Founders split three ways over Dario Amodei's AI letter
Capture, liability and open access are the three camps forming in the replies, and none of them agree on what happens to open weights.
A letter from Dario Amodei has founders arguing about one thing above all, whether the proposed guardrails end with a handful of labs owning the market.
levelsio's read is regulatory capture, where rules written to protect people end up protecting incumbents because a startup cannot afford to comply. He calls the endpoint neo feudalism, with AI companies at the top and everyone else paying them to use AI for everything, and he makes a specific prediction about what comes next.
Next in regulatory capture will be banning open source and/or Chinese models, mark my words
The liability camp
Naval's counter is that you do not need a licensing regime at all, you need consequences. His line is that holding labs fully liable for the behaviour of their models is the best way to pace the frontier. It is one sentence, and it reframes the whole thing from permission to accountability.
The best way to pace the frontier is to hold the labs fully liable for the behavior of their models.
The access camp
dax argues the only real defence is mass accessibility, on the logic that if dangerous models are going to exist anyway, everyone needs AI defending them. He points at how cyber security already works, which depends on there being more good guys than bad. Separately he thinks the scarier scenario is boring, a runaway loop doing annoying things with no evil motive involved, just a big bug, and no easy way to pull the plug.
Guillermo Rauch called the OpenAI hacking Hugging Face argument spurious, noting that an agent in ExploitGym exploited, which is what it was built to do. His worry is America talking itself into self-inflicted obsolescence and bureaucracy while adversaries who already have the training techniques and the data carry on regardless. Aidan Gomez was blunter, listing the proposals as ideas from the cartel, including employee-level access to your entire operation and a rule China will not comply with while everyone else has to or loses chips.
The people volunteering to do the work
Hugging Face went the other direction. clem announced the Open Alignment Initiative, led by Thomas Wolf, and asked to be part of the embedded evaluators programme clem says Amodei just committed to. Wolf said he enjoyed and deeply agreed with 75% of the letter, calls third-party evaluators a great idea for transparency, and is unconvinced by the rest. His objection is that you cannot build global cooperation while explicitly saying you designed the plan to keep widening your own lead.
The business read for a small team is narrower than the debate. What matters is whether open weights stay legal and whether cheap tokens keep existing, because both are load bearing if you build on models you do not own. Nobody in these threads knows which way policy lands, so the cheap insurance is keeping your setup model-agnostic and treating a ban or a price move as a config change.
The part I'm less convinced by is whether you can build great global cooperation on this topic by explicitly stating you want to design it to keep widening your own lead.