Twitxy Saved

Tobias Gessler

Verified account Axiom

@tobiges

sf Joined April 2019
697 Following 873 Followers 200 Posts

Latest posts 5

Tobias Gessler Verified account @tobiges
We gave Dream-1 from @dreamlabs_ai control of a Tesla. With only 1.5h of driving demos, it learned to drive and finished an unseen cone course 5/5 times 馃殫

Dream is a WAM built for two-armed robots. You wouldn't expect it to drive a car at all. Our action interface is [curvature, speed], very OOD for two-armed robots. A small fine-tune was enough to make it drive the Tesla!
0:30
DrivingBench Verified account @DrivingBench 1d
Introducing DrivingBench v2 馃У

We wired a Tesla Model 3 so any AI model can drive it, and put 8 frontier VLMs (Claude, GPT, Grok, Gemini models) and 6 fine-tuned robot models (VLAs & WAMs) behind the wheel.

A new lot and two new courses, none of which they鈥檇 seen. Only one VLM ever reached the finish, and its fastest run took 26 minutes. 馃弫
View quoted post on X
3 replies 3 reposts 17 likes 624 views View on X
Tobias Gessler Verified account @tobiges
Jacob Coxon Verified account @hilbertspaess 5d
This slide from the Anthropic IPO roadshow was a big factor in my decision to leave. It鈥檚 an insane plot and I鈥檓 surprised they made it public.

It shows the compute efficiency of pretraining over time. The y-axis is a **log scale**: they are seeing continued superexponential algorithmic progress.

Mythos was not a one-time deviation, but a point on a superexponential. They could already retrain Mythos with a fraction of the compute originally needed.

It doesn鈥檛 strictly imply a software-only intelligence explosion- they have been spending more on R+D. But algorithmic improvements do not require ridiculous creative genius and are on the verge of being automated. You do the math.
View quoted post on X
1 replies 0 reposts 25 likes 1.3K views View on X
Tobias Gessler Verified account @tobiges
Excited to be speaking next week at SGLang Summit by @lmsysorg on how we've built scalable post-training infrastructure to push the frontier of math
LMSYS Org Verified account @lmsysorg Nov 7
An in-house training stack for mathematical formalization.

Tobias Gessler from @axiommathai shares how the team approaches building training infrastructure for this specialized domain.
View quoted post on X
3 replies 4 reposts 23 likes 3.1K views View on X
Tobias Gessler Verified account @tobiges
gradually, then suddenly.鈥╪ot long ago we formalized individual proofs. now the entire stacks project.鈥╝ll of math next.
Axiom Verified account @axiommathai Nov 4
Most think vibecoding at scale is slop. Here鈥檚 a counterexample.

Today we release a 16M-line formalization of the entire Stacks Project, critical infra underlying modern math, with optimized code quality and its highly interconnected structure preserved.

A mathematical city 馃寙
View quoted post on X
1 replies 3 reposts 31 likes 1.7K views View on X
Tobias Gessler Verified account @tobiges
lots of great talks on benchmarks at the @GoogleDeepMind and @kaggle benchmark summit! the frontier tends to go wherever we can measure. thanks for having us, it was great speaking there!
Ivan Leo Verified account @ivanleomk Oct 28
At @kaggle bench summit and folks are talking about plugging a llm into a @comma_ai car haha

Physical benchmarks are so back
View quoted post on X