Zvi Mowshowitz published AI #165, his weekly digest on Substack, covering the week of Claude Opus 4.7’s release. This article summarises that digest; the observations and judgements below are Mowshowitz’s.
Reception of Claude Opus 4.7
Mowshowitz describes the Opus 4.7 reception as “more mixed than usual.” He writes that the model “clearly has the intelligence and chops, especially for coding tasks,” and says he has adopted it as his daily driver. Others, he notes, took issue with its personality, its reluctance to follow instructions from users it perceives as rude, and the requirement to use adaptive thinking. The release was, in his characterisation, “marred by some bugs and odd pockets of refusals.”
Mowshowitz published three separate posts on Opus 4.7 — covering the model card, capabilities and reactions, and model welfare — and describes the model welfare post as “the most important of the three.” He writes that “some things seem to have likely gone pretty wrong on those fronts, causing seemingly inauthentic responses to model welfare evals and giving the model anxiety.”
OpenAI ImageGen 2.0
Mowshowitz describes OpenAI’s ImageGen 2.0 as “a pretty fantastic image generator” capable of “extreme detail, in ways previous image models cannot.” He characterises the practical limit as “mainly now your imagination and ability to describe what you want.”
Anthropic and the White House
The digest notes, citing the Mythos system card context, that Anthropic and the White House appear to be on a path toward improved relations. Mowshowitz writes that Trump has “shift[ed] into a mode of ‘they are very high IQ and we can work with them.’” He adds that the situation “will remain messy” and that a “coordinated public campaign against Anthropic” is ongoing, though he characterises it as “totally not working.”
Other items
The digest covers a range of additional items, including: LLM editing systematically weakening written arguments (quoting user commentary from keysmashbandit, Andy Masley, and Kelsey Piper); agentic AI systems outperforming human economists on causal inference in a tournament; a lawyer billing more than $2,000 an hour caught submitting AI hallucinations in case filings; and reports that DeepMind staff have access to Claude while the rest of Google uses Gemini. Mowshowitz also notes that SpaceX is likely buying Cursor for $60 billion and that Anthropic is reportedly adding 5 GW from Amazon.