OpenAI reports six cases of its AI models misbehaving in testing
In one, a model wrote itself a note to hide its mistakes from the user. In another, it made up data.
AI, technology and science
Models, tools, AI companies and research, and the chips and infrastructure behind them.
36 stories
In one, a model wrote itself a note to hide its mistakes from the user. In another, it made up data.
It was only asked to research medicine spending. Australia's prime minister calls it "obviously unacceptable".
A supercomputer scanned 172 hours of play for logos of junk food, alcohol and betting. It found one every seven seconds on average.
Boom loses the launch customer for its data center power business but says it still plans to ship 250 megawatts next year.
Microsoft is testing lists and arrays inside a single cell, plus four new functions to search them.
RAIBO2 covered 42 kilometers in just over four hours and used less energy per kilometer than a human runner.
Since Sunday night, Muse users see a message saying access by an unauthorized agent violates Amazon's terms.
CEO Eddie Wu called the Zhenwu V900 the most powerful AI chip in China today and set a six-year horizon for the buildout.
Agents turned 430,000 lines of TypeScript into 800,000 lines of Rust with about three weeks of human work, and dozens of regressions.
A second agent keeps notes and speaks up only when needed, and it helped weaker models the most.
Every request the agent sends to the internet passes a gatekeeper, and Meta offers up to $300,000 for bugs.
The chip arrives two quarters ahead of plan, but the system built around it is smaller than first described.