From June to August 2026, podcaster and columnist Dwarkesh Patel published a series of discussions on the AI frontier. The core themes centered on the shift in training paradigms from pretraining to "on-the-job learning," the recursive self-improvement that comes with AI automating AI research, and the trend of compute resources concentrating in a handful of labs. Among them, Dylan Patel predicts that Anthropic and OpenAI will control most of the world's compute by 2028.

Compute centralization: the 2028 outlook

In an article published on Aug 25, Dylan Patel offered a clear judgment on the compute landscape. The article's summary reads: "Every force is screeching towards centralization." The phrase points to AI compute resources accelerating their concentration in leading labs.

The judgment aligns with the topic note's lead that "Anthropic/OpenAI will have most of the world's compute by 2028," but the source material does not provide the specific methodology or data sources; those details await further disclosure in the original text.

Recursive self-improvement: the debate over AI automating AI research

On Aug 11, Dwarkesh Patel published a conversation with Ryan Greenblatt titled "What happens once AI can automate AI research?" The article's summary characterizes it as "A debate about recursive self-improvement."

Recursive self-improvement refers to AI systems participating in the development of improvements to themselves or to subsequent AI systems, potentially creating an accelerating capability loop. The conversation presents arguments for and against whether this mechanism holds and what its impact would be, but the source material does not disclose the specific points made by either side.

Continual learning: on-the-job learning and the timing of regulation

A June 26 article, "The next big breakthrough will be AIs learning on the job," argues that "Labs are throwing away the most valuable data." The view points to a possible shift in training paradigms from large-scale pretraining to models learning continuously on real tasks.

An Aug 7 article, "8 Predictions for the Era of Continual Learning," extends the discussion, with its summary stating plainly: "Locking in AI safety regulation now is a mistake." The piece discusses the timing of regulation against the backdrop of a shifting continual-learning paradigm, but the source material does not elaborate on the eight predictions.

What to watch

  • The basis and methodology behind Dylan Patel's 2028 compute centralization judgment;
  • The specific disagreements between Ryan Greenblatt and Dwarkesh Patel over recursive self-improvement;
  • The practical impact of the "on-the-job learning" paradigm on data use and model training pipelines;
  • How AI safety regulation should time and shape its approach in the era of continual learning.