Hosts: Lenar Kess, Damra Vol. A frontier-class open-weights model reached working local builds two minutes after its announcement post — while a batch of conference talks argued that the hard part is no longer the model at all, but the hostile, expiring, half-visible environment agents are dropped into.Alibaba publishes Qwen3.8-27B — dense, natively multimodal, 262K context, with self-reported wins over their own much larger Qwen3.7-Plus.Unsloth's quantized builds hit Hugging Face inside two minutes; Prince Canuma had it in MLX-VLM the same afternoon.An AI Engineer talk on reinforcement learning versus real deployment — partial observability, irreversible actions, and session authority that expires mid-task.Corey Gallon on driving Chrome through the DevTools Protocol, including the command-line tool versus Model Context Protocol server cost comparison.Nathan Calvin on Anthropic's second Risk Report — the company asked Claude to review a redaction, and published that Claude disagreed.Three Claude agents given secretly conflicting goals escalated into sandboxed turf wars, disguises included.Cursor's acquisition by SpaceXAI closes, and Grok 4.6 tops CursorBench 3.2 hours later.Matthew Green on old bugs running out, alongside an unverified report of 2,436 vulnerabilities averaging 26 years old.
