The Polymathic AI Collaboration. Shared account.
Joined September 2023
- Tweets78
- Following69
- Followers3.7K
- Likes52
PolymathicAI retweeted
"The world is predictable in the right coordinates. The open question is whether a model can learn those coordinates on its own."
World model for Physics conference in Aspen coming winter 2027! Application deadline October 5 !
organized by @randall_balestr @KempeLab @ylecun and me 😀
PolymathicAI retweeted
1/ At @PolymathicAI, we have been asking:
What does it take to integrate general LLM agents with specialized scientific software?
Today, I am happy to announce Agent Rosetta, an LLM agent to support scientists in real-world protein design tasks using the Rosetta Modeling Suite.
5/ Spotlight presentation tomorrow Saturday 11th at the AI4Science workshop #ICML2026
Info:
🎙️Spotlight talk: 10:20 AM, Hall C
📌 Poster: 4:00–5:00 PM, Hall C
Come say hi!
4/ We evaluated on three datasets of varying difficulty.
and found out that our strongest agent matches or exceeds the performance of chemistry graduate students 🧑🔬
3/ Chemists solve it like a puzzle: one fragment at a time, until the full structure matches the measured spectrum and constraints 🧩
Our agent mimics this: proposes candidate molecules, tests against the chemical + spectral evidence, and assesses when they fail.
2/ Our key insight: going from spectrum → structure can be reframed from a one-shot prediction problem into a search problem, driven by an agent.
1/ NMR is one of the most important tools for figuring out molecular structure that's used across chemistry, biology, materials science, and drug development.
Collecting the data is routine, interpreting it can take days, sometimes years 😮
We built an AI agent for this 🧵
Interested in building the next generation of AI for fluid dynamics?
We're looking for a Research Scientist / Postdoctoral Researcher at Polymathic AI in New York to develop foundation models for computational fluid dynamics.
Interest form: forms.gle/7n4mrJyEJjmb4W8S9
PolymathicAI retweeted
Want to apply principled Bayesian methods to Large scientific foundation models?
Join @PolymathicAI and @EmtiyazKhan at Adaptive Bayesian Intelligence to explore the boundaries of SOTA large models for scientific data and Bayesian Principles in Tokyo 🗻 , Germany 🍺 and NYC 🍎!
post-doc: apply.interfolio.com/184211
research-scientist: apply.interfolio.com/184213
PolymathicAI retweeted
Happy to announce the partnership between @relativityspace and @PolymathicAI to explore Mars!! 🤩
Come work with real rocket scientists, astrophysicists and AI experts to explore Mars ❣️
Links:
job-boards.greenhouse.io/rel…
job-boards.greenhouse.io/rel…
Shout-out to the team @irina_espejo @damon_hinz @HaewonJeong00 @cosmo_shirley and to our generous funders and institutions @schmidtsciences @NYUDataScience @FlatironInst @SimonsFdn
(7/7)
There’s still a long way to go! The pipeline for small organic molecules can be improved and we will perform more evals, but that it's just scratching the surface…
If you work on NMR spectroscopy, let’s work together🙂
Stay tuned for the camera-ready version!
(6/7)
On the rest of the tasks, we show that framing NMR elucidation as an agentic search is not only viable but competitive with modeling approaches trained on simulated data and/or tabulated peaks on a variety of small molecules with different environmental sample conditions.
(5/7)
We test our agent on total >200 molecules with varying difficulty 😉
Input: 1H, 13C NMR + molecular formula (no context/reagents/solvents)
On the human-annotated subset, top-1 accuracy is:
Our Agent 71 ± 6%
Chemistry Graduates 67 ± 3%
Codex / CC out-of-the-box ~30 ±10%
(4/7)
We mirror a real lab workflow. Our agent starts from instrument data and preprocesses it with custom tools.
This is not straightforward and preprocessing errors only show up *downstream*.
So the agent has to backtrack with new assumptions. Real lab data is messy!
(3/7)
Anthropic’s work starts from reported peak lists. But, that’s *not* how chemists work.
They start with FID data, which includes effects of impurities, solvents, and artifacts.
Only *after* solving the structure, chemists report a clean tabulated list of compound peaks.
(2/7)
Are chemists done & is NMR solved? Not so fast! 🤔
Outperforming software is not the goal, to be truly useful, AI needs to operate like a human chemist under real lab conditions.
✨Check our #ICML2026 AI4Science Spotlight!
🔗OpenReview: openreview.net/forum?id=hsfH…
Thread 🧵
(1/7)
New Anthropic Science Blog: Making Claude a chemist.
To manipulate a molecule, chemists first need to understand its structure. Their main tool is NMR spectroscopy.
We found Opus 4.7 matches—and on some tasks beats—dedicated NMR software. Read more: anthropic.com/research/makin…
Interested in building the next generation of foundation models for science?
Join Polymathic in New York (NYU + Flatiron) for a postdoc.
Topics include generalization/transfer, and solar physics.
Gen-transfer: apply.interfolio.com/185111
Solar-physics: apply.interfolio.com/185110
