Snippets.

// AI Briefing · October 16, 2026

AI Briefing

Watch on YouTube ↗Subscribe ↗

Anthropic is embedding six researchers from the evaluation nonprofit METR with internal tools, laptops and office space, and METR can publish its assessments without Anthropic being able to block them. Meta released Muse Spark Open, an 816B-parameter mixture-of-experts open-weight model with a 1M-token context, tuned with Apple to run locally on paired M5 Ultra Mac Studios. In a long essay that drew heavy discussion on Hacker News, Goodfire's Nick Cammarata argues that most "messy" neurons come from fixable problems like outdated network design and weak visualization tools. Beijing proposed adding 23 AI technologies to its export-control list, including large language models, agents, chip designs and computing clusters.

In this video

  1. 01Anthropic Gives METR Employee-Level Access and Gives Up Its Veto Over METR's Findings
  2. 02Meta and Apple Launch Muse Spark Open, an 816B Open-Weight Model Built for Mac Studios
  3. 03Goodfire Researcher Argues Neural Networks Are Already Easy to Read, Challenging a Core Belief of AI Safety Research
  4. 04China Proposes Export Controls on AI Models, Agents and Chip Designs

The stories

01PolicyTechCrunch ↗

Anthropic Gives METR Employee-Level Access and Gives Up Its Veto Over METR's Findings

Anthropic is embedding six researchers from the evaluation nonprofit METR with internal tools, laptops and office space, and METR can publish its assessments without Anthropic being able to block them. The move comes days after reports that Anthropic and OpenAI wanted control over what outside auditors could publish. Anthropic is now committing to the strongest independent-oversight model of any frontier lab. It can still require redactions in narrow categories such as customer data and model architecture details.

02Open SourceTechCrunch ↗

Meta and Apple Launch Muse Spark Open, an 816B Open-Weight Model Built for Mac Studios

Meta released Muse Spark Open, an 816B-parameter mixture-of-experts open-weight model with a 1M-token context, tuned with Apple to run locally on paired M5 Ultra Mac Studios. Meta calls it the most capable open-weight model in the US, and it arrives as Meta keeps its top Muse Wonder models closed. Running it locally takes about 1TB of RAM, but the Apple partnership makes high-end Macs a serious platform for frontier-class local AI.

03ResearchGoodfire ↗

Goodfire Researcher Argues Neural Networks Are Already Easy to Read, Challenging a Core Belief of AI Safety Research

In a long essay that drew heavy discussion on Hacker News, Goodfire's Nick Cammarata argues that most "messy" neurons come from fixable problems like outdated network design and weak visualization tools. Many researchers instead blame an effect called superposition, where unrelated concepts get packed into the same neurons. He reinterprets famous examples, such as the "cat and car" neuron, and in early checks of Qwen3-32B he estimates that about 90% of neuron directions have a single main job. Why it matters: if he is right, understanding what AI models do inside may be far easier than the field assumed. He stresses that the numbers are rough working estimates.

04PolicyBloomberg ↗

China Proposes Export Controls on AI Models, Agents and Chip Designs

Beijing proposed adding 23 AI technologies to its export-control list, including large language models, agents, chip designs and computing clusters. Model weights and source code above certain performance thresholds would be restricted. It's a notable shift because it targets trained AI systems themselves, not just hardware, and could reshape the open-weight ecosystem Chinese labs have fed.

Previous digests

Let's talk on WhatsApp