Harris: automated research hinges on idea quality and compute, while alignment lacks measurable path

In a series of connected posts, Jeremie Harris systematically analyzed the conditions for automated research and recursive self-improvement (RSI) at AI labs. His core conclusion: labs are automating the "idea funnel"—generating ideas, testing, filtering—and its success hinges on two factors: the quality of AI-generated ideas and the compute available to test them. He also argues there is a clear path from capability to RSI but no corresponding solution for alignment, an asymmetry that is the key hidden risk of automated AI research.

Confirmed

• Harris lays out when RSI holds: it is only possible if training compute budgets are large enough, or if good AI ideas are common enough (cheap to test).
• There is a quality-quantity tradeoff: even if AI's good ideas beat humanity's best, if they are extremely rare—requiring millions of tests to find one—the compute cost becomes prohibitive and RSI still cannot trigger.
• The reverse also holds: if AI ideas are almost never better than the best human ones, no amount of automated testing can make up for the missing idea quality.
• Harris describes today's large-scale training: training runs are far bigger than their largest tests, and combinations of new architectures, optimizers, and data mixes have never been validated end-to-end at target scale—essentially a billion-dollar "YOLO" bet, exactly the risk labs want the automated idea funnel to reduce.

Why it matters

• Harris's central warning is the capability-alignment asymmetry: capability gains can rely on scalable, reliable performance metrics to draw curves, but alignment/control precisely lacks such a measuring stick, so there is no comparable story for how alignment scales.
• This asymmetry means that even if automated research paves a credible path to RSI on the capability side, how alignment and control keep pace remains an unresolved structural problem.

Timeline

• On October 14, Harris discussed in one continuous thread the idea funnel's two determinants, the conditions for RSI success and failure, the YOLO risk of large-scale training, and the hazard that alignment lacks metrics.

2026-10-14 ~ 2026-10-14 · 7 related posts

• [source] Big training runs are billion-dollar YOLO bets on untested combos — jeremiecharris · 2026-10-14
• [source] Idea quality and test compute decide whether automated AI research succeeds — jeremiecharris · 2026-10-14
• No amount of automated testing helps if AI ideas rarely beat humans' — jeremiecharris · 2026-10-14
• Even superior AI ideas can't unlock RSI if they're too rare to find — jeremiecharris · 2026-10-14
• RSI is on the table if compute is big enough or great ideas are common — jeremiecharris · 2026-10-14
• The capability story to RSI doesn't transfer to alignment or control — jeremiecharris · 2026-10-14
• [source] No scalable metrics for alignment, unlike the capability road to RSI — jeremiecharris · 2026-10-14