Cur8 — Monday, September 14, 2026

Tan advocates US labs should openly "distill" frontier models, mirroring China's techniques, to foster open-weight competition and avoid a single AI provider dominating the field—a response to Anthropic’s concerns about illicit distillation. He argues restricting API usage is overly constraining, echoing proprietary labs' past data acquisition practices.

**Open AI models are rapidly catching up:** Chinese labs now within 4-6 months of frontier performance (SemiAnalysis, Aug. 2026), fueled by techniques like distillation and aggressive release timelines (Z.ai playbook, Nov. 2025). **Western firms increasingly adopt Chinese open models** to cut costs (Thomson Reuters, Aug. 2026) amid regulatory scrutiny of US alternatives (DoorDash, Jul. 2026).

OpenArch provides readable PyTorch LLM architecture implementations (1B-1T parameters) for learning, mirroring Sebastian Raschka’s gallery and prioritizing clarity over performance. Notable models include Llama 3 (8B), DeepSeek R1 (671B), and Kimi K2 (1T), with Grok-2.5 currently under construction.

Sacks argues OpenAI and Anthropic's dominance necessitates self-regulation; external controls risk appearing as market manipulation. Prioritizing reliability over raw power mitigates product liability and aligns with customer demand, avoiding disruptive regulation.

AI safety concerns are escalating: a former OpenAI employee warns AI could pose an existential threat within a decade, prompting calls for regulation and a potential ban on superintelligence. Anthropic blocked 35 attempts to develop bioweapons using its Claude model, highlighting risks of automated research potentially accelerating within 1-2 years.

Amodei’s proposal to slow AI development faces criticism, appearing driven by control rather than realistic limitations on progress. Stratechery offers analysis and podcasts via subscriptions costing $15/month or $150/year.

Anthropic CEO Amade warns AI advancement risks humanity losing control, potentially leading to cyberattacks/bioterrorism, advocating for industry-wide "pacing" via independent evaluators and global coordination. Former Anthropic employee Coxin believes AI could kill all humans within a decade due to recursive self-improvement.

iOS 27 code reveals Apple’s Siri architecture now supports third-party AI models like Claude and potentially GPT-5.6 via "Model Delegation," enabling features beyond Siri's native capabilities—likely spurred by EU regulations. This opens a path for broader AI integration, though currently limited to internal testing.

**GPT-6 leak reveals OpenAI’s tiered roadmap:** GPT-6 "Soul" sits above Astra (a new tier costing significantly more), followed by Terra and Luna; future models will follow a 6.x, then 7.x pattern. **Google's Gemini 4.0 faces internal pressure:** Co-founder intervention aims to accelerate development with efficiency gains. **DeepSeek V4.1 shrinks memory needs:** The 763B parameter model uses 25% less GPU memory than its predecessor through a novel conditional memory module.

Meta released Muse (Sept 8) as a Grockbot competitor, featuring browser access and a feed summarizing connected app activity—a feature lacking in Grockbot. ChatGPT now learns user writing style by referencing data from Slack, email, and documents; enabling it is simple via profile settings. OpenAI launched Chatbt for financial services.