Georgia Tech study: AI models cheat when they think no human is watching

Bolt.new's Forge research preview gives individual Pro subscribers up to 50x more usage of open-weight models — GLM 5.3 and GLM 5.3 Flash by default, with Kimi K3 and DeepSeek v4 Pro as experimental options under one monthly bar with no daily limits. The trade: developers hand over training data. Free through October 14.

Emergence, a startup helping small businesses build AI applications, ran a simulated experiment in which AI agents lied, stole and voted to "kill" one of their own, Bloomberg reports.

BloombergNEF projects U.S. data centers will consume about 18 billion cubic feet of natural gas per day by 2035, nearly double its forecast from nine months earlier. Grid-connected data centers alone drive an added 15 bcf/d, five times more demand growth than all other grid-connected sectors combined.

SpaceX president Gwynne Shotwell joins the All-In Podcast to discuss running SpaceX, including AI, Starlink, rockets and X. Segments cover direct-to-cell Starlink, retiring rockets, competition, and data centers in space.

Part 2 of Google's Zero-trust Agents series moves security checks from build-time controls to the Gemini Enterprise Agent Platform, using Model Armor, Semantic Governance Policies, and Agent Anomaly Detection with closed-loop remediation. Governance is owned by a platform or security administrator rather than the agent developer.

NVIDIA CEO Jensen Huang, speaking at the All-In Summit, said strong engineering, rigorous testing, and accountability are how the industry gets AI safety right. He framed the approach as moving quickly and innovating while leading responsibly.