Our readingWhat is a korrent?
Drawn from what Yoshua Bengio said
neural networks The architecture almost all modern machine learning is built on, and what changes when it gets deeper.
At the same time, in order to reduce the probability of someone intentionally or unintentionally bringing about a rogue AI, we need to increase governance and we should consider limiting access to the large-scale generalist AI systems that could be weaponized, which would mean that the code and neural net parameters would not be shared in open-source and some of the important engineering tricks to make them work would not be shared either.
Added to korrents 29 Aug 2026 · How quotes work · Something wrong? Tell us
Do you hold this korrent?Do you also believe this?
I hold this tooCannot both be true
Keeping powerful models available only to a chosen few is not a good strategy.
Altman says keeping powerful models available only to a chosen few is not a good strategy. Bengio says access to systems that could be weaponised should be limited, and the weights not released. A chosen few is exactly what limiting access produces, and only one of them thinks that is the failure mode.
Cannot both be true
Ball argues that no ban on open source can prevent self-sovereign AI and that more open-weight models may be the better defence. Bengio argues that weights capable of being weaponized should be withheld. They cannot both be acted on: one treats openness as the mitigation and the other as the hazard.
Cannot both be true
Bengio says the code and parameters of large generalist systems that could be weaponized should not be released. Huang says open models are fundamentally necessary for most industries, countries and researchers to take part at all. This is the same release decision answered in opposite directions.
Cannot both be true
Restricting US AI models does not work: China builds better ones and everyone switches to them.
Levels says restricting US models simply moves everyone onto better Chinese ones, so the restriction fails on its own terms. Bengio says the dangerous weights should be withheld anyway. The disagreement is factual before it is moral — whether withholding accomplishes anything at all.
In tension with
Toner, Brundage and Yudkowsky hold jointly that any given capability becomes cheap and widely available within a couple of years of first being reached. Bengio argues that weights capable of being weaponised should not be released. If the diffusion claim holds, withholding buys a delay rather than a prevention — which is an argument for doing it carefully, not an argument that it is pointless, and neither side says which.
In tension with
Ravikant’s claim is about motive: labs are incentivised toward fear because it justifies staying closed. Bengio, who runs no lab, argues for exactly the closure Ravikant says is being sold under false pretences. The tension is that Ravikant’s explanation does not reach the person making the strongest version of the case.
Shared subject first, then similar wording. They may agree or disagree.