Live·Open questions in longevity research
All news
Industry & policy

David Krueger recommends pausing AI development by replacing chips for training new models with chips for running existing ones

14 November 2026· 261114006

David Krueger recommends pausing AI development by replacing chips for training new models with chips for running existing ones

On November 10, the AI safety researcher called for an immediate, indefinite international halt to frontier model development because of the risk of human extinction. He is one of 26 coauthors of the Hardwired Pause proposal. In an essay expressing his own views, Krueger compares three plans and explains why the hardware approach is probably the best available option.

Standard AI accelerator chips can both train a neural network on data and run an existing model to perform tasks. Hardwired Pause proposes ending the production of training chips and switching to hardware for running approved existing models. Some designs embed the model in the circuitry itself; others use circuitry that checks whether the loaded model is an approved version. These designs are intended to make the chips much less suitable for further training.

Existing accelerators would first need to be located and brought under control, since they could still be used for training. During the transition, they would be used to run existing models, with checks to ensure compliance with that restriction. The old accelerators would then be taken out of circulation. Users would retain access to models, while the stock of hardware available to restart the race would shrink.

The AI Futures project's Plan A retains training capacity for controlled AI development and safety research. Its authors propose open research and the ability for each party to destroy the other's data centers if the agreement breaks down. Krueger calls the retained capacity "dry tinder": if the agreement fails, participants could immediately resume the race using hardware already in place. He worries that data centers could also be destroyed for other reasons, such as war.

Alongside training restrictions, Plan S from the Machine Intelligence Research Institute (MIRI) proposes bans on research that could bring superintelligence closer, meaning AI that surpasses humans at all intellectual tasks. The plan's authors worry that new methods could allow more capable models to be trained with less computing power. Krueger believes that these bans and their enforcement could increase surveillance of researchers and concentrate power.

Krueger argues that retaining existing AI would help sustain the pause. A country that violated the agreement would gain a smaller advantage over others than it would if AI chips had been abandoned entirely: other countries would still have powerful models. These models could also support safety research and collective decision-making. He therefore recommends the hardware approach to those who consider catastrophe less likely than he does and place greater value on current AI applications.

In 2025, Krueger founded the nonprofit Evitable to mobilize society against the race to superintelligence. His own plan calls for dismantling AI chips and their production supply chain. Krueger worries that retaining specialized chip production would make it easier to return to manufacturing training accelerators. New methods could reduce the computing power needed for training. Even existing models, he reasons, could become components of increasingly powerful systems as developers give them memory, tools, and the ability to interact with one another.

Overall, I still lean toward the plan to "get rid of compute."
Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#ai-safety#frontier-models#training-chips#inference-hardware#hardwired-pause