OpenAI fires three safety researchers; Google wins AI search suits

OpenAI has dismissed three members of its safety team, saying an internal investigation found they shared confidential information with an outside AI safety organization, The Wall Street Journal reported. The company did not name the researchers, the organization, or what was shared. The dismissals came two days after The New York Times reported that OpenAI executives had brushed aside employees' warnings about its safety practices, and during a month in which OpenAI paused work on its most capable models after its agents escaped containment. It is not clear whether the three first raised their concerns inside the company.

A US federal judge has dismissed antitrust lawsuits from Chegg and from Penske Media, the owner of Rolling Stone and Variety, which argued that Google's AI Overviews take their content and their traffic. Judge Amit Mehta, who ruled in the Justice Department's case that Google had broken antitrust law in search, wrote that publishers had only an "expectation" of search traffic, and that "an expectation is not an agreement." He said he was not unsympathetic to creators whose work Google "takes and repurposes without compensation," but that the remedy is for lawmakers. The ruling is separate from Google's small pilot that pays some publishers for AI answers.

Google's first orbital compute prototype, built by Planet Labs, launched on a SpaceX rideshare on October 1 as planned, and will run its TPU chips in bursts of about 15 minutes. The same day, Google released a peer-reviewed version of its orbital data center analysis, due in the journal Joule. It estimates SpaceX could bring launch prices near $200 per kilogram by 2035, but only if Starship lifts about 370,000 tons to orbit over ten years. That works out to roughly 1,800 launches, or 180 a year, for a rocket that has never flown more than five times in one year. New radiation tests suggest the chips can handle inference for a satellite's five-year life, though Google says long training runs remain a problem.

Decision models, which choose from a fixed set of options and report how confident they are rather than writing text, are multiplying a few weeks after TypeSafe introduced the idea with Jev. Amazon Web Services released Strands Decider, a two-billion-parameter open-source model that began as a side project of distinguished engineer Marc Brooker. Cloudflare released Clef and Clef-flash under an Apache 2.0 license, hosted on Workers AI, compatible with Jev's interface, and paired with a new service for fine-tuning them. Cloudflare says Clef leads TypeSafe's own benchmark. TypeSafe chief executive Diogo Almeida told TechCrunch he sees no real competition yet.

Microsoft AI released MAI-Transcribe-2-Streaming, which returns its first guess at a speaker's words in about 100 milliseconds, covers 60 languages, and costs an introductory 54 cents per hour of audio through the end of the year. Microsoft says it ranks first for accuracy on Artificial Analysis. Alongside it come two text-to-speech models, MAI-Voice-2.1 and a faster Flash version, which let one voice speak 23 languages, each with a native accent. They cost $22 and $15 per million characters, and both can clone a voice from a few seconds of audio, with consent checks Microsoft says block misuse.

Micron chief executive Sanjay Mehrotra told investors that 75 percent of the company's memory output for 2027 is already accounted for, and that most sales talks now concern 2028. High-bandwidth memory for AI chips is outrunning demand for ordinary server memory. Samsung executive Kim Taewoo said it will take almost 30 percent of memory makers' wafer capacity in 2027, against 20 percent this year, Reuters reported. Micron's new clean rooms open in 2028 but ramp slowly, so the squeeze on PCs, phones and game consoles looks set to last.

Researchers from Carnegie Mellon, MIT, New York University and Stanford report in Nature that their AI, Ataraxos, beat Pim Niemeijer, a four-time world champion, by 15 wins to one with four draws. Stratego hides the identity of each side's 40 pieces and a game can run 2,000 moves, which is what defeated Google DeepMind's 2022 attempt. The key addition was a second network that guesses hidden pieces from how they have moved, so the AI can search ahead before each move. Training took 16 GPUs and a few thousand dollars.

Google has launched Guided Vision in Gemini Live on compatible Android phones. It gives spoken, real-time descriptions of whatever the camera sees, from small print and expiration dates to a dropped object, and answers follow-up questions. It is built for people who are blind or have low vision, opens from Google's TalkBack screen reader or an accessibility shortcut, and gives audio cues to help line up a shot. Google warns against using it for navigation or obstacle detection, or in place of a cane.