Simplified illustration of AI progress as a multi-step evolutionary process, accompanying this blog post. The process centers around agent development as the mutation method. Candidate agents then pass through iterative rounds of selection: first inside the lab developing the agent (lab optimisation), then in public by the users of the agent (public optimisation). The final output is the inference compute allocated to the agent program, correlated with real-world impact and opportunity to influence the design of future agent generations. Originally hand-drawn on an iPad, then made interactive with AI (Fable 5.1, Opus 5.5).