Join our growing #compchem team as an Assistant Prof UoBChemistry incl David Scanlon, Adam Michalchuk, yours truly, and others! Based on my first 6 weeks here it’s an amazing place to do great science! edzz.fa.em3.oraclecloud.com/hcmUI/Candidat… #chemjobs #AcademicChatter

We created a materials science QA benchmark spanning characterization, applications, properties & behavior at the #LLMs for #Materials and #Chemistry Hackathon (Mohd Zaki, Abhijeet Gangan, Hargun Singh Grover, Faisal Anees).

it was wonderful to have the fearless Melissa Ramirez with us in the The Gomes Group at Carnegie Mellon University for the past few weeks. just a bit too short. amazing chemistry to come soon, in collab with Olexandr Isayev 🇺🇦🇺🇸 and Brian Stoltz!

ML and #AI Twitter summary of this week 😎

Tool-augmented LLMs 🤝 Chemistry 🚀

Congrats to Andrés M Bran, Sam Cox (with Andrew White crow/acc), and the whole team! I'm proud of everyone who contributed to this fruitful collaboration!

Artwork by our favorite scientific illustrator Ella Marushchenko - thank you!

ChemCrow is out today in Nature Machine Intelligence! ChemCrow is an agent that uses chem tools and a cloud-based robotic lab for open-ended chem tasks. It’s been a journey to get to publication and I’d like to share some history about it. It started back in 2022. 1/8

This is the first paper from my lab at EPFL🙂 🎊. It conducts a detailed analysis of the representation dynamics of PPO to understand the reasons for poor performance that happens under certain conditions. We also conducted detailed analyses on ways to mitigate it.

How overfit are popular LLMs on public benchmarks?

New research out of @scale_ai SEAL to answer this:

- produced a new eval GSM1k
- evaluated public LLMs for overfitting on GSM8k

VERDICT: Mistral & Phi are overfitting benchmarks, while GPT, Claude, Gemini, and Llama are not.

New preprint! “Mixed Continuous and Categorical Flow Matching for 3D De Novo Molecule Generation” arxiv.org/abs/2404.19739… FlowMol is a flow matching model that jointly samples the geometric and topological structure of molecules. We also present surprising results on flow…

Match made in heaven! Probably among the largest H100 per researcher count in our field together with probably one of the largest proprietary structural datasets. If you are a researcher - work with us to solve the big remaining outstanding challenges in computational structural…

Anthropic Security will be at BSidesSF on May 4-5th! We have a number of roles open on my App Sec, Security Software Engineering, Detection and Response, IT, and Compliance teams. We'll be hosting a Capture the Flag contest with some unique prizes! anthropic.com/careers

THEY DID IT

Invalid SMILES is all you need 🤣

The White House Synthesis Screening Framework just dropped!

* requires providers to follow 2023 HHS guidance
* biofoundries, cloud labs, core facilities and CROs count as nucleic acid providers
* adherence via self-attestation

friends in Seattle and adjacent areas: I’ll be over there in about two weeks. it would be awesome to meet and visit some of you and your groups. reach out if interested.

Very cool paper in Science by the Baker Lab at UW. Screening of 15M macrocycles is enabled by our AIMNet machine learning potential! #compchem Can you imagine what we can do with a 2nd generation of AIMNet2 💪🏻🦾 science.org/doi/10.1126/sc…

Advice to my younger self:

- surround yourself with truly excellent people. top 1% in some important way

- find something to work on for a LONG time (10+ yrs). jumping around doesn’t beget mastery

- work harder than you think you should. hard work is a habit best learned young

Very, very true. When asked to list “research outputs” we can or should list papers, software contributions, open data sets, media participation and all other types of outreach.

Someone asked me yesterday why I would contribute to open-source software rather than writing another paper.

I personally feel that my impact is much larger if I provide tools to researchers that they can use to investigate materials than if I look at one more material

We are looking for a talented and motivated postdoc to work in a collaboration between Frank Noe and my group at Freie Universität Berlin on the development of machine-learned coarse-grained protein models. Please email [email protected] if interested.