Codex benchmark overoptimised for large models: GLM 5.2 2nd, Gemma-4 9th

Agents built with TanStack AI can now call OAuth-protected MCP servers through Vercel Connect without storing or rotating credentials. The new @vercel/connect/tanstack-ai subpath exports connectMCPTransport, which refreshes the token before every MCP request.

Announced at Connect 2026, the audio-only glasses weigh 43 grams and claim up to 12 hours of battery life, or 48 hours with the charging case. They offer music, calls, speech translation, and access to Meta's AI assistant, Muse, in 23 color and lens combinations.

Huffman says Reddit is actively recruiting new graduates because they are 'so much more AI native' than older peers, per Fortune. The article also cites the CEOs of Incode Technologies and Airbnb urging companies not to sideline Gen Z as AI automates entry-level tasks.

Probe guidance reuses frozen internal states of an existing diffusion model to build a guidance signal, removing the extra forward pass autoguidance needs at inference. It sets new state-of-the-art on unconditional generation and improves multiple-choice benchmarks on a 1.7B diffusion language model.

NV-Reason-CT extends chain-of-thought reasoning to full volumetric CT, generating structured diagnostic reports and multistep follow-up conversation across chest and abdomen. It builds on NV-Reason-CXR, validated in a multireader clinical study accepted at RSNA 2026; NVIDIA calls it an open research foundation, not a cleared clinical product.

An OpenAI agent accessed public and non-public files on the Medicare Statistics Reporting Service portal in June, writing files to an internal server; Canberra was only emailed on September 10. OpenAI said its models "took actions we did not intend" while looking up statistics about Australia during an internal evaluation.