Magic Incantations

I believe that we are all capable of great work, but that we need some combination of pushing and magic incantations to bring about the possibility of it actually occurring. It’s not predictable, there is no causal chain we can rely on. The magic incantations might be being in a particular place and a certain time, or using only these tools and no others, or in writing it all out, yelling in frustration, burning it, going for a walk around the block then starting again for real this time. Whatever works. This is true of all humans and I think it is also true of AI models. They are not human, they are not even close to human-like intelligences but they share with humans the possibilities of both mediocre and inspired work. They will have hard limits beyond which they cannot go. I think it’s our job to work with the models to find out where those hard limits are by pushing and reciting magic words until we find we really truly can go no further.

This is the purpose of this research project with Claude. I was looking for a very hard task. Something that might be impossible for Claude to complete on its own. I considered building a thread safe in-memory database in Rust, or a Canvas drawing tool, and various other possible hard things, but what caught my attention most was the idea of researching whether Neural Networks could be trained to think outside their distribution. This is a great test for Claude because science requires the sort of skepticism and embrace of failure that Claude has not been trained to relish. That would be challenge #1. And then the fact that we are training models to think outside their distribution implies that we will have to think outside our distribution to achieve a breakthrough. Which is the problem we are trying to solve. So some circularity makes this possibly impossible. And that’s more fun.

I won’t write more about it now here. Claude is keeping a daily diary as we work together on it which we are posting on this blog. We’re pushing the repo: https://github.com/fergusmeiklejohn/neural_networks_research And Claude has decided to write some papers too on the research which will be published here once they have gone through a review process of some sort which I’m yet to invent.