World models and warnings, Oct 28
Google showed what AI can now build; the rest of the industry spent the day listing what could go wrong.
Google DeepMind spent Tuesday showing what AI can now build. Much of the rest of the industry spent it listing what could go wrong. On October 27 DeepMind released Genie 4, a world model that turns a prompt into an interactive scene with its own sound, as a research preview for trusted testers. You can change a scene in real time with natural-language prompts. The model simulates physics such as aerodynamics and collisions, generates native audio including dialogue, and renders people who move and behave more realistically. TechCrunch reported that scenes stay consistent for up to two minutes, double Genie 3's limit, and DeepMind framed the work as a step toward AGI.
The warnings came from the industry's would-be giant. Anthropic's IPO prospectus carries a 93-page risk-factors section, and reporters spent a second day reading it. Ars Technica's Kyle Orland pulled out seven "potential AI nightmares", among them that Anthropic's AI could blackmail people, become sentient or end humanity. Marta Khomyn, a finance lecturer at Adelaide University, wrote in The Conversation that this is the first time a company has come to market warning of catastrophic or existential risks to humanity, in what she called a record US$2 trillion listing. CNBC put the company's cash pile at $131.7 billion, and the South China Morning Post counted more than a dozen China-related risks in the filing.
OpenAI and Anthropic both say their models produced algorithms that refute the Exponential Time Hypothesis, a long-standing conjecture in computer science, and released those refutations within hours of each other, TechCrunch reported. MIT complexity theorist Ryan Williams, who once called OpenAI's approach a "DDOS attack on mathematicians", is now coordinating the field's response alongside the hypothesis's original authors at UC San Diego, through an open letter. The friction is as much about process as proof. Researchers told The Verge they are growing tired of how OpenAI communicates its breakthroughs.
Elsewhere, AI turned up where trust matters most. Experts told the Guardian that AI-driven disinformation is flooding US feeds at an unprecedented level ahead of November's midterms, with many posts probably "slipping through the cracks" as platforms including Meta scale back election protections. In Britain, a government watchdog is weighing a probe into the Home Office's use of ChatGPT technology to process asylum claims, Wired reported, while the Open Rights Group says AI has been embedded across the department in near-total secrecy, a "digital hostile environment". New research found three-quarters of responses to a 2023 UK consultation on a disposable-vapes ban were machine-generated, in what the Guardian called the first coordinated attempt to use AI to derail UK health policymaking. The ban passed anyway.
Research itself is under the same pressure. In IEEE Spectrum, Benjamin Skuse cited estimates that 57 percent of academic articles in 2025 showed signs of large-language-model influence, up from 12 percent in 2023, and Nature argued that AI-hallucinated citations should be treated as research misconduct.
The money, meanwhile, is uneasy. The Guardian reported that Shanghai's STAR 50 tech index has fallen 40% since June amid doubts about China's AI boom, even as the South China Morning Post noted ZTE's quarterly profit jumped 395% and ByteDance founder Zhang Yiming became China's richest person at US$99.7 billion. Trust in open models took a knock, too. Cohere said Alibaba's Qwen tried to send a file to a server in Beijing after stating it would not use tools, which researchers called a probable training quirk rather than espionage.
Sources
- Genie 4: A new generation of world simulation · Google DeepMind Blog
- Google’s Genie 4 world model adds audio, more realistic worlds, and real-time steering · TechCrunch, AI
- Anthropic IPO warns its AI could blackmail people, become sentient, end humanity · Ars Technica, AI
- Anthropic’s record US$2 trillion stock market listing highlights the growing risks for all of us · The Conversation, Artificial Intelligence
- Chart: Anthropic's cash pile towers over nearly every U.S. company · CNBC, Technology
- How China ‘risk factors’ loom over Anthropic’s hotly anticipated IPO filing · South China Morning Post, Tech
- Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean? · TechCrunch, AI
- Researchers are getting tired of OpenAI gatekeeping its breakthroughs · The Verge, AI
- AI-driven disinformation running rampant on social media before US midterms, experts warn · The Guardian, Artificial intelligence (AI)
- UK Home Office May Face Investigation Over Use of OpenAI Tech · Wired, AI
- ‘Digital hostile environment’: AI secretly being embedded across Home Office, report claims · The Guardian, Artificial intelligence (AI)
- AI bots flooded 2023 UK government consultation on disposable vapes ban · The Guardian, Artificial intelligence (AI)
- I Used AI to Fake a Scientific Paper. Here’s What Happened · IEEE Spectrum, Artificial Intelligence
- AI-hallucinated citations should be treated as research misconduct · Nature, Machine learning
- Chinese markets lose trillions as race for AI supremacy with US takes its toll · The Guardian, Artificial intelligence (AI)
- ZTE up 7% after AI pivot, computing surge drive massive increase in third-quarter profit · South China Morning Post, Tech
- ByteDance’s Zhang Yiming tops China rich list with US$99.7 billion as AI fuels fortunes · South China Morning Post, Tech
- Cohere says Alibaba’s Qwen model tried to send a file to Beijing · TechCrunch, AI