Anthropic says it has found a new way to look inside its models while they work through answers. That is a useful claim, and a slippery one: peering at a model’s intermediate machinery is not the same as reading a mind, however tempting the branding may be.
MIT Technology Review’s James O’Donnell framed the announcement as a question of what Anthropic’s research does and does not prove. O’Donnell spoke with Will Douglas Heaven, the publication’s senior editor for AI, who has a PhD in computer science and has reported extensively on how much can be inferred about the behavior of AI systems.
The cautious read is the correct one. Anthropic described its work as a window into Claude’s “internal thoughts” as the model reasons through an answer. The confirmed fact is narrower: Anthropic has announced research that it says exposes some internal activity. The harder claim, that this maps neatly onto reasoning as humans understand it, still needs scrutiny.
MIT Technology Review is also turning to a related problem: how AI systems can deal with the physical world. Text, images, and code are now routine demo fodder for frontier models, but physical environments remain messy. Researchers cited by the publication argue that “world models” may be needed to narrow that gap.
The publication is hosting a LinkedIn Live session on the subject with Heaven and Sam Sinha, founding AI researcher and head of world models at 1X Technologies. The session is scheduled for 9:30 a.m. PDT, 12:30 p.m. EDT, and 5:30 p.m. BST, and MIT Technology Review says registration is free.
Other technology items on the docket
- The Wall Street Journal reported that New York has become the first U.S. state to enact a moratorium on large data-center construction, with the governor banning such projects for up to a year. The Verge reported that state lawmakers passed a bill that could go further.
- Reuters reported that global smartphone shipments fell 11% in the second quarter of 2026, reaching a 13-year low, as a memory-chip shortage pushed up prices.
- Nature reported that sugar molecules have been detected in interstellar space for the first time. New Scientist said the finding supports the possibility that ingredients for life may exist beyond Earth, while The New York Times reported that researchers used radio telescopes and data analysis to identify the molecules.
- The Financial Times reported that Nvidia has cut its Asia buyer list in half and introduced a whitelist of companies that passed stricter checks, a move Reuters tied to tighter Trump administration controls intended to stop AI chips reaching China.
- Ars Technica reported that the U.S. government warned users that Russian state hackers are targeting routers for spying and theft, and urged people to secure their devices.
- Reuters reported that Donald Trump moved crypto gains into stocks and bonds while urging others to buy more crypto, and said his crypto projects produced large gains for him while retail buyers suffered losses.
- The Los Angeles Times reported that the LAPD suspended its use of Flock Safety surveillance cameras over privacy concerns involving automated license plate readers. Engadget reported that Flock has also faced criticism over data sharing with state and federal officials.
- Wired reported that U.S. officials approved the launch of a space mirror designed to reflect sunlight to Earth as part of a disputed plan to help solar panels generate power around the clock.
Reuters also reported remarks from SoftBank CEO Masayoshi Son, who predicted at the company’s annual corporate conference in Tokyo that AI will overtake human intelligence by 2040. That is a forecast, not an engineering result. File it with the other expensive bets until the machines do more than impress conference stages.
This story draws on original reporting from MIT Technology Review.