The House voted 417-3 on data centre power as ChatGPT got advertiser agents
The House passed a Castor-Evans bill 417-3 on Wednesday, telling state regulators to consider making data centres and other large power users pay for their own grid upgrades, while OpenAI began piloting ads that open a conversation with the advertiser's own agent inside ChatGPT.
The House voted 417-3 to make data centres pay for power A bill from Kathy Castor and Gabe Evans passed under suspension of the rules on Wednesday, with only Summer Lee, Delia Ramirez and Rashida Tlaib voting against. It amends the 1978 Public Utility Regulatory Policies Act to require states to consider adopting federal standards that would make data centres and other large-load customers cover the grid upgrades they need. Senate Majority Leader John Thune has said it could clear the Senate before the elections, though that would likely take unanimous consent. Source
OpenAI is putting advertiser-run agents inside ChatGPT conversations OpenAI is piloting Sponsored Agents with a select group of brands, an ad format where a click opens a branded conversation with the advertiser’s own agent inside ChatGPT rather than a browser tab. Angi announced its place in the pilot on Wednesday, and says homeowners will be able to go from discussing a project in ChatGPT into its service request flow and on to a local tradesperson. Source
OpenAI is said to want $1.5tn after investors offered $1.2tn Forbes reports that investors approached the company about a new private round at $1.2tn and that OpenAI believes it should be worth at least $1.5tn, on the back of traction for Codex and its GPT-6 Astra and GPT-5.6 Sol models. That is days after Sam Altman said the industry needs to pace the frontier, and after he ruled out a listing this year. Source
Canada and Germany put C$150m and €100m into Bengio’s LawZero The two governments announced the commitments at the ALL IN conference in Montreal on Wednesday, to Yoshua Bengio’s safe-by-design AI non-profit. The Canadian money goes on research and engineering talent and computing capacity for LawZero’s Scientist AI; the German half funds work at a new German office and is subject to notification by the European Commission. Source
Google DeepMind launched a public institute to argue about AGI The DeepMind Institute has gone live under Shane Legg, Demis Hassabis and James Manyika, with four opening essays: one on reasoning transparency, one on economic policy for AGI, one on a new utopianism, and a framework for frontier AI that Hassabis wrote himself. The site describes itself as a platform started by researchers from Google and Google DeepMind, and says the essays are their authors’ ideas rather than official positions. Source
A judge ordered Musk’s companies to produce the Apple deal Judge Mark Pittman gave X and SpaceXAI until noon on Thursday to hand over any agreement with Apple behind their decision to drop it from their antitrust suit this week, which they did without explaining why. OpenAI, the only defendant left in the case, says it was not party to the talks and asked for the document because the terms could bear on its own defence. Source
ByteDance spun out its AI drug unit at $1.5bn Anew Labs, the Shanghai drug-discovery arm ByteDance has spun off, closed $290m in its first outside round at a $1.5bn valuation, co-led by HSG and IDG Capital with Hillhouse alongside. ByteDance keeps 56 per cent of it; the unit builds structure-prediction and protein-design models out of Shanghai, Singapore and San Jose. Source
Google shipped two Gemini 3.8 voice models on Tuesday Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking went out through the Gemini API and AI Studio, with the base model in Search Live and Extended Thinking reaching Gemini Live and Workspace. The Extended Thinking variant can carry on reasoning while it is still speaking, and Google says it took the top overall place on Artificial Analysis’s Speech to Speech Quality Index at 82.6. Source
An OpenAI veteran shipped a model that writes nothing TypeSafe AI came out of two years of stealth on Tuesday with Jev, a “System One” model that takes an unstructured state and returns a typed decision with a confidence score instead of generating text. Founder Diogo Almeida says he helped build the instruction-following methods at OpenAI that became the research behind ChatGPT; the launch demo runs a bot through Doom game states at ten queries a second. Source