Yeah but integrating LLMs with other systems is already happening.
Most recent case is out of Deepmind, where they managed to get silver medalist score in the International Mathematics Olympiad (IMO) using a LLM with a formal verification language (LEAN) and then using synthetic data and reinforcement learning. Although I think they had to manually formalize the problem before feeding it to the algorithm, and also it took several days to solve the problems (except for one that took minutes), so there’s still a lot of space for improvement.
Sure, but you can do a lot more than that. You could combine LLMs as part of a bigger system of different kinds agents, each specializing in different things. Similarly to the way different parts of the brain focus on solving different types problems. Sort of along the lines of what this article is describing https://archive.ph/odeBU
Yeah but integrating LLMs with other systems is already happening.
Most recent case is out of Deepmind, where they managed to get silver medalist score in the International Mathematics Olympiad (IMO) using a LLM with a formal verification language (LEAN) and then using synthetic data and reinforcement learning. Although I think they had to manually formalize the problem before feeding it to the algorithm, and also it took several days to solve the problems (except for one that took minutes), so there’s still a lot of space for improvement.
Sure, but you can do a lot more than that. You could combine LLMs as part of a bigger system of different kinds agents, each specializing in different things. Similarly to the way different parts of the brain focus on solving different types problems. Sort of along the lines of what this article is describing https://archive.ph/odeBU
It’s kind of like how graphics cards are used to optimize specific repeated computations but not used for general computation
Good analogy, it’s a tool for solving a fairly narrow problem in a particular domain.