Rayan Krishnan, co-founder and CEO of Vals AI, joins Theo Jaffee to discuss the RSI Index and the possibility of automating AI research. Rayan Krishnan offers a summer 2027 forecast for Anthropic based on proxy tasks, while Theo Jaffee challenges whether those measurements justify a claim of full recursive self-improvement. The date is a contested prediction in the interview, not an established outcome.
Rayan Krishnan describes five proxy tasks across pre-training, post-training and harness engineering, including work involving held-out benchmarks. Rayan Krishnan distinguishes question-answer exams from measuring professional work products, and discusses custom benchmarks that change with a company's workflows. Theo Jaffee presses on whether strong benchmark scores can still coexist with incomplete job automation.
Rayan Krishnan and Theo Jaffee examine the difference between improving a task and replacing a job. A scheduling example illustrates time saved without eliminating a doctor's role, while the discussion considers whether increasingly capable models could also accelerate their own adoption. The conversation leaves room for new bottlenecks and does not equate recursive improvement with an immediate fast takeoff.
Rayan Krishnan also describes concerns about unequal access to advanced agents, including the possibility that better-resourced people could win more negotiations. Rayan Krishnan outlines evaluation interests spanning frontier risks, scientific work and access to public services. Theo Jaffee questions both the economics of cheap intelligence and what it would mean for models to automate diffusion itself.
Watch on YouTube
