What Anthropic's new finding on Claude shows (and what it doesn't)

By James O'Donnell (The Download, MIT Technology Review) · 14 July 2026.

This edition of The Download focuses on Anthropic's recent claim to have found a new window into the "internal thoughts" of its models as they reason to produce a response. According to the newsletter itself, when Anthropic announced this finding last week, journalist James O'Donnell decided to discuss it with Will Douglas Heaven, the publication's senior AI editor, who also holds a PhD in computer science and has spent time investigating what can really be said about the inner workings of AI models.

The email does not develop in its body the technical details of Anthropic's finding or Heaven's specific conclusions on what this discovery does and does not show; for that it points to the full article published on technologyreview.com, drawn from its weekly newsletter The Algorithm. It is therefore not possible to reproduce here a detailed analysis of the research without speculating beyond what the email states literally.

As an additional detail from the newsletter itself, a live LinkedIn event is announced for that same day at 12:30 PM ET, in which Will Douglas Heaven will talk with Sam Sinha, founding AI researcher and head of world models at 1X Technologies, about how this technology could transform robotics and help unlock a new generation of intelligent machines, given that current AI systems skillfully generate text, images and code but still struggle to understand the complexity of the physical world.

The rest of the email consists of an advertising block from Elastic on the fundamentals of AI architecture (AI-ready data, context engineering, governance and observability, and human expertise) and a "must reads" section with general technology news —a data center moratorium in New York, a drop in smartphone shipments, sugar molecules found in interstellar space, Nvidia's restrictions on chip sales in Asia, Russian hackers targeting routers and Trump's cryptocurrency moves— that are not focused on agentic AI or on the main topic of this edition.