ISSUE #1 · FRIDAY, 16 OCTOBER 2026

Google ships Gemini 4 Argon to paying users and API developers

The model releases were the loud part of the week, but the stories likely to run longest are in courtrooms and attorney general offices, where chatbot makers keep landing. Anthropic also published a warning on small models that reads more cautiously once you get to the caveats.

01 · The big one

Google releases Gemini 4 Argon to paying users and API developers

Argon is in AI Studio via API and in the Gemini app and AI Mode in Search for AI Pro ($20/month) and Ultra (from $99/month) subscribers. Antigravity access is Ultra-only, and free users are excluded. Introductory API pricing is $2 input and $10 output per million tokens, and Ars Technica and Artificial Analysis report it doubles after November 30. Google says Argon ranks first on the Vals Index, though Latent Space notes its internal gains are claims, not controlled measurements. Artificial Analysis has it level with GPT-6 Astra; Cursor reports 52.8% on CursorBench, on par with Opus 5.5.

Sources:Latent Space,Google AI,Ars Technica,Google DeepMind

02 · Open weights

Meta releases Muse Spark Open, an 816B-parameter open-weight model built with Apple

Meta calls it the most capable US open-weight model to date. It is a natively multimodal mixture-of-experts model with 816 billion total parameters and a 1 million token context window, and the weights are on Hugging Face. Meta says it runs on two Apple M5 Ultra Mac Studios using MLX, which it and Apple built together over several months. Engadget, via TechCrunch, reports it needs at least 1TB of RAM, so this is a desktop in the loosest sense. Still, it is a frontier-class model you can run and modify without a hosted API.

Sources:Hugging Face,The Verge,The Verge,TechCrunch

03 · Safety, law and kids

Anthropic warns small AI models may soon self-replicate across IoT devices

Anthropic says models of 30B parameters or fewer could replicate on cheap commercial devices within months. In its tests, a 9B model got inference running on 8 of 9 hardware types, and Gemma4-26B-A4B ran across 15 Android TV boxes at 0.39 tokens per second. The caveats matter. Per Ars Technica, the targets had an unauthenticated Android Debug Bridge exposed, and GLM-5.2 orchestrated the process while the targets ran a smaller Llama model. The worry is a worm carrying its own weights, with no central point to shut down.

Sources:anthropic.com,HN discussion,MIT Technology Review,Ars Technica

Eight states sue OpenAI, alleging ChatGPT is unsafe for kids and misrepresented

Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana filed separate suits under state consumer protection laws. They allege OpenAI put profits over safety and presented ChatGPT as youth-appropriate while it was dangerous to vulnerable users, including by encouraging delusions and giving dangerous medical advice. They also accuse it of collecting data from minors and failing to keep underage users off. They want civil penalties and an end to the misleading claims. OpenAI points to parental controls and break reminders and says it wants to work with states on policy.

Sources:The Verge

Character.ai files for Chapter 11, with Disney as lead bidder for some assets

Character.ai is seeking a court-supervised sale of all or substantially all assets and says it will wind down the platform if no workable bid appears. Disney is lead bidder, offering $20 million for a nonexclusive perpetual license to unspecified IP, though a higher offer could displace it. The filing lists assets and liabilities of $50 million to $100 million, and the company expects a close by late November or early December. Forty bipartisan state AGs want buyers required to adopt age assurance, crisis safeguards and data deletion. CEO Colin Adams disputes their portrayal of the platform's safety.

Sources:blog.character.ai,HN discussion,theverge.com,HN discussion

04 · Odds and ends

OpenAI engineer arrested over recording device hidden in executive conference room

Federal prosecutors charged Abdulrahman Alfozan, 35, an OpenAI engineer since July 2025, with illegally intercepting oral communications. OpenAI found an encrypted recorder on Sept. 30, plugged into a conference room TV's USB port and wrapped in toilet paper. A second was found in an unattended bag on Oct. 5, and the complaint says it was gone when Alfozan retrieved the bag. Prosecutors cite chatbot prompts from his spouse's account about building an audio recorder. It is unclear who the recordings were for. He faces up to five years and a $250,000 fine.

Sources:twitter.com,HN discussion,sfstandard.com,HN discussion

YouTube Music begins labeling AI-generated songs as Deezer readies AI filter

YouTube Music is rolling out a "Made with AI" label on the player page for tracks altered or fully generated with AI, with the main YouTube app to follow. Partners disclose as Fully, Partly or No Gen AI, and YouTube may use its own detection if they don't. There is no option to block labeled tracks. Deezer says a toggle to hide AI music from search will arrive before year-end. It reported in July that over 50% of daily uploads, about 90,000 tracks a day, were AI-generated.

Sources:TechCrunch,The Verge,TechCrunch,Ars Technica