Today on AI:AM — “RSI Gets Real, the Context Bet, and the Benchmark Anthropic Fails.”
Prakash Narayanan and Nathan Labenz open with Fable, Recursive, token anxiety, and the way frontier models are changing the scale of work people are willing to delegate. The hosts frame the morning around a practical question: if the models can run longer, remember more, and coordinate more work, what parts of the organization and media stack get remapped first?
Andrew Moore (Lovelace AI) on context engineering — Moore argues that the next enterprise AI bottleneck is not simply bigger models or more compute. It is retrieval, recall, data corroboration, metadata-rich graphs, and the unglamorous work of organizing old data so agents can act safely in high-stakes domains.
prinz on legal AI benchmarks and governance — prinz walks through why legal research is a revealing testbed for model capability, why OpenAI’s unit-distance result matters, and why nationalizing frontier labs could concentrate dangerous state power rather than solve AI risk.
The close turns back to the week in AI: how contrarian benchmark graphs change the discourse, which models fit which jobs, why subscription products keep finding retention tricks, and how IPO liquidity could feed the next wave of venture-backed AI launches.
