Nicholas Runcie

🚀 LLMs can now do chemistry! Our new preprint shows that state-of-the-art reasoning models can now perform advanced chemical reasoning tasks, without any assistance from external tools. Here’s what o3-mini can already do👇 (1/🧵)

@thsottiaux @thsottiaux I love the codex app! (would love it even more if these bugs were fixed).

@thsottiaux @thsottiaux All these problems have started at some point in the last ~10 days.

@thsottiaux I'm having an issue with subagent threads being lost but still open. Subagents randomly disappear at end of turn, becoming inaccessible, but still contribute to thread limit. Currently need to restart the codex app every time to reset the subagent threads.

Assessing the Chemical Intelligence of Large Language Models #LLMs pubs.acs.org/doi/10.1021/ac… @NicholasRuncie @fergus_imrie @OPIGlets @UniofOxford #JCIM Vol66 Issue1 #MachineLearning #DeepLearning

@andrewwhite01 🙋‍♂️ I'm working on some stuff I think you will like.

Who are great young investigators doing AI in chemistry and automation in chemistry? I occasionally get asked for young people in the field and all the people I know are getting old. Feel free to self-nominate

If you’re interested in LLMs for chemistry, my ChemIQ preprint is here (accepted to JCIM): arxiv.org/abs/2505.07735

ChemIQ benchmark results: GPT-5: 70.2% GPT-5.2: 74.8% Solid jump. @sama @gdb

GPT-5.2 has impressive molecular understanding. ⚗️ SMILES → IUPAC translation: ~0% last year (GPT-4o). 67% today (GPT-5.2 high). This is a significant improvement! @sama @gdb

AIR Street Capital's "State of AI" 2025 Report references Nicholas Runcie, Charlotte Deane, and Fergus Imrie's work on Assessing the Chemical Intelligence of Large Language Models. Read the whole report here: stateof.ai And preprint here: lnkd.in/gbuNXR93

GPT-5 sets new SOTA on ChemIQ - our chemical reasoning benchmark. 🧪💥⚗️ The model scored 70.2%, with a 14% jump from the previous best model. A thread on the chemical intelligence of GPT-5 🧵 @sama @gdb (1/n)

I have summarized all my findings in a blog post (including some additional case studies). 📄 You can read more GPT-5 chemistry analysis here: blopig.com/blog/2025/08/g…

If you are interested in the ChemIQ benchmark, you can read a preprint here: arxiv.org/abs/2505.07735