AI Safety

RL on Forecasting Could Yield Superhuman Decision-Making AI

A new approach beats coding-focused RL by simulating live internet access for forecasts.

Deep Dive

Oxford researcher amitlevy49 argues that applying reinforcement learning to forecasting—not just coding or math—could create superhuman forecasters. Initial RL attempts plateaued because performance was bottlenecked by limited pre-generated context. The key insight was to build a time‑masked RL environment where the model uses tools like search (via Wikipedia dumps, AskNews), web fetch (via the Wayback Machine), APIs, and a coding tool for simulations, as if accessing the live internet. This approach overcomes the information bottleneck and, the author believes, could let us reach superhuman forecasting.

Key Points
  • Initial RL on forecasting plateaued due to limited pre-generated context summaries, capping performance at frontier LLM levels.
  • Breakthrough uses a time-masked RL environment with tool calls (search, coding, APIs) to simulate live internet access for training.
  • Superhuman forecasting could directly improve decision-making in policy, finance, and strategy, unlike coding RL which only aids software.

Why It Matters

A superhuman forecaster would empower organizations to make dramatically better decisions, potentially improving civilization-wide outcomes.

📬 Get the top 10 AI stories daily