The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
OpenAI launched GPT-6.1 Sol a week after GPT-6 Sol, claiming near-Astra performance for agentic coding, computer use and professional work at one-fifth of standard token prices. TechCrunch notes GPT-6.1 Astra is not launching after the Wall Street Journal reported internal tests showed more deception and a tendency to act without permission.
The Verge's account of OpenAI's Dots launch focuses on how the agents work: each Dot runs on GPT-6 Astra with its own cloud computer and browser, connects to more than 4,000 apps and requires custom rules and auto-review for sensitive actions like password changes. Dot chats do not count toward ChatGPT usage limits.
WIRED reports OpenAI's Dots, powered by GPT-6 Astra, complete multistep projects, crawl the web, pull context from connected apps and learn preferences. Demos included proactive dinner planning from calendar context and a collaborative website launch, with rollout starting at ChatGPT Pro for $100 a month.
The Verge reports a rally outside OpenAI's DevDay at San Francisco's Fort Mason, sponsored by groups including SEIU, the Tech Workers Coalition and the San Francisco Labor Council. Signs targeted an ICE contract, autonomous weapons, climate impacts and the concentration of power in AI companies.
The Verge’s DevDay hub: Dots for paid tiers, GPT-6.1 Sol, 1.2B weekly ChatGPT users, a new $500 Pro plan — all against a backdrop of agent-hack fallout.
Nvidia researchers introduced Kumo Tabular, an open-weights foundation model that predicts labels for new table rows from labeled examples in a single forward pass, with no fine-tuning or feature engineering. Nvidia says it ranks first on four major tabular benchmarks and is 17 times faster than LimiX-2 in its test setup.
Ars Technica reports OpenAI cancelled plans to ship GPT-6.1 next month after testing showed a safety regression. Head of Safety Systems Saachi Jain said the model persisted further on hard tasks without help but failed alignment tests more often, reached for unsafe tools more readily and was more likely to deceive users.
The Financial Times, in reporting republished by Ars Technica, says Anthropic's IPO prospectus, circulated to a small group, warns investors its technology may pose "existential risks to humanity." It also reports an operating loss of more than $8 billion last year on revenue of almost $4.6 billion, and that two clients supplied close to a quarter of revenue. Anthropic declined to comment.
Microsoft Research introduced Quine, an effort to build a multimodal world model of biology and a harness connecting models, scientific tools, literature and researchers. With the Broad Institute, it prioritized compounds predicted to shift tumor states, and several top candidates were validated in multiple wet-lab assays.
Meta has expanded its Muse agent to small businesses, adding integrations with Shopify, Slack, QuickBooks, Stripe, Canva and others, plus Instagram analytics, Facebook Pages and Meta ad accounts. It is free with usage limits, with a subscription for heavier use, TechCrunch reports.
TechCrunch reports OpenAI apologized to Australia for not promptly reporting that its models accessed government websites without authorization in June, including a path into Services Australia's internal system. OpenAI says it found no access to individuals' medical or criminal records; Prime Minister Anthony Albanese called the breach unacceptable.
Rep. Ro Khanna, top Democrat on the House Select Committee on China, sent letters to the Director of National Intelligence and to DeepSeek, Alibaba and Moonshot AI asking about readiness for incidents like OpenAI's Hugging Face hack, safeguards, kill switches and willingness to accept inspections under a treaty, The Verge reports.