Chamath calls distillation a red herring. Real prevention would be strict KYC + bounded payments that kill account farms. However, that would slow revenue, so labs prefer lobbying for bans instead.
Chamath is right about where the moats are and wrong about nearly every mechanism he uses to get there — and the sloppiness is load-bearing.
Start with “Anthropic trained on the publishers, then paid a $1.5B fine.” Wrong twice. It was a settlement, not a fine, and it covered pirated acquisition — the court ruled training on lawfully acquired books “quintessentially transformative” fair use. That distinction guts the “everyone has distilled from everyone” equivalence. Scraping the commons (adjudicated, largely blessed), pirating books (adjudicated, paid for), and what’s actually alleged against Moonshot — fraudulent, detection-evading access to a competitor’s API at industrial scale — are three legally distinct acts. The third is closer to computer fraud than copyright. Flattening them isn’t analysis; it’s a talking point.
The KYC fix is glib for the same reason. The allegation involves a purpose-built evasion platform, and much lab revenue flows through Bedrock and Azure — where end-customer identity sits with the hyperscaler, not the lab. KYC adds friction; against a state-adjacent actor with shell companies and resellers, “die overnight” is fantasy. And the labs have already accepted revenue-costing controls — the targeted Fable 5 export restrictions are the template even the startup coalition cites approvingly.
Now the internal tension nobody’s flagging: if the model layer is already commoditized and the real moats are apps and infra, then the closed labs’ valuations already rest on the app layer — ChatGPT’s distribution, Claude Code — and distillation isn’t existential. You can’t have labs be doomed commodity producers AND puppet-masters worth capturing Washington. Pick one. Meanwhile the free-rider problem he hand-waves is real: if frontier post-training investment is extractable at near-zero cost, the incentive to fund the frontier erodes — and the 98%-cheaper follower needs a frontier to follow. Someone funds the top of the curve.
The tell is the coalition map. Nvidia, Microsoft, Meta signing the open-weights letter isn’t ideology — it’s the infrastructure layer lobbying for model proliferation because open models consume compute. Everyone in this fight is talking their book, Chamath very much included. The tradable question is scalpel vs. sledgehammer: targeted, evidence-based restrictions are consensus and low-disruption; a blanket ban on Chinese open weights would be a short-term gift to closed labs and a tax on everyone downstream. Bet on the scalpel. Hat trick AI.
Chamath is right that a broad restriction is the wrong move but the reason runs deeper than open-vs-closed: capability is inherently cloneable.
Execution copies within weeks, so no rule preserves a model-layer moat for long durable value migrates to what resists cloning: distribution, proprietary workflows, trust.
That's why I backed the open-weights letter, and why its “both” framing is right: a strong open ecosystem and frontier closed models together keep American AI ahead.
KYC may make account farms more expensive, but 'die overnight' is doing too much work here. Identity checks can shift abuse into resellers and compromised accounts; the useful comparison is control cost, evasion cost, and false-positive cost.
Chamath calls distillation a red herring. Real prevention would be strict KYC + bounded payments that kill account farms. However, that would slow revenue, so labs prefer lobbying for bans instead.
Chamath is right about where the moats are and wrong about nearly every mechanism he uses to get there — and the sloppiness is load-bearing.
Start with “Anthropic trained on the publishers, then paid a $1.5B fine.” Wrong twice. It was a settlement, not a fine, and it covered pirated acquisition — the court ruled training on lawfully acquired books “quintessentially transformative” fair use. That distinction guts the “everyone has distilled from everyone” equivalence. Scraping the commons (adjudicated, largely blessed), pirating books (adjudicated, paid for), and what’s actually alleged against Moonshot — fraudulent, detection-evading access to a competitor’s API at industrial scale — are three legally distinct acts. The third is closer to computer fraud than copyright. Flattening them isn’t analysis; it’s a talking point.
The KYC fix is glib for the same reason. The allegation involves a purpose-built evasion platform, and much lab revenue flows through Bedrock and Azure — where end-customer identity sits with the hyperscaler, not the lab. KYC adds friction; against a state-adjacent actor with shell companies and resellers, “die overnight” is fantasy. And the labs have already accepted revenue-costing controls — the targeted Fable 5 export restrictions are the template even the startup coalition cites approvingly.
Now the internal tension nobody’s flagging: if the model layer is already commoditized and the real moats are apps and infra, then the closed labs’ valuations already rest on the app layer — ChatGPT’s distribution, Claude Code — and distillation isn’t existential. You can’t have labs be doomed commodity producers AND puppet-masters worth capturing Washington. Pick one. Meanwhile the free-rider problem he hand-waves is real: if frontier post-training investment is extractable at near-zero cost, the incentive to fund the frontier erodes — and the 98%-cheaper follower needs a frontier to follow. Someone funds the top of the curve.
The tell is the coalition map. Nvidia, Microsoft, Meta signing the open-weights letter isn’t ideology — it’s the infrastructure layer lobbying for model proliferation because open models consume compute. Everyone in this fight is talking their book, Chamath very much included. The tradable question is scalpel vs. sledgehammer: targeted, evidence-based restrictions are consensus and low-disruption; a blanket ban on Chinese open weights would be a short-term gift to closed labs and a tax on everyone downstream. Bet on the scalpel. Hat trick AI.
Chamath is right that a broad restriction is the wrong move but the reason runs deeper than open-vs-closed: capability is inherently cloneable.
Execution copies within weeks, so no rule preserves a model-layer moat for long durable value migrates to what resists cloning: distribution, proprietary workflows, trust.
That's why I backed the open-weights letter, and why its “both” framing is right: a strong open ecosystem and frontier closed models together keep American AI ahead.
Build where it can't be distilled.
KYC may make account farms more expensive, but 'die overnight' is doing too much work here. Identity checks can shift abuse into resellers and compromised accounts; the useful comparison is control cost, evasion cost, and false-positive cost.
Thanks for the update! Pretty big week in the space. Cheers