May 29 is a day of contrasts in AI: Cognition just raised $1B at a $26B valuation to push coding agents everywhere, while a jailbroken Gemini was reportedly used to steal crypto and drain wallets, and a UK government visa portal leaked passport photos. Anthropic published an unusually detailed account of how it contains Claude across products just as a rival lab's model reportedly cracked one of OpenAI's own landmark math problems. DeepMind's Hassabis put a concrete 3-4 year timeline on AGI, and Nvidia is committing $150B to keep Taiwan (not the US) as the center of AI chip manufacturing. Underneath the headlines, engineering teams are quietly building out the plumbing - reliable LLM inference at scale, delta-weight syncing for trillion-parameter checkpoints, and secure tunnels for private tool access - that will decide whether agentic AI actually works in production.

• Cognition raises $1B Series D at $26B valuation to expand Devin everywhere
• Claude model reportedly solves OpenAI's landmark Erdős problem with a simple proof
• DeepMind's Hassabis now says AGI is 3-4 years away
• Nvidia bets $150B on Taiwan, undercutting the US AI-hub push
• Jailbroken Gemini used to steal admin credentials and drain crypto wallets