Thoughts on AI Progress (Dec 2025)
Source: https://www.dwarkeshpatel.com/p/thoughts-on-ai-progress-dec-2025 Author: Dwarkesh Patel Date: 2025-12-03
Summary
Dwarkesh Patel’s end-of-2025 assessment of AI progress. Known for interviewing frontier researchers; his views synthesize many private conversations. Key question: is the pace of fundamental progress slowing or accelerating?
Key Claims
- Capability progress in 2025 was “real but uneven” — massive gains in coding/math, slower gains in general reasoning and world knowledge.
- The “pretraining cliff” concern: the data that pushed GPT-4 → o1 was different from GPT-3 → GPT-4. The next step requires new techniques, not more data.
- Inference scaling (o1-style) is the clearest path to continued progress — but it has diminishing returns for some task types.
- Most underrated development: the quality of open-source models (Llama, DeepSeek) catching up to closed-source much faster than expected.
- Prediction: 2026 will see the first genuinely useful autonomous research agents (weeks-long tasks, not just hours).
- Concern: economic translation of capability gains is lagging — most capability improvements aren’t yet captured in GDP.
Entities
- Dwarkesh Patel — Dwarkesh Podcast
Concepts
- Scaling & Compute — pretraining cliff thesis
- Test-Time Compute — inference scaling as primary near-term lever
- Autonomous Research — 2026 prediction for weeks-long agent tasks