The Week in Open-Source AI | 8.21

What is it with AI CEOs and their manifestos? Last week it was Zuckerberg’s 6,500 words. Back in June, Dario Amodei published a full tome. But this week Dario did something rarer than a manifesto - he logged on, posted all weekend, and diagnosed the AI backlash as “a crisis of trust.” Two days earlier, a CNBC poll had asked young Americans about nine AI leaders by name. Every single one came back majority-distrusted. So the polls say he’s right - which is the most dangerous kind of right there is.

So this week is about trust: who asks for it, and who builds it. Two fintechs went to war over who meters AI. And the story I can’t stop thinking about? The safety trained into open-weight models strips out in minutes. The same week, Google shipped the other kind of protection: encryption that computes on your data without ever seeing it. Not “promises not to look.” Cannot look. One is a promise. The other is architecture. Words don’t mean much. Architecture does.

Tell me what you think.

I’m reading everything:

• Zuckerberg’s 6,500-word manifesto (last week’s tome, for the rhyme)

• Dario’s June tome: “Policy on the AI Exponential”

• The weekend thread: “fundamentally a crisis of trust”

• Same weekend: “I’m not an AI pessimist”

• The thread went everywhere

• The CNBC/Generation Lab poll: nine AI leaders, nine majority-distrusts

• The per-leader breakdown, legibly

• Same poll, second read

• The livestream (Tuesday — guests, open Q&A)

• This week’s essay: one model name, four different outcomes

• Andrew Carnegie, the man, before the essay

• The Gospel of Wealth, in his own words

• The Gospel, with its publication history

• Homestead, 1892 — the stress test

• Homestead, with the contemporary illustrations

• Anthropic’s usage policy (the switch, in writing)

• Download a model today — a file on your disk is architecture

• Stripe buys OpenRouter, north of $7 billion

• The 400+ models behind one API

• Robber barons — the ORIGINAL ones

• Pfalzgrafenstein: a toll castle in the middle of the river — the picture that explains everything

• The Rhine Gorge: castle after castle. There was never just one.

• The narrow spot itself

• Ramp opens router.com — free through 2026, and the CTO quote that gives the game away

• The second castle

• Hugging Face’s State of Open Models: the 83%

• LiteLLM — downloadable code that routes on your terms (this kind of thing; much more next week)

• Fool’s Gold: the safety strips out in minutes — the abstract says it plainly

• Kerckhoffs’s principle, 1883: the only secret is the key

• The man himself

• Google’s HEIR: the model computes without ever seeing your words

• HEIR on GitHub — the downloadable thing

• Mozilla’s State of Open Source AI report