AI will be doing most AI research at lower time horizons than most expect. Generally superhuman AI researchers aren't necessary. AI just needs to be roughly human-level at the tip of one of the spikes of its jagged frontier.
Serious AI research takes hundreds of hours to do. This doesn't mean we need a 50% time horizon of hundreds of hours. We need a 5% time horizon of hundreds of hours. The common wisdom that what really matters is the 80% time horizon is inverted when it comes to automating AI research. Here, what matters is that insight is achieved and recognised.
Recognition might be a hindrance—today's models often miss when their work is flawed (although they're improving rapidly at this—but autonomous experimentation and proof automation etc. address this sufficiently. I'll write more about that tomorrow I think. It's important.
Once AI is at this point the nature of AI research changes dramatically. Instead of having 10,000 human AI researchers with some coding assistants working on AI, you have millions of GPUs across the top labs working on AI. Perhaps a 10x increase in the pace of algorithmic progress over a very short period of time.
One might respond that the low-hanging fruit in that spike will quickly be picked by the millions of GPUs. Whether this is true (unclear to me) 2.5 months later, time horizon will have doubled and suddenly the models spikes will be further out, human-level in more areas, and now superhuman in a few important domains of research. Then the pace of AI development really picks up. But that isn't what this post is about. This post is just to point out that automated AI research at the datacentre scale is coming earlier than most expect.