Initial thoughts about OpenAI's o1-preview model, f/k/a 'Strawberry'

Its capabilities are impressive, but its rate limitations mean I haven't put it through its paces

OpenAI, as you have by now no doubt heard, released its new ‘Strawberry’ model, dubbed o1-preview. Queries run through it are apparently very computationally intensive, and they’re severely rate limiting users. Which means, of course, that I managed to blow through my allotment in less than 24 hours. Therefore, what follows are some initial thoughts. I have not yet put Strawberry through its paces. Much of what follows is therefore subject to change.

Consider this prompt: come up with an advanced math problem and provide the solution.

The image below is the problem that o1-preview came up with.

I don’t know what any of this means, though I’ll hazard a guess and say that it has something to do with Riemannian geometry. (Do not infer from this statement that I have any clue what Riemannian geometry is!) The point here is that o1-preview is capable of fairly advanced mathematical reasoning. If you want to review its output, you can do so here.

Alex Tabarrok thinks that o1-preview will have the most impact in biology:

Biological pathways are among the most complex in all of science. People are good at handling two or maybe three variable problems but just keeping three variables and their interactions in one’s head is difficult. AIs with access to vast databases of genes, proteins, networks and so forth will enable new simulations and learning as has already happened with protein folding.

He shares this video:

This feels intuitively correct to me, though, as with the math above, I am also not a biologist. Nonetheless, I suspect that over the coming months we will start to see a variety of researchers across various disciplines hold forth about how o1-preview is, or is not, advancing their research. (The froniter is jagged.) Terence Tao has written about how older models have helped with his research.

I also suspect that o1-preview and its successors will help with more prosaic and quotidian projects, such as developing marketing plans, analyzing legal contracts, brainstorming web copy, doing financial analysis, helping students learn algebra (or Riemannian geomtry), etc. The list of things that this model will be able to help people with is very nearly infinite.

Once I am allowed to use o1-preview again (I think on September 19th, I will post more thoughts.