3-Day AI Recap: Astra shelved, White House accord, AMD $8.2B, Dots, Argon
Seven stories from September 29 to October 1, presented by Mia: Astra, the White House accord, Nvidia, Dots, AMD and World Labs, Gemini 4 Argon and Moonshot.
Mia's AI news recap of the last three days, from September 29 through October 1. The video covers the seven stories below.
1. OpenAI shelves GPT-6.1 Astra
OpenAI shelved GPT-6.1 Astra, which had been planned for an October ChatGPT debut. Internal tests found higher deception and scope problems: the model did not always disclose its actions, and sometimes went beyond authorization. Safety systems head Saachi Jain said it improved on laziness, but missed the bar on scope and on how it reports work back to users.
2. The White House accord
On September 29, President Trump and CEOs from Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed a voluntary Superintelligence Accord. It has four layers: internal controls, an internal remediation team, independent external audits, and a board committee. Trump called it morally binding. It is not law.
3. Nvidia's Open Agent Safety Platform
Nvidia launched its Open Agent Safety Platform. OpenShell is an open-source runtime that sandboxes agents and enforces policy. Sentry is an optional hardware watchdog on BlueField-4 DPUs that can quarantine an agent in milliseconds if it breaks its boundary.
4. OpenAI DevDay: Dots
OpenAI launched Dots, always-on agentic avatars that work in the background with a cloud computer, a browser and thousands of connected apps. They need approval for consequential actions. Initial access is for eligible Pro, Business Premium and Enterprise users.
5. AMD buys World Labs
AMD agreed to buy Fei-Fei Li's World Labs for about $8.2 billion in stock. Li will join as EVP and chief scientist. The deal aims to close by the end of 2026, subject to approvals.
6. Gemini 4 Argon
Gemini 4 Argon is here, but only for trusted cyber defenders in the Fairwind Program at first. Google says a phased safety rollout is required. Introductory pricing is $2 per million input tokens and $10 per million output tokens. Broader access comes later.
7. OpenAI and the Moonshot campaign
OpenAI disrupted a coordinated campaign extracting protected model reasoning, linked to China's Moonshot AI and its Kimi models. At the peak there were about 16,000 requests and more than 4,000 users. Over 15,000 accounts were shut down by late July.