Recursive Self-Improvement (RSI)

The loop has three steps. A system gets better at the task of building AI, uses that skill to produce a more capable successor, and the successor is better again at the same task. If each round makes the next one faster, capability could climb quickly. That is the “intelligence explosion” I. J. Good described in 1965, and the path that some expect to lead from AGI to superintelligence.

Whether that happens is disputed. Improvements may get harder to find at each step, and real progress still depends on compute, data and human-run experiments. Treat RSI as a forecast with conditions attached, not a description of current systems.

Today’s agents use narrower learning loops. An agent that writes memory or rules from past work improves its context, not the model’s underlying ability to improve itself. In episode 71, at 37:28, Conor asks how Slack supports “recursive self-improvement or learning loops if you want a more simplified version.” Slack’s Jaime DeLanghe answers at 38:14 that the archive teams build in Slack gives agents “a self-improving loop”: one agent already writes per-channel rules to get better at being tagged there. Slack doesn’t build that in yet. That is agent memory, the simplified version Conor names, not RSI in Good’s sense. Episode 71 credits Svix and Walrus Memory as Season 4 presenting sponsors in its opening read.

In episode 72, at 46:14, Conor asks whether these loops lead to a broader takeoff in model capability. AMD’s Anush Elangovan declines to predict. He argues that the bigger near-term change comes from applying today’s intelligence, held constant, to new markets. Episode 72 names Svix and Walrus Memory as presenting sponsors in its opening read.