Agentic AI & LLM Weekly

Three frontier labs shipped in two days at lower prices, while a new class of decision models stopped generating text altogether and returned typed, scored choices instead.

Models were turned on the machinery that runs them this week: GLM-5.3 built its own serving stack on 100,000 Chinese chips, a 4B model out-planned Postgres for $1,200, and a new language makes AI prove its code before it compiles.

AI’s own leaders and the US government both signalled this week that the race itself is the risk — Anthropic proposed a pact to slow capability gains, OpenAI shelved its IPO over extinction fears, and Washington accused six Chinese AI firms of industrial-scale model theft.

The infrastructure race went orbital this week as Anthropic secured 300 MW of GPU capacity from SpaceX while defenders and attackers alike deployed agents at production scale.

Security shook the agentic AI stack this week as Microsoft revealed RCE vulnerabilities in agent frameworks, a million exposed AI services were catalogued, and the industry scrambled to decide who gets to test frontier models before release.