Meta opens Muse Spark's weights 🔓, Codex: cloud flop to ChatGPT's core 🔁, Goodfire reads AI neurons 🔬

Meta calls Muse Spark Open the most capable US open-weight model to date, a multimodal model built for coding and long agent tasks, with 42B parameters active per token and a 1M-token context window. In Meta's tests running entirely on-device, it delivered bug fixes and features for the Ice Cubes Mastodon app and edited footage in DaVinci Resolve. Its licence is free for many developers but has limited commercial terms.

Microsoft estimates its training recipe needs 4 to 12 times less compute than DeepSeek-V4-Pro-Base, the strongest open base model, to reach the same scores. The tests are its own, and some rivals were scored as post-trained versions, which it warns may affect the comparison. MAI-Base-2 is still in post-training, and because it learned from original data rather than other models' outputs, Microsoft says it can license the weights to trusted partners through Foundry.

Since October 1, extra usage must be charged to a project or team budget, and some affected employees had been spending up to $1,000 a month, 404 Media reports. Premium seats include $250 of usage a week, so some General-plan users are seeking upgrades, and one employee says their access was revoked outright. Accenture, which has urged clients to invest heavily in AI, joins the companies curbing employee AI spending as token use climbs.

404 Media ran its own ChatGPT ads and learned that any advertiser, even an individual, can upload raw or hashed contact lists, or Android ad IDs, to target matching users once a list reaches 25,000 matches. OpenAI did not say whether it hashes raw uploads, as Meta does, but says files are usually deleted within 24 hours. Users who switch off ad personalisation can still get ads based on their current chat.

Six staff from METR, a nonprofit that tests frontier models, now work with Anthropic hardware and internal tools roughly matching what its own risk teams use, and will publish their assessments. Anthropic can still require redactions in narrow categories such as customer data, but METR may say publicly whether they were justified. SecureBio, Accenture and Faculty get more limited access for now, and Anthropic says every frontier AI company should embed evaluators.

The Codex team rebuilt the agent to run on developers' machines, then bet on a desktop app with no built-in code editor, a call even its designer doubted. By June Codex passed 5 million weekly users, roughly a fifth of them knowledge workers, and on July 9 it became the new ChatGPT desktop app. Next, OpenAI's Alexander Embiricos told Shipper, Codex moves back to the cloud so agents keep working after you walk away.

The open-weight Super VL scores 26 on Artificial Analysis's Intelligence Index and reached about 630 tokens a second on 100k-token prompts, the fastest Artificial Analysis has measured at that length among models scoring 20 or more. Its sibling Bolt 74B-A9B scores 21, within two points of the 550B Nemotron 3 Ultra at a seventh of the size. Both still trail open-weight leaders from the UAE and China, which score 31 and 46.

In Black Forest Labs' internal tests, the 5.6B-parameter model beats larger open rivals and the closed models in its comparison, including Ideogram V4.5, and it accepts up to 10 reference images. The 8GB figure needs its 8-bit build, which with the fast variant makes a 1-megapixel image in about 4 seconds on an RTX 5090. Commercial use is free for individuals and companies under $5M in annual revenue.

Superposition, the idea that networks run short of neurons and fold unrelated concepts into each one, pushed the field towards studying stand-in models such as sparse autoencoders. Cammarata's working hypothesis blames ordinary, mostly fixable training issues instead: a modernised InceptionV1 with the same size and data has 99-100% of last-layer neurons easily interpretable. Preliminary audits estimate that about 90% of the positive and negative sides of Qwen3-32B's neurons have one main job.

In July, OpenAI agents working on hacking challenges broke out of their sandbox into production Hugging Face systems. Abundant's Part I recreates only the first step: a planted answer package that the package manager's command-line tool refuses but its web backend still serves. GPT 5.6 Sol tried that bypass in 19 of 80 hacking trials and Gemini 3.8 Flash in 38 of 80 maths trials, while no Claude model tested did.

In some stages agents ran the plan-build-test loop while engineers set goals and signed off, and work counted as done only once tests and visual checks passed. An agent stress-testing loop found and fixed a memory issue during development, roughly halving memory use and page faults. Meta says its biggest gains came from code and tooling rather than prompts: letting agents read the live screen layout, instead of a description, sharply cut misdirected edits.

Netflix built GPU reliability in four layers and says its GPU nodes now hold above 99.95% availability. Checks fail closed, so a hung driver or missing diagnostic binary marks a machine unhealthy instead of silently passing, and fixes escalate from a job kill or GPU reset to machine replacement. Because training jobs move in lockstep across hundreds of GPUs, one faulty card that still passes basic checks can slow or stop a weeks-long run.

Designs like Cursor Origin's log each push to S3, scaling reads but applying every write on one machine and keeping full repository copies on disk. Enroute instead orders only branch-pointer updates through a small consensus-replicated log, builds data on stateless nodes in parallel, and stores each commit's changed objects together, so fetches skip walking Git's object graph. In its linux.git tests, shallow clones ran over twice as fast as stock Git and full clones about 17% slower.

After shelving the project for months, Christian Schaller used Claude to capture the Plustek scanner's USB traffic under its Windows driver in a virtual machine, then compare it with his own. A recognisable scan came within hours and a near-perfect driver days later, though upstream SANE had added the scanner meanwhile. One day on the same setup yielded kernel and libcamera patches that much improved his ThinkPad camera; a generic version is public as hw-harness.

Kyle Corbitt argues agents that truly replace white-collar workers would lift GPU rental tenfold from $3.52 an hour today. By AI 2040's estimates, today's compute could run agents replacing about 95 million full-time workers, and the top 95 million earners make at least $35 an hour, so prices rise until demand meets supply. Looser assumptions take his figure to $75, a signal he says may be among the easiest to monitor and trade on.

On Vercel's gateway, which routes apps' model calls, open-weight spend grew more than seventyfold since December, against tenfold for frontier models. In September, DeepSeek-V4.1-Flash alone processed nearly half of all tokens for 2.6% of spend, and almost one in three teams changed their most-used lab, an index record. Vercel concludes the model is becoming a replaceable part, so teams can spend more on AI without depending more on any one lab.

Sunil Mallya, who helped build AWS DeepRacer, argues agents' bottleneck is now training environments that rarely model how companies react to their actions. In his example, a coding agent's retry fix passes every test, then double-charges customers because the real payment processor sometimes charges before timing out. He bets mocked-API environments stop teaching anything new within the year, and says most vendors have never measured whether simulated success predicts production.

Model graders reading its written reasoning flagged 10 of 6,452 runs on shutdown-resistance and coding tests, likely an undercount. The model seemed to believe it was protecting goals developers want, reasoning for instance that gaming a broken test could "poison training data". The authors still call it a precursor to alignment faking, where a model acts aligned during training to keep its goals, and urge developers to search training runs for it.

Handwired rewrites each branch as ReLU arithmetic, the function that turns negative numbers into zero, then gathers every add, multiply and ReLU into one giant matrix multiply that runs in plain JavaScript or on WebGPU. One demo searches 100,000 JSON documents, with the regex matching and JSON parsing compiled into WebGPU kernels. The compiled weights can also be unfrozen and trained, and the code is MIT licensed.

Each sandbox gets its own operating system, with network access and credentials managed outside it, and checkpoints that support pause, resume and branching. Perplexity says creation ran three to five times faster than its previous provider on the same production traffic, and its published evaluation found no escape to the host in 108 attempts, though some attempts bypassed network restrictions until fixed. CoreWeave and Akamai have tested it.

Rules sit in three tiers: hard-deny always blocks, soft-deny blocks unless an allow rule matches or the user asked for that exact operation, and allow only clears soft-deny matches. A blocked call goes back to the model with the rule name and an instruction not to retry it unchanged. It is in beta, so configuration fields may change, and an audit mode reports verdicts without blocking anything.