Anthropic research PM Alex Albert gave a detailed account of how frontier model development actually works from the inside, covering the full pipeline from user feedback collection through eval design to model training. The conversation is unusually specific: Albert explains how Claude Cowork is used internally to pressure test documentation, how the eval process for new models is structured at timestamp 17:05, and how character training differs from capability training at 21:02.
Two technical threads are worth watching in full. First, the adaptive thinking section at 04:58 describes how the team is building reasoning that shifts based on task context, not just a single inference mode. Second, the consciousness question at 32:39 is not a philosophical sidebar. Anthropic is reportedly doing active work on this, and Albert's framing of it as a research problem rather than a PR problem is the most substantive public statement the company has made on the topic.
The decision-making framework Albert describes at 11:15, treating reversible choices as essentially free, explains a lot about why Anthropic ships certain features incrementally while blocking others entirely. If you want to understand the logic behind what Claude can and cannot do, that section alone justifies the watch.
[WATCH ON YOUTUBE →]