Ajeya Cotra's 2026 AI forecast review: 'plan a birthday party' has already fallen
ajeya_cotra · x · 2026-10-02
AI researcher Ajeya Cotra made ten qualitative capability forecasts for 2026 in mid-January (five categories: gameplay, logistics, video, game design, math, each with an 80% low-end and 20% high-end prediction) and is scoring them five months early:
- Math: Her forecast that AIs couldn't place a paper in a top math journal has fallen amid AI cracking dozens of esoteric theorems, though she notes the situation was more ambiguous than assumed.
- Software tasks: Her 50%-time-horizon prediction already looked too conservative by early March.
- Games: The low-end forecast (AI beating Pokemon like a typical ten-year-old) is falsified — Claude Opus 4.7 beat Pokemon Red in May (112,000 steps), and Fable beat FireRed in 50 hours of gameplay in June, close to human wall-clock time.
- Latest update: The "plan a birthday party" forecast has basically fallen, and "plan a wedding" may be within striking distance, driven by major gains in computer use.
More from AGI Musings
- Terence Tao: Today's AI doesn't feel like intelligence, it's more like a trick — rohanpaul_ai · 2026-10-02
- Researcher JJitsev: Frontier Labs Fear Collapse of Their Trillion-Dollar Exclusivity Claim, Not Extinction — JJitsev · 2026-10-02
- Open weights will be good enough: why frontier AI smarts won't matter for daily life — Dan_Jeffries1 · 2026-10-02
- Human-like information processing is a tiny sliver of all possible systems — burny_tech · 2026-10-02
- Reddit asks: when AI denies you service, who do you argue with? — yi111 · 2026-10-02
- AAA games still cost hundreds of millions despite LLMs, as marketing eats half the budget — willcb · 2026-10-02