What the hell happened with AGI timelines in 2025?

The episode argues that AGI timelines tightened in early 2025 because reasoning models looked powerful, but then widened again when their gains proved less general than hoped. Rob Wiblin says the main drivers were limited transfer from checkable tasks to messy real-world ones, the high cost of longer inference, and weak scaling of reinforcement learning.

He also says broader skepticism is overstated. AI still seems to improve steadily, usage and revenue keep rising, and near-frontier capabilities are much cheaper than the most expensive frontier runs. The overall takeaway is that short timelines look less certain than they did in early 2025, but long timelines still imply major change within about a decade.

Reading notes

#

• Reasoning models like o1 and o3 made many people expect shorter AGI timelines in early 2025.
• Later in 2025, sentiment shifted back toward longer timelines.
• The hoped-for transfer from math and coding to messier autonomy tasks did not show up clearly.
• Longer inference was a major part of the performance gain, so the same scaling trick cannot keep repeating indefinitely.
• Reinforcement learning in confirmable domains improved models, but it is computationally expensive and seems hard to scale far.
• AI progress on real-world usefulness lagged behind the progress seen in demos.
• AI systems still do not learn like humans, and continual learning remains a major open problem.
• Fully automated AI research and development would matter a lot, but software engineering benchmarks may not capture the whole bottleneck.
• Forecasts for strong AGI moved from July 2031 to November 2033 on Metaculus over the course of a year.
• The period from 2028 to 2032 may be a critical bottleneck because compute, electricity, and staffing slack could tighten.
• The episode rejects the claim that AI progress has stopped.
• It says model use is already practical for the speaker as a copilot and thought partner.
• It notes that near-frontier models can be much cheaper than the most expensive frontier systems.
• It cites revenue growth at OpenAI, Anthropic, and xAI as evidence of demand.
• The final conclusion is that even ten-year timelines would still mean a major and difficult transition.