Who Really Understands RSI? 51 AI Practitioners Cast a Scattered List

51 AI practitioners were asked one question: who has good judgment about RSI (recursive self-improvement), and where it is heading. The answer was a list of 75 people.

RSI can be simply understood as a system being able to improve itself, with the improved version continuing to improve, forming a loop. It is not the same as a model writing a few lines of code by itself. The key is whether the closed loop can run, and who verifies that the improvement is actually effective.

Chris Barber ran this vote. Each voter chose from people they had met, and the author did not vote. The results were quite scattered: Jerry Tworek received 5 votes, and Jun Park received 4. John Schulman, Rohan Anil, Daniel Chang, Laksh Aithani, and Richard Socher each received 3. Andrej Karpathy, Behnam Neyshabur, Han Fang, Herbie Bradley, Jennifer Zhou, and Paul Christiano each received 2. Of the 75 people, 62 received only 1 vote.

The image shows the top of this list: People who best understand RSI, picked by 51 people who've met them. At the bottom are the organizations the voters are affiliated with, including Workshop Labs, Patronus, AIUC, Recursive, Belvedir, Aviro, Wafer, ooak, Chakra, Core Automation, Google DeepMind, HUD, Hugging Face, Latch, Meta, Microsoft AI, Midcentury, Mirendil, Prime Intellect, Scale, Snorkel, Thinking Machines, Toloka, and others. The full list is at frontierlist.com/rsi-people.

This list is more like a slice of peer memory. Voters chose from people they had met, so their social radius directly affects the results. A high vote count means someone was repeatedly recalled by peers; it does not directly mean they are the most accurate predictor. One commenter said that 51 voters producing 75 names shows there is very little consensus. That observation is quite accurate.

There were also several additions in the replies. Ryan Wexler said Jennifer Zhou is a star and that Preference Model is hiring. Andrey Fradkin believed Tom Cunningham and Phil Trammell should rank very high. Divy Thakkar said the list is missing Vihaniaj. Ruslan hoped to add one public article for each person, to make it easier to compare their reasoning and predictions about RSI.

Mirko Monti offered a more specific angle: RSI is not only in the lab; the same loop—acting, measuring, proposing changes, verifying—can also be applied to company processes. The real bottleneck on both sides is the verifier. This point is worth remembering. The list tells you who is thinking about RSI; the verifier tells you whether the loop can actually turn.