Twitxy Saved

Robert Joseph

Verified account This account is verified. Learn more

@Robertljg

math+cs phd @caltech
Los Angeles, CA robertj1.com Joined February 2026
305 Following 596 Followers 397 Posts

Latest posts 5

Robert Joseph Verified account @Robertljg
What an incredible weekend at the @Caltech × @SAIRfoundation Mathathon!

Huge congrats to the student organizers (seriously go thank @_sathvikr, @Mav_Gators, @Brian_bfz, mayla and the whole @math_hackathon team), @xtxmarkets and all the sponsors. For an experimental event at this scale, everyone did an amazing job!

I enjoyed seeing/judging and discussing what people built in just 40 hours :) The best part wasn't the AI, it was the creativity and the effort to make hard math easier to understand.

I personally preferred the original theme of picking your own problems and solve open ones, but this was much needed for the math community and a fantastic experiment and I'm glad I got to be part of it. Congrats to all the round 1 winners!
3 replies 4 reposts 36 likes 1.3K views View on X
Robert Joseph Verified account @Robertljg
really fun doing this with @srush_nlp!! softmax was just the start. we did a full numerical audit of ML ops across PyTorch, JAX, TF, NumPy, CuPy, ONNX Runtime and MLX, checked against TorchLean, and found a lot more!!!
Sasha Rush Verified account @srush_nlp 4d
Lean Verified Precision (w/ @Robertljg)

Recently, "A Fix for Softmax" demonstrated a surprising bug in the textbook softmax implementation. How do we avoid having subtle bugs in the world's most expensive software? 🧵
View quoted post on X
7 replies 5 reposts 70 likes 4.6K views View on X
Robert Joseph Verified account @Robertljg
so many formalization announcements this week!! Signal verified down to the machine code @GoogleDeepMind, the @axiommathai Stacks Project and Weil conjectures in Lean, a Lean proof of Aaronson–Ambainis, anthropic hill climbing on FLT etc super exciting :)

but Lean only checks that the proof proves the statement. it doesnt check that the statement is what the paper actually meant (semantic issue)! thats still on us (or agents) to review!

so I wrote up how I believe we should do this properly, for math, physics and code and other fields. most of it I learned from the @leanprover community, talking to people in the field and my own personal research so I would love feedback!

robertj1.com/ai4science/get…
www.robertj1.com Getting the (In)formal Statement Right Proofs got cheap; statements didn't. My checklist and habits for autoformalization in Lean, from Comparator and the kernel to levels of a formalization.
2 replies 7 reposts 45 likes 2.7K views View on X
Robert Joseph Verified account @Robertljg
spent pretty much the whole day testing inkling 2 small on different lean tasks, and i'm genuinely impressed! it's really good at lean, especially using tools to work with mathlib and other lean libraries. glad to see this much attention going into these capabilities especially compared to the last model!
Thinking Machines Verified account @thinkymachines Nov 5
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.

Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.

thinkingmachines.ai/news/inkling-2…
View quoted post on X