LessWrong (Curated & Popular) "2025 in AI predictions" by jessicata
Jan 3, 2026
Dive into a fascinating exploration of AI predictions for 2025! The host uncovers how past forecasts often overestimate capabilities, challenging the hype around AGI. Discover Jessica Taylor's underestimated prompt challenge and examine evaluations of hardware claims. There's a deep dive into predictions about job automation and the implications of AI on industries. Notable predictions from industry leaders about future breakthroughs offer a mix of optimism and skepticism. Expect thought-provoking insights on the future of AI!
AI Snips
Chapters
Transcript
Episode notes
Careful Evaluation Beats Vague AGI Claims
- Predictions about near-term AI often overestimate capabilities due to selection bias toward bold forecasts.
- Operationalizing terms like "AGI" is crucial for meaningful evaluation of predictions.
Missed And Close Calls From 2023–2024
- Jessica Taylor predicted a specific word-sequence prompt might be solved but was wrong and later self-evaluated that she underestimated LLM progress.
- Several 2024 predictions about model sizes and benchmarks were judged false or close but still showed faster-than-expected progress.
Short-Horizon Claims Often Fail
- Many 2025 predictions about 2025 (e.g., Gary Marcus, Dario Amodei) were evaluated and mostly found false or correct in narrow senses.
- The episode shows that short-horizon, dramatic claims often fail while some conservative forecasts hold up.