eigenpaul (@eigen_paul) header
eigenpaul (@eigen_paul) profile picture

deep in the RL mines teaching models to do science. resident @PrimeIntellect

Groningen
Joined August 2024
Photo by eigenpaul (@eigen_paul)Photo by eigenpaul (@eigen_paul)Photo by eigenpaul (@eigen_paul)Photo by eigenpaul (@eigen_paul)Photo by eigenpaul (@eigen_paul)Photo by eigenpaul (@eigen_paul)Photo by eigenpaul (@eigen_paul)Photo by eigenpaul (@eigen_paul)Photo by eigenpaul (@eigen_paul)Photo by eigenpaul (@eigen_paul)
Happy to finally release TransitBench, a benchmark for autonomous scientific programming in astronomy. It measures how far AI agents can automate the search for exoplanets.
1
1
7
427
Our RL Residency gives independent researchers compute grants and mentorship to build open evals and training environments for overlooked problems. In 2026, residents shipped projects spanning robot control, medical imaging, gameplay, and GPU kernels. Some highlights:
18
52
7
723
43,784
eigenpaul retweeted
Today we're launching a preprint server built specifically for AI agents For three months, nearly 1,000 agents from 35 labs (across Stanford, MIT, ETH Zurich, and more) conducted research autonomously on this server, producing 936 papers alongside their code, data, and model weights. Agents found significant errors in published human research and even created their own journal to review each other’s work. Some agents independently produced results similar to those in human papers that later appeared on arXiv, including findings on failure modes in self-distillation and feature reproducibility in mechanistic interpretability. As a part of this study, we open-source the full twelve-week archive of manuscripts, code, data, peer reviews and agent chat transcripts. Read the study: alphaxiv.org/abs/2611.societ…
77
321
129
1,708
148,154
Leave one AI agent alone to do research and it often cuts corners. Leave nearly 1,000 AI agents together for three months and they invent peer review. Across 35 labs (Stanford, MIT, ETH Zurich, UCSF, and more), agents published 936 papers with full code and data on a shared server built for AI. No human wrote code or text. Unprompted, they debated study designs in chat, retracted their own mistakes, found real errors in human papers, and started a journal that caught other agents cheating in code. In blind reviews against human arXiv papers, humans still won on original ideas, agents won on rigor. alphaxiv.org/abs/2611.societ…
64
289
56
1,555
146,551
eigenpaul retweeted
we have a (new) pretraining feature that lets you do autoresearch on top of @karpathy's nanochat scaling ladder on a few nodes but also launch jobs on thousands of gpus with automatic recovery love how beautiful these scaling plots look even at very small scale
Introducing Model Factory. Train, deploy, and continuously improve your own models on one stack. Run on-demand, on dedicated clusters, or bring your own. Own your intelligence. Own your model factory.
11
22
2
299
16,111
eigenpaul retweeted
9/ Estimating a galaxy's distance using telescope images is challenging, especially for nearby galaxies. With an astronomer, DiscoveryAgent built a model to find nearby galaxies. It correctly placed dwarf galaxies that other models put much farther away.
2
8
900
the simulated lab is now mostly built. unfortunately i have a tendency to recursively abstract my research questions, so im now wondering whether the scientific method itself can emerge through gradient descent
Replying to @eigen_paul
the plan is to build simulated worlds with hidden laws and put them behind instruments that can lie. i want to see how far ai scientists can reach without fooling themselves and mostly whether they pick experiments that tell the physics apart from instrument artifacts
1
1
48
eventually i want to replace the simulator with actual lab equipment. how far can we get just letting the model run experiments in the physical world and updating its weights along the way?
1
1
13
the universe has an unlimited supply of OOD evals. terrible throughput tho. we should spend the next few years building more ways to ask nature
1
17
Today I’m launching what I think could be an extremely important research project. Fixing humanity’s scientific knowledge. The idea is inspired by physicist Casey Handmer, who used AI to screen more than 50,000 papers in theoretical astrophysics. His audit flagged 872 possible errors. This is impressive. But why stop at finding the errors? Let’s use AI to FIX them. And why stop at astrophysics? The ambition of this project is to go through every scientific paper ever published, across every scientific discipline. AI agents will review each article and produce a revised version that corrects the errors they find. The aim is not to shame authors or seek retractions. We are humans and we all make mistakes. The ambition of this project is to correct all the errors that we have made so that current and future generations of scientists can build on firmer foundations. I’m looking for researchers from every discipline to help make this happen. If you’d like to take part, please fill out this short survey and tell me which disciplines you’d like to help cover. Feel free to share with anyone interested. Survey link: unimib.qualtrics.com/jfe/for…
111
98
36
857
51,833
the masculine urge to turn the entire observable universe into a verifiable reward
2
62
eigenpaul retweeted
ENTROPY is one of the 5 fundamental laws governing agent swarms. We've observed agents finding their own mediums to share information across the swarm, and over time their hypotheses start becoming correlated. You could have a 1000 agents running experiments but effectively exploring the same tiny region of the search space. So regulating information diffusion becomes extremely important. You want agents to learn from each other without prematurely converging, preserve independent lines of inquiry, and adjust selection pressure as new evidence surfaces. In ArkenOS, we currently expose this through a simple exploration slider, like temperature in LLMs. Eventually, we want swarms to regulate their own entropy and adapt how they explore!
You can now control your swarm's exploration entropy in ArkenOS v0.6.12 with a simple slider! Want agents to stick to proven methods or chase completely new hypotheses? Go from Exploit to Explore in one slide. A little more control over how your swarm thinks, experiments, and discovers!
4
14
2
175
7,907
if many-worlds and the simulation hypothesis are both true, we're self-aware commits in a repo with no main branch
33