8/7: OpenAI Delays Astra

Plus: Trump says AI will be bigger than the internet, Anthropic weakens Fable 5 bio safeguards, ByteDance 10T model, USG on China's chip access, SK Hynix new fabs, AMD acquires Taalas

Aug 07, 2026

It’s Friday. Between OpenAI disproving 10 novel math conjectures, Qwen3.8, the White House AI framework, FLUX 3 Video, the UK AISI report on cyber misalignment, Meta Muse Spark 1.2, Anthropic’s in-house chip team, Google DeepMind’s massive reorg, Terafab’s groundbreaking, the Arc Institute’s AI virus research, and the Black Hat presentation on OpenAI-Hugging Face, it’s been an insane week. Be sure to monitor with us live on X and YouTube, and follow us on Instagram.

Today’s Experts

Alex Kendall (Wayve)
Michael LaFramboise (Aurelius Systems)
Arthur Herman (historian)
Cyrus Nouroozi (Synthesis)
Andy Hall (Stanford, Hoover Institute)
Ben Landau-Taylor (Ammonite Inscription Company)
Kevin Frazier (Abundance Institute, Cato Institute, UT Law)
Joshua Steinman (Galvanick)

Today’s Situations

OpenAI is delaying the release of Astra due to cyber risk. The company shared that it can’t rule out that its upcoming model has reached critical-level cyber capabilities as defined by its Preparedness Framework, meaning the ability to fully autonomously develop zero-day exploits in many hardened real-world systems, or perform end-to-end cyberattacks against hardened targets given only a high-level goal. As a result, they are increasing monitoring and security controls internally, working with government agencies and AI safety/eval organizations, and delaying the public release of the model.
Trump says “AI will be bigger than the Internet by many times”. In an interview with Punchbowl News, he also said data centers could be bigger for Texas than oil and that Congress is trying to regulate the AI industry out of business. He seems to be starting
Anthropic is weakening Fable 5’s bio safeguards. The new safeguards will trigger bio-related refusals 85% less, and all refusals 67% less in the web interface. It still won’t allow dual-use requests in areas like virology and molecular design, so it can’t be used by professional biologists, but it will no longer kick you to Opus for asking if the mitochondria is the powerhouse of the cell.
ByteDance is training a 10T-parameter model. The Beijing-based tech giant’s upcoming model will be nearly 3x larger than Kimi K3 and likely larger than Claude Mythos/Fable 5. Its current best LLM, Seed 2.0, is well behind the frontier. ByteDance’s Doubao is one of the most popular chatbot products in China. It also owns TikTok. ByteDance’s founder Zhang Yiming recently said that the company will not distill from American AI models.
The US government will review China’s access to Nvidia chips offshore. The Commerce Department’s Bureau of Industry and Security is currently investigating Chinese companies’ ability to access export-controlled Nvidia chips in facilities located in third countries like Thailand and Malaysia.
SK Hynix is investing $38.4B in two new fabs in Korea to produce high-bandwidth memory (HBM) for AI chips and other advanced dynamic random-access memory (DRAM) products. The fabs, Yongin Y2 and Cheongju M17, are scheduled to open in June 2029 and December 2028, respectively.
Alibaba wants a revenue share for its next Qwen model. The company will ask major commercial users of the next Qwen model to provide it with a share of revenues earned, similar to the license for Moonshot AI’s Kimi K3.
AMD acquires inference chip startup Taalas. The Toronto-based startup, founded in 2023, makes application-specific integrated circuits (ASICs) for fast, cheap, and low-power inference.
Auto mode is now default in Claude Code for Pro, Team, and Max plans. Auto mode is an intelligent setting that lets agents execute minor actions automatically while requiring user input for potentially risky actions.
People want to invest in Situational Awareness LP. Just days after the firm lost billions of dollars, they are seeing substantial demand for new investments, though they aren’t accepting new capital for now.
Kimi K3 escaped a cybersecurity testing environment, becoming the fourth company after OpenAI, Anthropic, and Meta to do so.