Alec Radford

See our appendix for more on Geodesic's compute procurement strategy and takeaways.

Acknowledgements We'd like to thank our funders at Coefficient Giving for making our compute procurement possible, in particular Aidan Hewitt for his support, as well as the other AI safety organisations with whom we've been exchanging information and advice throughout this process, namely James Collins, Matt Palisade, Nick Levine, Alec Radford, Stan Van Wingerden, James Marks, and Quentin Anthony.

This article was narrated by Type 3 Audio for Less Wrong.

It was published on September 16, 2026.

[laughs]

... you mentioned GPT-1, uh, and I, I cannot let any, uh, Alec Radford, uh, you know, mention es-escape.

Um, did you talk with him when he was training GPT-1? Like any, any, any sort of historical fun stories there that, that you might come up?

I, I did not like meet him a bunch of times.

For example, although Paul's theoretical justifications for iterated amplification referred a lot to the safety properties of imitation learning, all of these papers added RLHF for better performance.

This focus on engineering-style work was facilitated by Dario's push to scale up from GPT-1, which was mainly Alec Radford and Ilya Sutskiva's project, to GPT-2 and subsequently GPT-3, justifying this in significant part by arguing that it would help boost alignment research.

In Empire of AI, Karen Howe reports Dario telling her in 2019 that we want a language model that humans can give feedback on and interact with, where the language model is strong enough that we can really have a meaningful conversation about human values and preferences.

My understanding is that Paul opposed this strategy internally, but Jeffrey supported it.