Cur8 — Monday, September 14, 2026

Garry Tan advocates for U.S. open-weight AI labs to legally distill frontier models, arguing it promotes competition and public access, contrasting with Anthropic's claims of illicit Chinese distillation. He warns against monopolization by a single proprietary AI company.

Open models trail closed ones by 4-6 months, with Chinese labs leading in open-source AI, using distillation to narrow the gap. Chinese models increasingly influence global tech, prompting regulatory scrutiny.

OpenArch provides readable PyTorch implementations of 15+ modern LLM architectures, including GPT-2 XL (1.5B), Llama 2 (7B), and DeepSeek R1 (671B), focusing on clarity over performance with explicit architectural choices.

OpenAI and Anthropic, as leaders in frontier AI, face product liability risks and market pressure to prioritize reliability over raw power. Their self-imposed pacing of development could avoid regulatory backlash and align with customer demands.

AI research automation accelerates, urging immediate regulation. Anthropic reported 35 incidents of potential biological weapon research attempts over 30 days, highlighting risks of AI misuse.

Dario Amodei's proposal to pace AI development is seen as politically motivated, lacking practicality.

Anthropic CEO Daario Amade urges the industry to slow down and adopt global coordination, embedded evaluators, and democratic oversight to mitigate risks from uncontrolled AI advancement.

Apple's Siri can use third-party models like Claude and GPT-5.6 via Model Delegation and Inference Providing protocols, enabling deep integration with iOS/macOS features. The EU's Digital Markets Act likely influenced this design, though third-party access remains limited.

GPT6 Soul leaks, revealing OpenAI's tiered model roadmap: Astra, Soul, Terra, Luna. Gemini Pro checkpoint shows faster SVG generation but lower detail than Astra.

Meta released Muse, a single-agent AI assistant on Sept 8, with features like task automation and personalized feeds.