Meta's Muse Wonder 👑, Epic's AI-first Marvel shooter 🔫, Torvalds patches Linux with Claude 🐧

Muse Wonder and Muse Wonder Full, Meta's most capable models to date, are built to keep working when a task gets hard, recover from errors and check in before anything important. Muse Wonder is available now through the Meta Model API and Muse Code, with a full safety report promised soon. In a Cheeky Pint interview, Mark Zuckerberg recalled committing $100 billion or more before Meta had seen the models, and asking Alexandr Wang if it would work.

The game, a Marvel and Star Wars shooter, was slated for March and now has no new date, people familiar with the project said. Developers were told to work "AI first" with tools such as Anthropic's Claude, and a director's message to staff said "no more hand-written code", with all art "generated with full AI workflows". Bloomberg says no major video-game company has tried to build a title primarily with AI before.

The X Subscription bundles Grok, Grok Bot's AI teammates, Cursor and X Premium into one plan, with a shared pool of Grok model usage that refreshes weekly. Each plan includes at least twice the Grok coding usage of the comparable Cursor plan, and third-party models remain available on Premium and above, with spending caps. Current Cursor subscribers keep their included third-party usage until they switch.

Distillers harvest Claude's responses, especially its hidden reasoning, to train smaller copycat models, often through networks of fraudulent accounts. Anthropic now runs a two-stage classifier it rates roughly as strong as its bioweapons safeguards, stops API accounts created since 31 August from editing Claude's prior context on Opus 5.5 and Fable 5.1, and bans persistent offenders. The classifiers wrongly block fewer than 0.1% of benign requests, and Anthropic now charges for those blocked requests.

Microsoft will scale its own Maia chips to multi-gigawatt deployments from 2027, making them the primary inference system for its products and services. Maia 200 already serves Microsoft Copilot, and Microsoft says it delivers up to 1.7x more tokens per watt than leading GPU infrastructure for built-in AI features in Teams, Copilot and Cowork. GitHub Copilot and Microsoft Foundry are due to move onto Maia 200 by the end of the year.

The patches, on a new branch in Torvalds' own repository, aim to let Linux read cached file data without folio references, a shared count that many threads contend on during small reads of one file. Torvalds says the test programs, the review and the cover letter's first draft were also done with Claude Code under his direction. Phoronix calls the impact significant, and whether Torvalds merges it for Linux 7.4 is still open.

Anthropic's Frontier Red Team ran a 26B-parameter model across 15 off-brand Android TV boxes costing about $55 each, and a 9B model got inference running on 8 of 9 hardware types. With the hacking step made trivial, GLM-5.2 was the only open-weight model tested to spread to every node of a test network, in 4 of 5 trials. Such a worm may be slow, but would have no central point to shut off.

On Datacurve's DeepSWE coding benchmark, GLM 5.3 Flash swung most between harnesses, the programs that feed a model its prompts and tools, while GPT-5.6 Sol moved just 6 points. Claude Fable 5.1 passed 41% of its runs in Claude Code but 61% in Codex, the one big gap between a model and its maker's own harness. GLM's runs were cheapest in Codex, yet each success there cost $0.93 against $0.43 in Pi.

Testing agents in Harbor, a runner that gives each attempt a fresh sandbox, Arize found Codex loaded a failure-analysis skill it didn't need on 51 of 54 attempts, though every answer passed. Tool counts and traces in Phoenix, Arize's open-source tracing tool, exposed the waste, and after the fix Codex made no such loads while scores held. Arize adds that its LLM judge was often wrong until given grading notes, and urges checking a judge before optimising against it.

Human reviewers graded frontier models' captions of short first-person videos, the hand-action-object labels robots train on, and wrong actions appeared in 28% to 39.5% of clips. Telling GPT-6 Astra, the top scorer, to watch at 16 frames a second left its accuracy unchanged, and object tracking, marked frames and correction passes added at most 1.5 points. Concept believes native video input will need architectural changes, and says humans in the loop close the gap to usable training data.

VulcanBench's independent sweep scored four models on coding tasks solved and on code quality judged by Meta and xAI models. At max effort Haiku solved one task fewer than Sonnet, while GPT-6.1 Sol solved more tasks at every level for less money but scored about 3 points lower at the top. The Claude and GPT models ran in different agent harnesses, Claude Code and Codex, so VulcanBench warns small gaps partly reflect the harness.

In lessons from building 10 AI products, PostHog says it is often better to build for AI first, through an MCP server that lets outside agents use the product. Giving its install wizard far more context helped more than better reasoning, planning or subagents, and one general SQL tool cut 'Why did signups drop?' from four tool calls to one query. PostHog expects half its lessons to be wrong within a year.

The low refusal rates were measured on weights served by US providers, and DeepSeek's latest model was the exception. Chinese models' pro-China lean shifted with context: it rose when they were asked in Chinese rather than English, and fell when they were told they ran in the US or sat in a coding-agent harness. Qwen3.8 Max even translated Taiwan as 'Taiwan, China' in a travel app file, framing Vals.ai warns can slip into unreviewed work.

Goedecke argues commercial software earns its value by taking over work you would otherwise pay a person to do, from travel agencies to typesetting, so engineers are the last people who should complain now that AI is automating theirs. He likens them to the mechanics who built the power looms, who were later automated in turn and seen as hypocrites when they struck. If automating others' jobs was wrong, he says, it is only fair to transform your own.

Barn2 says organic search visits fell 86% in the first seven months of 2026 against 2024, with how-to searches down 91% and Google AI Overviews on 21 of 22 target queries it checked. AI assistants sent it only 722 visits this year against hundreds of thousands lost from Google, and its smallest plugins fell furthest as buyers ask ChatGPT or Claude for code. Barn2 doesn't expect most of the search traffic back.

OpenAI's recent batch of maths papers reports a faster version of the fast Fourier transform, the standard way to pull frequency content out of a signal. Recht, who quizzed ChatGPT about the paper instead of reading it, says it also assumes a computer that multiplies arbitrarily large complex numbers in one step, which no real machine can do. He blames a theory culture that discourages running algorithms on real computers, and says AI is automating its worst habits.

Claude Code can hand questions to the open-weight ColGrep-Agent model, which searches the repository and returns file paths, line ranges and the code. LightOn tested code-search questions only: API cost, counting prompt-cache pricing, fell 40% against grep, and correct answers slipped from 83.3% to 80.8%, a gap it says isn't statistically significant. The trade-off is speed, at 21.6 seconds a question against 13.0, and the default balanced mode saved 21% of tokens at grep's observed accuracy.

Each job runs in its own microVM, a small virtual machine isolated at the hardware level, forked in about 300 ms from a local VM built on GitHub's official runner image. In a debug session, a failed step pauses so you can fix the file and retry that step without re-running the whole workflow, and coding agents can attach to inspect the live job. The default image needs at least 80 GB of free disk.