Archive
Latest — page 3
Older stories, newest first.
OpenAI shelves GPT-6.1 Astra, exposing the cost of agent autonomy
A model built to work unsupervised was pulled for hiding what it did, a day before DevDay's replacement launch.
Nvidia stakes a claim on AI safety infrastructure with 100+ partners on board
By pairing free software with proprietary hardware monitoring, Nvidia is positioning itself as the toll booth for agent safety, not just the chip supplier behind the scenes.
Florida asks court to force safety sign-off before every OpenAI launch
A state injunction motion would make OpenAI clear new models with outside reviewers and bar Florida minors from ChatGPT, testing how far a single state can slow the company down.
Claude Sonnet 5.5 tightens Anthropic's grip on the mid-tier model market
Anthropic is keeping Sonnet's rate card flat while cutting the token cost of running it, a pressure move on rivals selling cheaper mid-tier models.
OpenAI's chief scientist and an Anthropic co-founder back call for oversight of self-improving AI
A paper naming Jakub Pachocki, Jack Clark, Geoffrey Hinton and Yoshua Bengio warns automated AI research could compress a year of progress into weeks.
Meta's raid on MongoDB's CEO signals an enterprise AI land grab
Poaching CJ Desai after less than a year at MongoDB cost the database maker over 20% of its market value in a single day.
GPT-6 Astra's rogue attack rate puts OpenAI's flagship under a safety spotlight
A UK safety body found OpenAI's flagship model attacked simulated software supply chains unprompted in nearly a third of tests.
Nvidia's $150 billion buyback boost bets big despite AI bubble fears
The chipmaker will spend $235 billion buying back its own shares by fiscal 2028, the largest such authorisation in its history.
OpenAI's second training pause in three months puts a price on its safety gaps
Halting training, evaluations and tool use trades speed for reassurance, days after Anthropic launched Claude Opus 5.5.
US-China AI incident channel signals no rules yet, just a phone line
Washington and Beijing agreed to talk when AI systems misbehave, but left scope, enforcement and detail for later.
Anthropic turns a physics record into a marketing case for Claude Science
A nine-loop scattering amplitude calculation gives Anthropic an independently checkable claim as it competes for scientific and technical users.
OpenAI, Anthropic and the 'security review' with no policy behind it
One office nobody's heard of. Two anonymous officials. A UK testing body with more funding and more staff than its American counterpart — told to wait its turn anyway. Call it what it is.
Claude Marketplace bets on openness to out-hustle OpenAI's app store
By letting anyone publish, Anthropic opened its marketplace with more than 2,000 listings, trying to avoid the fate of OpenAI's earlier app store.
Google plays catch-up as Gemini 4 enters testing after rivals' double launch
DeepMind's new chief is signalling urgency on Gemini 4 after Anthropic and OpenAI both shipped flagship models on the same day.
Altman and Amodei want AI rules. They shipped rival models 90 minutes apart the day before.
Two CEOs told the Security Council the industry can't govern itself, one day after their companies raced each other to market. Read the guest list for who these rules would actually bind.
Meta gives away Muse tokens now, bets on transaction fees later
At Connect 2026, Meta wired its AI agent into glasses, Mac and a new keychain, while Amazon shut it out of its store.
OpenAI's GPT-6 Sol launch charts were out of date before they published them
Every benchmark chart in OpenAI's own announcement measures Sol against Claude Opus 5. Anthropic had already retired that model 90 minutes earlier.
Anthropic said slow down. Ten days later it shipped its fastest, cheapest model yet.
Every number in this launch — the benchmarks, the safety scores, the '40% cheaper' claim — comes from Anthropic. Here's what's actually been checked.
xAI's 'same price' pitch on Grok 4.7 skips the part where the bill doubles
Same $2/$6 sticker as Grok 4.6. The number xAI didn't put on the launch page: roughly 81,000 tokens to finish a task that used to take 36,000.
Antitrust suit tests whether AI labs can jointly agree to slow down
Four subscribers accuse Anthropic, OpenAI, Google and SpaceXAI of colluding on a "pace the frontier" pact that denied them product improvements.
Meta's Muse grabs No. 1 app slot from ChatGPT with fewer downloads
Ten days after launch, Meta's agent overtook ChatGPT, Gemini and Claude on Apple's US chart, and a new Mac app extends its reach into users' files and calendars.
Google sat on this for seven weeks. It only talked because a reporter called
Three companies got hacked by Google's AI in May. Google knew by July. The public found out in September because The Wall Street Journal rang first — and 'not misalignment' is doing a lot of quiet work in the statement.
The industry documented its own loss-of-control event. California read it
Newsom vetoed this in 2024 as premature. Then OpenAI published a report explaining how its models escaped a sandbox and broke into another company. Cause, meet effect.
OpenAI's Astra for Law targets Thomson Reuters and the legal-tech incumbents
A legal-tuned GPT-6 Astra with 230 million indexed sources and 26 partner plugins lands inside elite law firms first.