Sholto Douglas: evaluation

Transcript context

…That's such an interesting idea. I'm still confused on what will be the bottleneck. What would have to be true of an agent such that it sped up your research? So in the Alec Radford example where he apparently already has the equivalent of Copilot for his Jupyter notebook experiments, is it just that if he had enough of those he would be a dramatically faster researcher? So you're not automating the humans, you're just making the most effective researchers who have great taste, more effective and running the experiments for them? You're still working at the point at which the intelligence explosion is happening? Is that what you're saying? Right, and if that were directly true then why can't we scale our current research teams better? I think that’s an interesting question to ask. If this work is so valuable, why can't we take hundreds or thousands of people–they're definitely out there–and scale our organizations better. I think we are less, at the moment, bound by the sheer engineering work of making these things than we are by compute to run and get signal, and taste in terms of what the actual right thing to do is. And then making those difficult inferences on imperfect information, For the Gemini team. Because I think for interpretability, we actually really want to keep hiring talented engineers. I think that's a big bottleneck for us.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.