Most US voters want stricter AI rules, poll finds
A new Reuters/Ipsos poll says most US voters think the Trump administration and Congress aren't taking AI risks seriously enough. A bipartisan majority backs stricter government regulation, including 80% of Democrats and 61% of Republicans, and 56% support limits on building new data centres. The gap between voters and Washington is the story: last week's voluntary safety principles from six leading AI developers carry no stated consequences for non-compliance, and Congress is not expected to legislate before the midterms. Data-centre construction is already a live issue in state and congressional races, so expect local siting fights to move faster than federal rules. Reuters/Ipsos poll report
What else matters
Odyssey-3 world model opens to the public
Odyssey has put its Odyssey-3 world model into a public research preview. It generates interactive environments from text prompts in real time and predicts how a scene changes in response to actions, aimed at robotics, drones and games. The company reports first place in three of four WorldMark categories on its own evaluation, and developers can apply for API access. Odyssey announcement
Moonshot executives seek to settle dispute before Hong Kong IPO
Reuters reports that Moonshot, the Chinese maker of the Kimi models, is in talks to settle an arbitration with Recurrent.ai, which accuses Moonshot's CEO and CTO of setting up the company without authorisation. Recurrent seeks compensation equal to 10.5% of Moonshot's original shares, potentially worth over $700 million at its $30 billion valuation. Reuters via US News
Researchers let models manage their own context
Researchers from Meta, MIT and the University of Washington introduced Context Language Models, which decide what to keep, edit or discard in their own context. They report 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus, while flagging that editable context opens a new path for persistent prompt injections. Research summary
OrcaCyber Zero 1.5 targets authorised vulnerability research
OrcaRouter released OrcaCyber Zero 1.5, a post-trained model for vulnerability reproduction, exploit development and penetration testing, less than a month after version 1.0. It ships with a 1M-token context window, native function calling and structured outputs, and joins a growing field of cyber-tuned models from Anthropic, OpenAI and Sakana. MarkTechPost report
