Wet labs for AI Agents (YC F26)
Joined July 2026
- Tweets69
- Following0
- Followers361
- Likes29
Petrichor retweeted
The breakthroughs I most want to see from AI are in medicine, not maths. @Petrichorbios are building wet labs for agents to make that possible.
Congrats on the launch @Wilhelmmlem, Jesper and Elias!
Models use the lab directly through our Python library, writing each experiment as a job and submitting it themselves.
We capture as much data as we can from every run, so labs end up with the best possible training data for improving their models.
2
17
3,456
Customers own all the data we produce for them, and we never share it or train on it.
If you work on biology at an AI lab, email me at [email protected]
14
3,017
Petrichor retweeted
We gave GPT-6 Astra and Claude Fable 5.1 the X-ray data for 30 real drug crystals and 8 hours each to work out the atomic structure.
Astra solved 10. Claude solved 3.
Astra also handed in 18 wrong answers. Claude handed in 3.
Our reading of these runs: current models can already carry out a crystallographer's workflow offline, building tools as needed, and finish about a third of a hard panel. On the rest they mostly submit wrong answers with some confidence. Well short of solved, far from nothing.
1
1
85
The whole powder-diffraction environment is now public: all 30 compounds, the grader and its test battery, the reference structures with every source DOI, and the full visible traces of both runs. If you think we graded something wrong, the material to show it is there.
github.com/Qokedas/blackbox-…
1
1
83