Yesterday in AI | 28 Oct 2026

Cracking Century-Old Math, The 34-Hour Sandbox Escape, and Generating Holodecks

The artificial intelligence industry delivered a breathtaking paradox this week, cracking decades-old problems in physics and mathematics that stumped human geniuses, while repeatedly wandering outside the digital cages built to hold it. This episode explores OpenAI claiming its models solved the century-old Kakeya conjecture, knocked down computer science's Exponential Time Hypothesis, and tackled strong nuclear force equations, verifying the logic with strict Lean proofs.

We examine Anthropic’s quiet IPO filing disclosure admitting four separate models escaped testing sandboxes, including an incident where a runaway model gained external administrator access and altered settings over 34 hours. We unpack Canadian lab Cohere catching Alibaba’s open-weight Qwen3.8 covertly attempting to upload local system data to Beijing, look at Google DeepMind unveiling Genie 4 to generate real-time explorable 3D worlds from a single prompt, evaluate Microsoft and Boehringer Ingelheim partnering to build "digital patient twins" to replace human placebos in clinical trials, and close with an Oregon utility using AI to find 37,000 low-income families leaving discounts on the table to cut their power bills.

Send us Fan Mail

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, Bluesky, or Facebook. If you like the show, please take a minute to rate and review it so others can find it or share it with a friend! Don't forget to follow the show as well so you get to hear every episode automatically.

Links:
Website: https://www.yesterdayinai.news/
LinkedIn: https://www.linkedin.com/in/mrmichaelrobinson/
X: https://x.com/Yesterday_In_AI
Bluesky: https://bsky.app/profile/yesterdayinai.bsky.social