An open letter to Dario: if you mean it, open the weights

If all publicly available AI models must be open-weights, the cost to train the next generation of frontier models will decrease.

An open letter proposes that AI frontier companies should be legally required to release the weights of any AI model offered to the public. This proposal, addressed to Anthropic CEO Dario Amodei, argues that open-weighting publicly available models would intrinsically slow down AI progress by reducing the financial incentives for massive training runs. The author suggests this approach is more effective than government regulation, which they claim is susceptible to regulatory capture by established labs. The letter highlights Amodei's past actions and Anthropic's Public Benefit Corporation status as reasons he might be amenable to such a drastic, self-sacrificing measure. This contrasts with current regulatory proposals like embedded third-party evaluators, which the author believes big companies can manipulate to maintain their lead.

The author advocates for mandatory open-weight releases of public AI models, positioning it as a more effective method for pacing AI development than regulatory oversight. By forcing the release of model weights, the economic model supporting frontier AI development, which relies on proprietary technology for high valuations, would be fundamentally altered. This could reduce the immense capital flowing into training runs and thus decelerate progress across the industry more broadly and equitably than current regulatory proposals.

This proposal directly challenges the current paradigm where advanced AI models are kept proprietary to secure funding and competitive advantage. If enacted, it would democratize access to state-of-the-art AI capabilities for researchers and smaller companies, while simultaneously disincentivizing massive, exclusive investments in future training. The success of such a proposal hinges on a key industry leader like Anthropic embracing it, despite the immediate financial implications for all major AI labs.