Musing 111

The big AI news of the week is that an OpenAI model solved a famous geometry problem - the planar unit distance problem.

What’s interesting is, in reading the proof, I’m reminded of my view of LLMs - they are fantastic at interpolating existing human knowledge and filling in the gaps between known points.

I’ve read a number of math papers in my life, and this was one of the most citation dense papers I’ve read in quite a while. And combined with the fact that there was no novel technique (as in, completely new, from scratch, with no clear source for reference in its function or description), and I feel very vindicated right now.

But it’s interesting to see how the accelerationists respond. They see it as “AGI is here baby!!!”, which, by some metrics we may very well be. Giving the sum of all human knowledge, indexing it for easy reference, then stick it into a machine that can effectively look at all of it will go a long way to mimicking at worst. But that aside, this exposes how little many of the lay level accelerationists really understand the world. They see “AI solved this problem!” and seem to automatically assume that the solution was some groundbreaking, completely innovative solution that had no or minimal grounding in human knowledge.

Such a view is so shallow though, because it implicitly functions as if existing knowledge is comprehensive. What LLMs have shown is that any idea our knowledge was comprehensive and fully applied is but an illusion, and now a shattered one at that.

I, for one, look forward to what LLMs can come up with as they’re able to cross reference ideas from the massive amounts of data in a given field and help us fill out our knowledge and application.