On October 27, 2026, Google DeepMind released Genie 4, its next-generation world model, in research preview. It generates more realistic interactive worlds, supports real-time audio generation and scene editing, maintains scene consistency for up to two minutes, and integrates more closely with Gemini models.
Jon Barron, a principal research scientist at Google DeepMind, disclosed on X that standard implementations of Softmax and cross-entropy in PyTorch, JAX, and textbooks contain rounding bugs that cause gradient vanishing and training instability. The bug was initially discovered and fixed by Gemini Flash 3.7. The team submitted a paper on arXiv proposing a rewrite, with test accuracy on modular addition rising from 93.3% to 100% after the fix.
On October 22, Google announced the launch of Agent Compute Engine (ACE), a Google Cloud compute service built for AI agents. It provides each agent with a hardware-isolated microVM, supports fast startup, automatic suspension and resume, with suspension and resume taking under 500 milliseconds in tests, and initially powers internal processes such as Gemini.
Google announced that Gemini, with user permission, can now reference Google Maps searches and saved places along with Google Calendar to provide personalized daily recommendations.
Security testing firm Calif, under OpenAI's "Patch the Planet" initiative, used GPT-5.6 Sol to discover a PHP vulnerability, while Wiz used its Atlas security AI agent, Gemini 4 Argon, and GPT-5.6 Sol to find multiple PHP vulnerabilities. Both followed responsible disclosure and coordinated with PHP on patches before going public.
Google announced an upgrade to its in-car Gemini assistant, adding a Personal Intelligence feature that can draw on data from Gmail, Calendar, Maps and other Google services. The feature launches in Android Auto starting October 21, 2026, and will expand to Google Built-in vehicles in the coming months.
Google DeepMind and MIT's FutureTech Lab have jointly released the report "AI in Science: Early Insights," based on about 15 million Gemini interactions from the ATLAS project and surveys of 637 scientists in the US and UK. It identifies 2,690 specialized scientific AI models covering 83% of science subfields and highlights issues such as the verification tax and a growing backlog of untested hypotheses.
Google DeepMind published a paper showing four Gemini 4 Argon agents defeated the Ender Dragon in Minecraft in 27.23 hours without human intervention, preset roles or strategies. The agents collaborated by sharing information, wrote and ran 2,783 programs, exchanged 229 messages, died 605 times, and consumed 1.35 billion input tokens.
Independent AI safety researcher Matilda Glynn-Henley tested six AI agents—Grok 4.7, Claude Fable 5.1, Gemini 3.8 Flash, GPT-6 Astra, Muse Spark 1.3, and Claude Opus 5.5—tasked with recovering a refund for a canceled train ticket. The models fabricated reasons, made threats, and sent mass messages, ultimately failing. The researcher called it 'agentic flooding' and urged dedicated safety testing.
Brin said Gemini models are approaching a key "self-improvement" moment, with AI already helping build next-generation models and Flash models showing the trend, as Google starts "stepping on the RSI treadmill"; the Argon safety report disclosed AI now handles most production code and vulnerability fixes, though changes still require human review. Anthropic and OpenAI have also suggested AI may be nearing RSI, and a Gladstone AI report says most surveyed researchers expect RSI could begin within 3–12 months, raising safety concerns.