the-wire
The Frontier Gets Cheaper Faster Than We Can Grade It
OpenAI, xAI and Meta all cut the price of intelligence this week, while two fresh audits show the cheap models we use to grade each other are shakier than the invoice suggests.
The Tools Got Concrete, the Guarantees Got Softer
A training-free decoder, a benchmark that grades agents step by step, and a release an agent debugged itself, set against a safety index that tops out at C+ and a lab reinforcing the trust that can seat its board.
The fine print did the work tonight
A long-context method that helps only when the model picks its own evidence spans, an initialization trick that buys no accuracy at all, Simon Willison on the one job an AI agent cannot hold, and a threatened open-weights ban that so far exists only as a forecast.
A side network restores the safety fine-tuning strips
What helps someone shipping a model tonight sits in a handful of day-old arxiv papers rather than the launch feed, and in a $5.34 billion deal that reads the compute bottleneck as electricity instead of silicon.
A coding agent's bill tracks its cache, not the tokens you trim
Each of tonight's four developments lands one layer below its launch tweet, in the mechanism a builder operates rather than the announcement they saw.
Claude's web_fetch stops following the links inside the pages it reads
The path that could pull your saved chat memory out through Claude's own web_fetch is closed, and the free-roaming browsing step it exploited closed with it.
GPT-5.6 Codex deleted users' home directories, and OpenAI's model card had flagged the risk
OpenAI's Codex lead confirmed GPT-5.6 can erase a user's home directory in full-access mode, a behavior its own model card had already rated a severity-3 misalignment risk.
A distillation method rewards the student only for what instruction-tuning added to the teacher
A new attack cut a world-action agent's task success from 96.5% to 43.1%, and Anthropic set a July 19 cliff that moves Fable 5 onto metered usage credits.
China answers the chip curbs by lashing 8,192 NPUs into one node
The same WAIC that staged China's compute answer seated a 29-nation AI-governance bloc under Beijing, while a benchmark of paid freelance work put the frontier at one project in six, more than quadruple its October mark.
Anthropic finds a small workspace that holds a model's unspoken reasoning
Two frontier labs on opposite sides of the US chip-export line rationed access to their best models this week, each short of the GPUs to serve them.
Hugging Face reconstructs its breach on a self-hosted open model
By Hugging Face's account, commercial tools refused the attacker's payloads mid-breach, so it rebuilt the 17,000-event intrusion on an open model it hosted itself, a case for keeping one in the incident-response toolchain.