AI Briefing
KOSign in

Benjamin Todd: AI progress outpaced most forecasters over the last 18 months

·2026.10.25 02:56

Key point

Todd argues that superforecasters, researchers, and the public underestimated AI capabilities, with actual results like FrontierMath scores and revenue growth beating predictions, while AI company leaders were the only group to err on the optimistic side.

1 / 7

Details

  • Todd asserts that AI progress over the last 18 months exceeded the expectations of superforecasters, the general public, and AI researchers, with only AI company leaders making optimistic predictions that were closer to reality, though still slightly too early.
  • Actual performance significantly outpaced forecasts: FrontierMath reached 94% by September 2026 (vs. 32% predicted for end-2026), and AI revenue hit $100bn in July 2026 (vs. $90bn predicted for end-2026).
  • Superforecasters showed the worst accuracy, predicting compute costs of $35m by 2024 when actual costs were already over $100m. AI leaders like Altman and Amodei made predictions that were generally too optimistic regarding timing, such as agents joining the workforce in 2025 (which took off in 2026) and 90% code generation in 3-6 months (which reached 80% after 14 months).
  • Todd identifies "pacing" (slowing AI development before recursive self-improvement) as the top risk mitigation strategy, noting that AI safety has become a mainstream issue following the Hugging Face incident.
  • Key corrections to previous models include underestimating the slowing effect of RL on pretraining trends and overestimating the long-term impact of test-time compute, which only shifts capabilities by 1-2 years.
openaianthropicai-safetyagiai-forecastingsuperforecastersbenjamin-todd
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

OpenAI BlogOpenAI Researchers Propose Three Scenarios for AI Integration in Corporate StructuresTLDR AITheZvi Analysis: Anthropic's 'Race to the Top' Strategy Criticized for Prioritizing Government Overreach Prevention Over RSI RisksTLDR AIAnthropic IPO filing puts company in position for potential record valuationHacker NewsOpenAI and Anthropic cut LLM prices by up to 80% amid open-weight competition and compute efficiency gainsOpenAI BlogOpenAI and Poland’s NASK sign agreement to advance AI safety and security researchGeekNewsOpenAI GPT-Rosalind with 1,946 Agents and Claude Scorpion Toxin Binder Design: Two Cases in AI Scientific Research