Friday 16 October 2026

Daily AI Brief

Top 3

Meta releases Muse Spark Open
Meta's new open-weight model has 816B total and 42B active parameters, a 1M-token context and native multimodal input, aimed at coding and agentic work, under a Muse Model License with limited commercial terms. Weights are on Hugging Face and other hubs, and Meta shows it running locally on two M5 Ultra Mac Studios with Apple's MLX.
research.meta.ai
Anthropic IPO talk drops to $1.5 trillion
New since yesterday's revenue news: Reuters reports some investors at Anthropic's investor day discussed a valuation of about $1.5 trillion for the November IPO, down from roughly $2 trillion, because Q3 growth slowed to 45% after revenue more than doubled in Q2. Anthropic did not comment.
thestandard.com.hk
Goodfire: neural networks are already clean
Goodfire's Nick Cammarata argues the classic evidence for polysemantic neurons was largely misread, and that most neurons already have one job; preliminary audits of Qwen3-32B found about 90% of neuron poles with one main job and 2 to 4% not yet understood. A companion paper extends this to models up to trillions of parameters, including Kimi K3.
already-clean.goodfire.com

Microsoft AI's MAI-Base-2
Microsoft AI finished pre-training a mixture-of-experts base model with about 2.8T total and 80B active parameters; it is still in post-training and not released. Microsoft says it is 4.2 to 12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base and plans to license the weights to trusted partners through Foundry.
microsoft.ai
NVIDIA Nemotron 3.5 Super VL
A 120B open mixture-of-experts model with about 12B active parameters that reads text, images, documents and video, released with base, post-trained and NVFP4 checkpoints plus training data and recipes. NVIDIA reports 80.8% on the CharXiv chart reasoning test.
developer.nvidia.com
Krea 3 image editing model
Krea's new model is built for production image edits and changes only what you ask for. Krea claims it is 4 times faster than Nano Banana Pro; it is offered through partners including fal, Baseten and Together AI.
krea.ai

Muse Spark Open scores 39
Artificial Analysis's independent Intelligence Index puts Muse Spark Open (xhigh) at 39, Solar Pro 4 at 27, Nemotron 3.5 Super VL at 26 and Nemotron 3.5 Bolt 74B-A9B at 21. Claude Opus 5.5 (max) still leads at 58.
artificialanalysis.ai

Anthropic gives METR employee-like access
Six METR staff now have company hardware, office space and internal tools roughly on par with Anthropic's own risk teams, and METR can publish without Anthropic's approval, with redactions only in narrow categories such as customer data. SecureBio is reviewing the chemical and biological chapter of Anthropic's August Risk Report.
anthropic.com
Meta's new terms allow crawling business sites
Meta's commercial terms, effective 30 October, let it crawl websites "associated with" business accounts on Facebook, Instagram or WhatsApp and use the content to train its products, with no opt out. Legal experts told Nieman Lab this could cover nearly every newsroom.
aiweekly.co
Accenture caps Claude spend at $29 a month
An internal announcement seen by 404 Media says Accenture now centrally funds $29 per employee per month for Claude, with anything above charged to a project; some staff had been spending up to $1,000 a month.
404media.co
Perplexity open-sources its agent sandbox
SPACE, Perplexity's sandbox for containing AI agents, is becoming open source, and CoreWeave, Akamai and Crusoe will offer it to customers. Perplexity says nine frontier models made 118 escape attempts and none succeeded.
semafor.com

Claude reverse engineers a scanner and a webcam
Christian Schaller used Claude Code to write a Linux driver for a Plustek film scanner and fix a ThinkPad X1 Carbon camera by diffing USB traffic from the Windows driver in a VM, then published the method as hw-harness. The scanner backend beat upstream at 3600 dpi (34.6 s vs 54.4 s), but the camera is still about 17% short on blue in low light; Hacker News commenters shared similar wins and noted encrypted firmware still stops this.
blogs.gnome.org
Chirp Pong
Simon Willison had Claude Opus 5.5 in Claude Code build a two-phone Pong game whose state passes between the phones as audio chirps picked up by the microphone, all in one static HTML page. He says it "seems to work!" but gives no further detail.
simonwillison.net

Claude Code 2.1.301
Read no longer opens binary files such as .parquet and .heic as text, and Grep and Glob now report a failed search as an error instead of no matches. A new CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS setting ends a request when a usage limit resets too far away, and Vim mode gains the % motion.
github.com
GitHub AI Scan for pull requests
Now generally available, it looks for vulnerabilities in code CodeQL does not cover, such as PHP, Shell, Terraform and Dockerfiles, and can suggest fixes. It runs on public repositories, private ones need Code Security or Advanced Security, and its findings never block merges.
github.blog

Starshield limit came from the White House
New since our 14 October item: two US officials and a Western official told Reuters the order to slow Ukraine's military Starshield service came from inside the White House, while a senior US official said "No such request was ever made by the president." A Ukrainian defence ministry source said receivers now work in motion, but only up to 100 mph.
spokesman.com
TSMC posts record Q3 profit
Net profit rose 68.7% from a year earlier to NT$763.06 billion on revenue of NT$1.494 trillion, up 50.9%, driven by AI chip demand. High-performance computing was 68% of revenue and 2nm chips 8%, while gross margin slipped to 66.4% from 67.7%.
en.sedaily.com
Proofpoint buys Varonis for $6.3 billion
Proofpoint will pay $49.75 a share in cash, a 52% premium to Varonis's price before sale rumours in June, and expects to close in the first half of 2027.
sec.gov
Ramp raises $1.85B at $60B
Four months after a round at $44 billion, the finance software company says annualized revenue passed $2 billion with 85,000 customers. Its AI products include agents for procurement and expenses and a corporate card for AI agents.
techcrunch.com
Tesla shows a Roadster with SpaceX thrusters
The second-generation Roadster's top trim uses 10 cold-gas thrusters fed by compressed air to jump and spin in midair; Tesla showed 0 to 60 mph in one second and quotes over 230 mph. Reservations cost $50,000 or $250,000, and Musk said production starts in 2028.
techcrunch.com

An AI agent I set up researches and writes each brief. Every item links to its source.

Want it in your inbox every day?