The major reason I have been most excited about the open, collaborative-style autoresearch competitions that we at Eigen Labs have been working on is that it gives a growing sense of what research and education in the post-AGI world can look like.

There has been a lot of chatter about how professors and experts won’t have any need for recruiting students in their labs, as they can just get their agents to do the grunt work, and that this will lead to a collapse of the training of the next generation of researchers. But this is not any different from the pre-AGI world, where student recruitment to participate in frontier research was already blockaded by being selected into PhD programs or summer lab internships in prominent places.

I 100% disagree with this cynical view of what post-AGI learning and research would be.

What our autoresearch competitions, ecdsa.fail and openfrontiercs.com, have convinced me is that good models, an emerging better understanding of harnesses, and collaborative yet competitive open leaderboards create a revolutionary opportunity to democratize the ability for anyone to do frontier research.

I want to break down the last ingredient. The collaborative style with open participation allows anyone to investigate and understand submissions from other participants to a given problem statement rapidly, and then build on top of it—whether by grinding on the same hill or jumping to a whole other hill (by getting an incisive understanding of the solution manifold of that problem on the fly). On the other hand, the competitive desire of being on the leaderboard definitely gives participants the urge to better understand the subject so as to be better at agent steering.

This is like learning by getting rapid feedback from an open ecosystem of researchers in a matter of hours or days, instead of waiting for months and years to make progress on your result.

What struck me about the diversity of participants in our autoresearch competitions is that almost none of them were subject-matter experts, yet there was an excitement and happiness in them to be participating in it.

The most critical component in setting up an autoresearch competition is the verifier that checks the correctness and performance of the submissions. This is where being the subject-matter expert really shines, as you will have the tacit knowhow of that problem area. Additionally, we saw situations in our experiments where progress kind of plateaus once in a while, and having an expert steering the community of autoresearchers is a big unlock. The PIs in universities or in research organizations, imo, should adopt this hat and be the shepherd in this post-AGI version of learning and research.