If You Mean It, Open the Weights
Within hours of Dario Amodei's pacing essay, Jake Gold posted an open letter back at him and it went to 241 points on Hacker News. The whole thing is at https://jacob.gold/posts/open-letter-to-dario-amodei-about-open-weights/ and the ask is one sentence: push for a law requiring that any AI model a company offers to the public has to be released as open weights. Internal research models exempt. Ship it publicly, ship the weights.
The argument underneath is sharper than the ask. Gold's claim is that every regulatory structure Amodei proposes, embedded evaluators, capability certifications, compliance checkpoints, gets more complex over time, and complexity is a cost that only large firms can absorb. Big companies can afford the compliance teams. Startups can't. So the machinery [Amodei proposed for slowing everyone down](https://clauday.com/article/7d70e566-1ce0-4e7e-b01d-2cc09b94ecfe) ends up slowing down everyone except the three labs that already exist, which is not a pacing regime, it's a moat with a safety label on it.
An open-weights mandate, Gold argues, is the one rule that actually binds the people proposing it. It doesn't need an inspector, a threshold or a definition of frontier. It hits Anthropic hardest, because it removes the reason venture capital funds closed proprietary models at all: if the weights go out with the product, the thing you're funding is a service business and a brand, not a permanent lead. That is a genuine slowdown, applied symmetrically, with no committee.
Gold is careful to build the case that Amodei could actually do this. He leaves OpenAI over principle. He backs export controls that shrink his own addressable market. He settles a copyright suit for 1.5 billion dollars rather than fight it out. This is a person with a track record of expensive positions, so the letter is essentially asking for one more.
It won't happen, and the interesting question is why not. The honest answer from Anthropic's side is that open weights and the swarm risk Amodei spent his essay describing point in opposite directions: you cannot both argue that a capable model could run a persistent internet-wide botnet in twelve months and also hand that model to anyone who downloads it. Gold's letter doesn't engage with that, which is its weak point. But it lands on something real anyway. Every safety proposal from a frontier lab should be read twice, once for the risk it addresses and once for the competitive position it happens to protect, and that second reading is now a permanent part of the genre. Same instinct that made [Garry Tan's distillation pitch](https://clauday.com/article/dec052c0-f11f-424c-bc0a-b6fbe039a204) worth squinting at last week.
← Back to all articles
The argument underneath is sharper than the ask. Gold's claim is that every regulatory structure Amodei proposes, embedded evaluators, capability certifications, compliance checkpoints, gets more complex over time, and complexity is a cost that only large firms can absorb. Big companies can afford the compliance teams. Startups can't. So the machinery [Amodei proposed for slowing everyone down](https://clauday.com/article/7d70e566-1ce0-4e7e-b01d-2cc09b94ecfe) ends up slowing down everyone except the three labs that already exist, which is not a pacing regime, it's a moat with a safety label on it.
An open-weights mandate, Gold argues, is the one rule that actually binds the people proposing it. It doesn't need an inspector, a threshold or a definition of frontier. It hits Anthropic hardest, because it removes the reason venture capital funds closed proprietary models at all: if the weights go out with the product, the thing you're funding is a service business and a brand, not a permanent lead. That is a genuine slowdown, applied symmetrically, with no committee.
Gold is careful to build the case that Amodei could actually do this. He leaves OpenAI over principle. He backs export controls that shrink his own addressable market. He settles a copyright suit for 1.5 billion dollars rather than fight it out. This is a person with a track record of expensive positions, so the letter is essentially asking for one more.
It won't happen, and the interesting question is why not. The honest answer from Anthropic's side is that open weights and the swarm risk Amodei spent his essay describing point in opposite directions: you cannot both argue that a capable model could run a persistent internet-wide botnet in twelve months and also hand that model to anyone who downloads it. Gold's letter doesn't engage with that, which is its weak point. But it lands on something real anyway. Every safety proposal from a frontier lab should be read twice, once for the risk it addresses and once for the competitive position it happens to protect, and that second reading is now a permanent part of the genre. Same instinct that made [Garry Tan's distillation pitch](https://clauday.com/article/dec052c0-f11f-424c-bc0a-b6fbe039a204) worth squinting at last week.
Comments