I’m incredibly proud to share what we’ve been building at the AItonomy Foundation: Science’s Last Exam (SLE). 🧪✨
Can AI do science? Instead of Q&A, we give agents actual research goals and real-world tools (lab data, software), then grade them against hidden expert checks.
Deeply grateful to the brilliant minds at 34 institutions (Harvard, MIT, Stanford, etc.) who helped craft 113 tasks across 9 disciplines. ❤️
The results? Science is hard. Of 8 frontier models, the best scored 37.5/100. On true end-to-end research? All scored 0. We have an exciting journey ahead! 🚀
(Claude Opus 5.5 & GPT-6 Astra coming soon—just clearing bio safeguards).
Everything opens Nov 6. 🗓️
🌐 sciences-last-exam.org
💻 github.com/Gen-Verse/Science…
🤗 huggingface.co/datasets/Gen-…
Thread below on how the exam works and where models break. 👇 (1/14)
#AGI #AIforScience #AIAgents #SciencesLastExam #OpenAI #Anthropic
The 6th capability is the one we are most careful about—and arguably the most exciting. We are moving from code to the physical world with Wet-lab Execution. 🧫
Computation can tell you if a protocol is plausible. Only the actual lab bench tells you if it works.
Here is how we are closing the loop safely:
1️⃣ The agent writes a protocol.
2️⃣ Strict safety review (BSL-1 only, whitelists, human sign-off). 🛡️
3️⃣ Blind execution by humans or robots who don't know the expected result.
4️⃣ Data goes back to the agent to troubleshoot, revise, and try again! 🔄
We want to see if an AI can actually figure out why a physical experiment failed and fix it.
🤝 Call for Partners: We are looking for labs to build this with us! If you have automated benchtop setups, highly refined protocols, or just want to see how an AI handles your hardest routine experiment—let’s talk. (13/14)
This brings us to our final ask: Bring us your hardest problems. 🧠
SLE is built by researchers, for researchers. We currently have 43 domain experts from 34 institutions, and we’re actively recruiting toward 200 across every scientific discipline! 🌍
What makes a great SLE task?
🔬 Real: From your actual experiments, code, or lab work.
🧩 Challenging: Requires deep domain judgment, not just basic execution.
✅ Verifiable: Has an objective way to score it outside the AI's reach.
If your task is accepted, you’ll join the author list of the official SLE paper! 📝 (You can also contribute by running your own agents or joining our wet-lab validations).
A massive THANK YOU to everyone who built the first 113 tasks. 🙏
Can an AI do real research on its own? We’d rather find out with concrete tasks than argue about it on the internet. That’s our invitation to you. Let's find out together. 👇
🌐 Contribute: sciences-last-exam.org/contr…
💻 GitHub: github.com/Gen-Verse/Science…
🤗 Hugging Face: huggingface.co/datasets/Gen-…
#AIforScience #OpenScience #SciencesLastExam (14/14)
On Monday, we hosted a private salon at Stanford titled "Towards Personal AGI." We were honored to welcome:
-Travis Oliphant @teoliphant, Founder of NumPy, SciPy, and Anaconda
-researchers from Columbia University’s Data Agent Process Lab @Zeming_Liang & @Yuting_Deng
-researchers from MIT, Microsoft, and Harvard @YuexingHao & @XiaominLi98 @MatrAIx2026
-researchers from UCSD @Ruiyi_Zhang
Our other guests hailed from academia—including Stanford, UC Berkeley, MIT, Harvard, Princeton, Caltech, Columbia, Oxford, and Yale—as well as the industry, representing organizations such as Anthropic, OpenAI, DeepMind, Meta MSL, Amazon AGI, Nvidia, Tesla, Apple, Microsoft Superintelligence, Together AI, ByteDance, AMD, and Samsung.
We are deeply grateful for everyone's support. We are also seeking additional collaborators for our project: aitonomy.org/projects/sci-ac…. Welcome to join us in building open-source infrastructure for AI for Science!
Announcing AItonomy Foundation.
We are PhD students and postdocs at Stanford, Princeton and Oxford, building a nonprofit research organization for AI4Science and Science4AI — and connecting early-career scientists across the world to work on it together.
Our conviction is simple. AI advances science, and science advances AI. Neither direction moves as fast as it should, because the open infrastructure both sides depend on — the benchmarks, the verifiers, the record of what is genuinely unsolved — does not yet exist. We are building it in the open, and giving it away.
Two projects are launching today.
ScienceAccelBench — scientific code, accelerated by AI and verified against the original code's own output. The faster code goes back to the lab that contributed it. aitonomy.org/projects/sci-ac…
ScienceMysteryBench — a finished discovery, rewound to the raw data it came from, and handed to an AI to attempt again. Eight scientific domains, one question each. aitonomy.org/projects/scienc…
We are looking for people to build this with.
Researchers — join a project. We provide the AI credits and the GPU time; you bring the science. Contributions are credited by name, and lead to a paper on a short timeline.
Frontier labs — we would like to work with you. Tell us what evaluations you need, and consider supporting the work with API credits.
Sponsors — compute and API credits are what we need most. Everything we build stays open, free, and behind no paywall.
Everyone else — a repost genuinely helps us reach the scientists we are looking for.
aitonomy.org
