The Week in Open-Source AI | 8.14

Mark Zuckerberg published a 6,500-word manifesto on AI this week, writing that instead of centralizing superintelligence, we should distribute it widely and give every person the ability to direct it. That could be my line! So I wrote a response in the Guardian. The short version: when the landlord starts quoting the tenants’ rights pamphlet, you should read the lease. Carefully.

So, this week is about reading the lease. Meta and NVIDIA released two different versions of open this week. Six giants who agree on nothing agreed on a folder format. And the one story I can’t stop thinking about? The approve button died this week. The last layer where you held the pen just moved from your finger into a model. A closed one. Run by a vendor.

Tell me what you think.

I’m reading everything:

• Zuckerberg’s 6,500-word manifesto

• My response in the Guardian (read the lease, carefully)

• The livestream (every two weeks, guests, open Q&A)

• This week’s essay: my genome, analyzed on a gaming PC

• Meta ships Muse Glimmer

• The demo: it finds your smart home hub and runs your house

• It fits on the graphics card a gamer already owns

• The license, read closely (Apache 2.0 — the file, still not the recipe)

• What Llama carried, for comparison

• NVIDIA ships Nemotron 3.5 Lightning

• The file, the training data, the recipes — actually open

• The whole kitchen

• Is Meta’s re-opening a PC play or a policy play?

• Jensen Huang: free AI should be great for chips

• And Kimi K3, the file that came with a leash (last episode)

• Agent Plugins (the folder format six rivals agreed on)

• The maintainers list (notice who’s missing)

• Vercel’s announcement (write once, run anywhere)

• MCP, invented at Anthropic

• Google’s walkthrough of the spec

• Tim O’Reilly: modularity, not features, was the moat

• 1998, the room where Netscape opened the browser

• That open sourcing has a name (the original Mozilla mission)

• The browser game: 40,000 plays, 409,000 decisions, one threat in three missed

• Anthropic’s own numbers (97% approved, 13.6% caught, 5% after fifty)

• Auto mode is now the default

• Cloudflare open sources its agent workspace (every agent starts with nothing)

• Mistral’s Shieldstral (an open safety classifier you run yourself)

• Mozilla’s State of Open Source AI report