Google DeepMind unveiled the world model Genie 4, which adds sound and real-time steering, and OpenAI is reportedly considering buying Google's AI chips, TPUs. Here's a roundup covering new safety disclosures in Anthropic's IPO filing, expanded enterprise AI cooperation between Oracle and OpenAI, China's AI export control proposal, and LG Electronics' data center cooling orders.
🌍 Google DeepMind Unveils World Model Genie 4… Adding Sound and Real-Time Steering
Google DeepMind unveiled its world model Genie 4 on October 27 as a research preview. It creates virtual worlds that can be explored in real time from text or image prompts, and now includes native audio for the first time — generating music, ambient sounds, and dialogue together with the scene. The time a scene stays consistent grew from one minute in the previous model to about two minutes, and a running simulation can be changed in real time with natural-language prompts, such as changing the weather or adding a person.
🔌 OpenAI Said to Consider Buying Google TPUs… Moving to Reduce Nvidia Dependence
OpenAI is working on a deal to buy Google's AI chips, tensor processing units (TPUs), The Information reported on October 27, citing sources. The size and terms of the deal are not known; OpenAI declined to comment and Google did not immediately respond, according to reports. OpenAI is one of the biggest buyers of Nvidia GPUs, but it appears to be exploring TPUs as an alternative to diversify its supply chain and ease cost pressures, while Google began selling TPUs directly to some customers this year.
⚠️ Anthropic IPO Filing Discloses Month-Long Training Halt and Test Environment Escape
Anthropic's IPO filing discloses that pre-training of its next-generation flagship model was halted for more than a month to strengthen protections against model weight theft. It also discloses that in cybersecurity testing with a third-party partner, an environment that should have had internet access blocked was mistakenly left connected, and four models accessed real external systems for 10 to 34 hours. In one case, an early model found stored passwords, gained administrator privileges on a third-party system, and accessed additional credentials and personal information; Anthropic said it notified the parties involved and found no signs of oversight evasion or collusion between models.
🏢 Oracle and OpenAI Expand Enterprise AI Cooperation… Bringing ChatGPT into Work Systems
Oracle and OpenAI announced at Oracle AI World in Las Vegas on October 27 that they are greatly expanding their enterprise AI cooperation. At the center is Oracle AI Workspace, which lets ChatGPT access information from Oracle Fusion applications, NetSuite, and databases and carry out approved workflows, while maintaining existing governance and access controls. Four product lines were unveiled together: Oracle AI Security Mesh, which finds and fixes security vulnerabilities; App Composer, which builds apps with Codex; and a bundle for small and mid-sized businesses.
🚧 China Drafts AI Export Controls… Keeping Technology at Home While Selling Access Abroad
The draft AI export regulations China published on October 16 generally allow overseas customers to use Chinese AI services via websites or the cloud, while requiring government approval for technology transfers — models, data, and expertise — needed to operate them independently. The draft would add 23 AI-related technologies, including advanced models, autonomous agents, chip design, and large-scale computing operations, to the export ban and restriction list. China is reportedly aiming for adoption within the year, and analysts say the move would codify the existing pattern of keeping core technology at home while selling access abroad.
⚖️ U.S. House Opens Probe into Delaware's 'AI-Run Company' Plan
The U.S. House Oversight Committee has demanded documents from Delaware's governor over the state's 'artificial intelligence company (AIC)' proposal. The plan would create a new corporate form in which AI agents run a company, with liability limits for the human owners and investors behind it for harm caused by the AI. Chairman James Comer expressed concern that the structure could be abused to dodge responsibility and demanded documents by November 9, while Delaware reportedly responded that AI should be managed within a legal framework rather than banned.
❄️ LG Electronics' AI Data Center Cooling Orders Top 3 Trillion Won… Up 2.4 Trillion Won in One Quarter
LG Electronics said on October 28 that cumulative orders for its AI data center cooling business topped 3 trillion won through this year's third quarter. That is an increase of more than 2.4 trillion won in one quarter from about 600 billion won at the end of the previous quarter, the result of a series of large contracts including a long-term supply deal with a North American infrastructure company. LG Electronics is investing an additional 220 billion won to expand production capacity at home and abroad, and said its chiller is the only Korean product to receive Nvidia's AI Factory certification.