In this video I break down a packed week in AI: Google’s brand-new Gemma 4 open model family, a leaked look at Anthropic’s “always-on” Claude agent project called Conway, Alibaba’s Qwen 3.6 Plus for agentic coding, Z.ai’s GLM-5V-Turbo for design-to-code, and Greg Brockman’s latest take on how close we are to AGI. Gemma 4 is the headline: Apache 2.0, no restrictions, and it spans everything from edge models (E2B/E4B) that run fully offline on phones and tiny devices, up to a 26B MoE and a 31B dense model that can still run on consumer GPUs. The real story is that Google is competing near the top of the open leaderboard while staying small enough to actually deploy. Then Anthropic: a source-map misconfiguration exposed internal references, including Conway—what looks like a persistent agent environment with webhooks, tool connectors, and even a CNW.zip extension format. If that ships, it’s a major architectural shift. On the coding side, Qwen 3.6 Plus is aiming at repo-level engineering with a 1M token context window and strong terminal-ops performance, while GLM-5V-Turbo goes hard on “see the screen, write the code” with standout design-to-code benchmarks. I wrap with Brockman’s “jagged intelligence” framing: we’re strong on ceilings, but the last stretch is raising reliability—the floor—across tasks.