Latest issue · May 26, 2026

The open-weights argument

Whether frontier models should ship with their weights open has hardened into two coherent camps that aren't really speaking to each other any more.

Two camps that don’t really meet

The case for open weights, made most loudly by Yann LeCun at Meta and by the a16z crowd, runs roughly like this. Closed models concentrate power in a handful of labs whose decisions about safety and pricing nobody outside those labs can audit. Open weights mean researchers can actually study the models, and startups can compete without paying rent to the frontier labs. Regulators get to see what is going into the systems they are being asked to govern. Meta has been publishing Llama since 2023 on roughly those grounds. Marc Andreessen pushes the broader version in his “Techno-Optimist Manifesto”: treat AI like every other general-purpose technology, and the gains from broad distribution swamp the risks.

The case against, made most carefully by Dan Hendrycks at CAIS and by Yoshua Bengio in the Science paper he co-signed with twenty-odd senior researchers, is narrower than it sounds. Nobody serious in that camp is arguing against open research or open code. The argument is specifically that releasing the weights of the most capable models is irreversible, and that “most capable” is a moving target. Once a model is out, the safety work you did before release is the only safety work that will ever apply to it. If a future Llama crosses one of the bright lines on biological uplift or large-scale cyber operations sketched in Anthropic’s Responsible Scaling Policy, recalling it is impossible. The downside, in their framing, is asymmetric in a way most technology arguments are not. An open model that turns out to be safer than feared costs little. One that turns out to be more dangerous than expected costs everything.

The two camps disagree less about the present than about the next two years. Most of the open-weights camp will admit that today’s Llama is not the model the safety camp is really worried about. Most of the anti-camp will admit that today’s Llama does not need to be pulled. They diverge on what to do about the trajectory: “decide later” is itself a decision, and the safety camp thinks it’s the wrong one. The cleanest version of the disagreement is therefore about thresholds. Bengio’s paper proposes specific compute and capability lines that, if crossed, should change how models get released. Meta has not offered an equivalent line, and neither has Andreessen. Asking for one is not a rhetorical trap. It is the form the argument has to take to actually be settled.