Thoughts on AI progress (Dec 2025)

Patel opens with a puzzle about a combination of beliefs rather than about a fact. Why do some people hold short timelines and simultaneously feel bullish about the current scale-up of RL on top of LLMs? His argument is that these two positions sit badly together: if we are genuinely close to a human-like learner, then the whole approach of training on verifiable outcomes is doomed.

The reason is what the labs are currently doing. They are baking skills into models through mid-training, supported by an entire supply chain of companies building RL environments that teach a model to navigate a browser or use Excel to build financial models. Patel frames the fork sharply: either models will soon learn on the job in a self-directed way, which makes all this pre-baking pointless, or they will not, which means AGI is not imminent. The premise is that humans do not go through a special training phase rehearsing every piece of software they might ever use.

He credits Beren Millidge with the sharper version of the point: when frontier models improve on benchmarks, the explanation is not only scale and clever research but billions of dollars paying PhDs, MDs and other experts to write questions and example answers targeting those precise capabilities. Benchmark progress and capability progress are not the same measurement.