Cur8 — Monday, September 14, 2026

U.S. open-weight AI labs should freely distill frontier models, argues Y Combinator CEO Garry Tan, countering Anthropic's call for regulation on Chinese labs' alleged illicit distillation.

Open models lag behind closed models by 4-6 months; Chinese labs lead in open model performance since 2024. Distillation aids Chinese labs but doesn't negate their innovation; regulatory attention grows on Western companies using Chinese models.

OpenArch offers PyTorch implementations of 15 modern LLM architectures, focusing on clarity and learning; includes models like GPT-2 XL (1.5B), Llama 3 (8B), and Kimi K2 (1T).

OpenAI and Anthropic can self-regulate frontier models without external oversight; market pressures already incentivize reliability over raw power. Demanding specific regulations may be seen as blackmail.

Lawmakers push for AI regulation after warnings from former researchers; Congress urged to act within 1-2 years to prevent uncontrolled AI progress. Anthropic blocked 35 incidents of potential biological weapon research over 30 days.

Dario Amodei proposes pacing AI development for political control, an unrealistic approach.

Anthropic CEO agrees with ex-employee's warning: AI could pose extinction risk within a decade due to recursive self-improvement. Proposes industry-wide coordination to pace model development, endorsed by OpenAI's Sam Altman and Elon Musk. Current models' risk deemed low.

Apple's new Siri architecture in iOS 27 and macOS Golden Gate supports deep integration with third-party AI models like Claude and ChatGPT via Model Delegation and Inference Providing protocols, enabling tasks like setting reminders and creating files. EU's Digital Markets Act may have influenced this interoperability.

GPT-6 Soul leaked, reshaping OpenAI's tier structure; Gemini 4.0 checkpoint shows efficiency gains but lacks detail; DeepSeek V4.1 Flash offers 763B parameters with reduced GPU memory.

Meta released Muse AI agent (Sept 8), competing with GrokBot; features include a secure browser, task management, and personalized feeds. ChatGPT now learns writing style from connected apps.