AMD agrees to buy Fei-Fei Li’s World Labs for $8.2 billion

AMD and World Labs. AMD has agreed to buy World Labs, the startup co-founded and led by Fei-Fei Li that builds AI models of interactive 3D environments. The price is about $8.2 billion, paid entirely in AMD stock, according to AMD’s Form 8-K dated Monday. The deal has not closed. AMD says it expects it to close by the end of 2026, subject to regulatory approvals and other conditions. The number of shares will be set by AMD’s volume-weighted average stock price over the 10 trading days ending two trading days before closing, the filing says. After the deal closes, Li will become AMD’s executive vice president and chief scientist, reporting to CEO Lisa Su, AMD says.

Thomson Reuters v. Ross. The U.S. Court of Appeals for the Third Circuit on Tuesday affirmed a trial-court order in Thomson Reuters’ copyright suit against Ross Intelligence, a legal-research AI startup. The one-page judgment (No. 25-2153) says the order “is hereby AFFIRMED.” The 32-page opinion is under seal. In the February 2025 ruling under appeal, Judge Stephanos Bibas rejected Ross’s fair-use defense for using material derived from Westlaw headnotes to train its AI search tool. He wrote that Ross’s AI “is not generative AI.”

GLM-5.3. Anthropic, which sells competing models, said Tuesday that GLM-5.3, an open-weight model from China’s Zhipu AI, or Z.ai, can build working cyber exploits and was released “without meaningful safeguards.” In Anthropic’s tests, it built full exploits of known flaws in Chrome’s V8 engine in 50 of 410 tries, versus 56 for Anthropic’s Claude Mythos Preview. Simple techniques beat its safeguards 64% to 100% of the time in simulated tests, Anthropic says, adding that defenders can use such abilities too. On Sept. 17, NIST’s Center for AI Standards and Innovation called GLM-5.3 “the most cyber-capable open-weight model released to date,” but said its cyber capabilities are “significantly lower than those of current U.S. frontier models.”

Evening update

AI companies could be held liable for recklessly designed agents, and charged with crimes, under a bill Sen. Josh Hawley says he is introducing

AI companies could be held liable, and in some cases criminally charged, over the actions of their AI agents under legislation Sen. Josh Hawley says he is introducing. The Missouri Republican described the plan in an op-ed that his office posted Tuesday and says ran in The Washington Post. He wrote that the bill would make companies liable for their agents’ actions when they design the agents “in a reckless fashion.” It would also make users liable when they deploy agents “in a reckless way.” Hawley wrote that federal anti-hacking laws should clearly apply to AI companies. He said the bill would let prosecutors charge companies “when the firms know their agents are capable of criminal conduct and fail to create reasonable safeguards.” Hawley argued that letting people harmed by AI sue the companies would change how the labs behave. As of 4:47 p.m. CT Wednesday, his office had not posted bill text, a bill number or cosponsors.

Earlier coverage

• Trump and six AI company leaders signed a White House pledge Tuesday to put frontier models through “four layers of controls and audits.”
• OpenAI began rolling out always-on ChatGPT agents it calls dots to Pro and Business Premium users Tuesday, the company said.
• Trump signed an executive order Tuesday making America.gov a “conversational point of entry” for federal services.
• Florida’s attorney general asked a state judge, in a motion not yet granted, to bar OpenAI from developing AI models without independent approval.
• Amazon’s Bedrock service now offers SpaceXAI’s Grok 4.7, AWS said Monday.

— Lark Pemberton