I don't see this creating new algorithms (as in, not in the training corpus), but maybe giving the kind of answer you would expect from Stack Overflow, without all the social fluff around it (comments, badges and so on).
The day one of these find new algorithms to solve problems with better complexity or simpler code that state of the art, I'll wake up. When I give a LLM a computational geometry problem, it's exactly like a student trying to bullshit his/her way through an exam without any actual deep understanding.
For example, I ask for an algorithm to compute Laguerre Voronoi diagrams (usually not available in books or code examples), and I get answers for plain Voronoi diagrams, because it's what you will find in many books and code samples. Generating boring but necessary code, in moderation, is a win.
A model for algorithms would not be trained on code but mathematics with some limited psuedocode.
IIRC there are some people looking at theorem proving with LLMs, but the reality is they don't have to do anything foundational or groundbreaking to be of value in assisting, supplanting, or replacing the vast majority of people who interact with computer code.
We are talking about the field that openly just memorized leetcode questions to "make it", right?
Why should we have such high standards for tools when we don't even apply them to the "best and brightest"
The day one of these find new algorithms to solve problems with better complexity or simpler code that state of the art, I'll wake up. When I give a LLM a computational geometry problem, it's exactly like a student trying to bullshit his/her way through an exam without any actual deep understanding.
For example, I ask for an algorithm to compute Laguerre Voronoi diagrams (usually not available in books or code examples), and I get answers for plain Voronoi diagrams, because it's what you will find in many books and code samples. Generating boring but necessary code, in moderation, is a win.