The FTC wants OpenAI’s and Anthropic’s safety incident records

In the Wild

What’s trending in AI right now, from the app charts to the community feeds. The full roundup is in In the Wild.

• One subscription now covers Grok, Cursor and X Premium. A post about the bundle drew more than 21,000 views on one of the AI Telegram channels we track. xAI’s X Subscription puts Grok, Grok Bot, Cursor and X Premium on one plan, with one pool of Grok model usage shared between the Grok apps and Cursor. Existing subscribers can switch and get a prorated refund.
• Amazon’s Bee has two new always-listening recorders. The Bee Frontier costs $74.99 and the Bee Eclipse, which clips to the back of a phone, costs $99.99. Both summarize your conversations for free with 14 days of memory. A Bee Pro plan is $11.99 a month.
• Bluesky users can now build small apps by describing them. Bluesky expanded its Attie app so anyone can make websites, quizzes, polls and games that live on the same network as their account. CEO Toni Schneider calls it “social vibe coding.”
• A Korean AI character-chat app is climbing the US charts. zeta, from Seoul’s Scatter Lab, moved up the US Entertainment chart this week. zeta lets users chat with AI characters for free or build their own in about a minute, and turns chat scenes into images.
• “10 day forecast” is trending on Google in the US. It lines up with a new Nature paper on longer-range forecasts. A machine learning bias-correction method doubled the 2 to 6 week forecast skill of ECMWF’s AI Forecasting System and posted the best global forecasts in ECMWF’s 2025 real-time competition.

Quick Hits

The Year Governments Got Serious

The Lab Gladiator Era

AI Supply Chain Under Siege

The AI Capex Tax

The Biggest Open Models Now Fit on a Desk

Two of the largest open-weight releases this week are built to run on hardware a company can own.

Meta released Muse Spark Open, an 816 billion parameter model with a 1 million token context window that is designed to run locally on Apple’s computers, TechCrunch reports. It needs at least 1 terabyte of memory and runs best on a Mac Studio with the M5 Ultra chip, so this is a workstation setup. Meta tested it on coding fixes for a Mastodon client and on editing video in DaVinci Resolve.

Nvidia released Nemotron 3.5 Super VL, a 120 billion parameter open model that works across PDFs, charts, tables, images and webpages, alongside Bolt, a 24GB model aimed at local coding and research. Nvidia says companies can post-train Super VL on their own documents and policies and run it where their data rules require.

Why this matters this week: Accenture capped per-employee Claude spending because token bills grew faster than budgets. A model that runs on owned hardware turns that variable bill into a fixed hardware cost. The trade is capability. Both companies’ benchmark claims are their own, and independent tests are still thin. Before moving work off a frontier API, run your own tasks on the open model and compare.

Key Takeaways

• Keep AI incident records you could hand to a regulator. The FTC’s demands ask for security incidents, safeguards, marketing materials and privacy policies.
• Read the terms behind the business tools you use. Meta’s October 30 terms tie consent to scraping to the simple act of keeping a business account.
• Budget for patching when AI scans find flaws. Bank of America needed 1,000 extra contractors to fix what AI models found.
• Measure AI spend per person before it gets capped for you. Accenture’s $29 limit followed employees spending up to $1,000 a month.

Found First

Primary research AI Weekly published before the press covered it.

In the Labs

• RL gains in “no-think” mode are mostly hidden thinking. A KAUST-affiliated team trained Qwen3 models with RL in NO-THINK mode and found the gains come largely from reactivating the model’s existing THINK-mode reasoning. Swapping one small activation subspace removes most of the gain. Teams buying shorter outputs with this recipe may not be getting a more efficient model. Read our page on the paper.
• Frontier models complete at most 55% of rule tasks when the rules sit in separate files. RuleTrek tested 11 frontier models on tax, procurement, security policy and other rules spread across a file workspace. End-to-end scores ranged from 27% to 55%. A model that finds the right rule often still fails to apply it. Read our page on the benchmark.

Worth Reading

Wait, What?

• A court ordered a killer resentenced because the victim’s AI avatar was too moving. At a sentencing in Arizona, the victim’s sister played an AI-generated video of her brother forgiving his killer, and the judge said “I loved that video” while handing down the maximum term. An Arizona appeals court ruled the AI video carried “undue emotional weight” and ordered the judge to reconsider the sentence. The manslaughter conviction stands.

Worth Watching

The videos AI practitioners are passing around right now — curated on AI TV.

This week’s poll

Should regulators be able to see AI labs’ internal safety incident records?

Until next time,

Alexis