doubleGraph, a drop-in replacement, 3.6x faster on average, with many algorithms exceeding 10x.

Full blogpost: doubleai.com/research/doubl…
www.doubleai.com Maxwell Beats Expert-Written GPU Kernels at Scale | doubleAI Maxwell built and verified a drop-in NVIDIA cuGraph that runs 3.6x faster on average, speeding up every algorithm across three GPU architectures.
Shai Shalev-Shwartz Verified account @shai_s_shwartz Mar 2
1/ Software was eating the world - and now AI is eating software.
AI already beats humans at math/coding (IMO, CodeForces). Right?

So let's test the strongest coding agents on a real domain: optimizing cuGraph (GPU graph analytics kernels).

Spoiler:
* The strongest coding agents crash.
* And @doubleai built WarpSpeed - an AI that beat a decade of expert-engineered GPU kernels.
🧵
View quoted post on X
0 replies 1 reposts 14 likes 737 views View on X
Alon Vinnikov @AlonVin
If you’re curious about the FormulaOne reasoning benchmark mentioned in the video, here’s a link: arxiv.org/pdf/2507.13337
Shai Shalev-Shwartz Verified account @shai_s_shwartz Dec 24, 2025
What does it take for AI to reason?
I break down how models can actually learn to reason, not just mimic. Based on the Diligent Learning paper with @AmnonShashua at doubleai.com. Link to the talk 👇
youtube.com/watch?v=QkQnIX…
View quoted post on X