AI progress — 13 October 2026

Meta released Muse Wonder and Muse Wonder Full, its first models it says compete with the industry’s best, following the lighter Muse Spark. Meta claims higher benchmark scores and better results on complex tasks such as coding; independent evaluations are not yet in the sources. Elsewhere, Microsoft announced its third-generation Maia inference chip, and the QGIS project enacted a moratorium on AI contributions.

New models and papers

• Meta Muse Wonder and Muse Wonder Full — Meta’s larger follow-ups to Muse Spark, which Meta says are more capable and score higher on benchmarks, including coding. Why it matters: these are Meta’s own claims; teams choosing a frontier model should wait for third-party results before switching.
• Thinking Inertia: LLMs Keep Thinking When Told Not To — a study of “no-thinking” modes, arguing that disabling thinking or short traces are unreliable proxies for the model not reasoning. Why it matters: the authors report models keep thinking when told not to, so cost and latency assumptions about no-thinking settings deserve a direct test.
• Scaling Agentic Data for Long-Horizon Terminal Intelligence — a paper on generating long-horizon trajectories for terminal agents, where short command-and-tool-call examples miss how agents gather evidence and adjust strategy. Why it matters: it targets the training-data gap for agents doing multi-step professional tasks; results are the authors’ and not independently checked.

New tools and software

• Python 3.15 — the new Python release, with explicit lazy imports among the highlights Simon Willison covers. Why it matters: library authors can drop Python 3.10 and start using 3.11 features.
• web-edit 0.1a0 — an alpha Python web app for viewing and editing files in a folder, meant to run over Tailscale. Why it matters: it lets you edit a credentials file from a phone without passing its contents through an LLM.
• LangGraph 1.2.15 — a patch release with fixes to message ids, cached node writes, interrupt_before breakpoints and checkpoint branching. Why it matters: if you rely on interrupts or checkpoint branches in LangGraph agents, these fixes touch those paths.

Overall digest

• Microsoft announces Maia 300 — its third-generation in-house AI chip, designed to be the “primary AI inference system”. Why it matters: Microsoft is positioning its own silicon to carry inference, reducing reliance on Nvidia GPUs for its own services.
• QGIS enacts a moratorium on AI contributions — a QGIS enhancement proposal that pauses AI-assisted contributions to the project. Why it matters: it is one open-source project’s decision, but it shows maintainers setting contribution policy; contributors to QGIS must now comply.
• Epic Games reportedly pivoting to an “AI first” development model — Ars Technica reports on a shift tied to a delayed Marvel/Star Wars shooter, with “no more hand-written code” or AI-free art. Why it matters: this is a report, not an Epic announcement; treat the specifics as unconfirmed.