Mesa-LLM developpement (Colin's version)

[colinfrisch]

Hello everyone !

We can discuss here about the Mesa-LLM project that I'll be developing with @sanika-n under the mentorship of @wang-boyu and @jackiekazil.

My proposal, some more detailed code examples for the project as well as some refining that I did on the subject are available in my public abm-mesa-experimentations repository.

• Proposal
• Proposal refining (post-submission)
• Code examples

The general idea to it is to build a derived LLMAgent(Agent) class, with three main attributes : self.memory, self.llm and self.strategy, each depending on external classes to make it as modular as possible. I also chose to separate paradigm (basically the behavior of the agent, meaning the way that it observes and acts) and the reasoning (the way the agent plans its actions with the LLM). All this is explained in more details in the proposal and in the refinement.

These ideas are not necessarily the exact way that Mesa-LLM is going to be developed : @sanika-n's proposal tackles this project in ways that are sometimes different but also very interesting, which is why we want to discuss it to get the better of both worlds, and also take into account other ideas coming from everyone !

If you have a new topic of discussion on Mesa-LLM that you would want to explore, please feel free to create a new discussion in the Mesa-LLM repository discussions. And if you want to be involved in this project, you can join our Matrix chat room in which questions, issues, and ideas can be (informally) discussed.

[colinfrisch]

Hey, after a few discussions with @wang-boyu, @jackiekazil and @sanika-n, we are starting to have a solid common roadmap, topics and other insights to work on.
Would it be possible to have a complete feedback on this proposal from the maintainers (@EwoutH, @tpike3, @quaquel) to dig deeper into what works with you and try to improve the rest ?

[colinfrisch]

I could schedule a call tomorrow sometime between 9:30 and 14:30 CEST.

That would be great ! 9:30 CEST is okay with me.

[EwoutH]

@tpike3 would you like to join, and if so what time would work for you?

Colin and Sanika, you're presenting your roadmap (together) tomorrow? If you have any specific points or questions you want to discuss please send them beforehand.

[EwoutH]

Okay let’s do 9:30 CEST.

[EwoutH]

@sanika-n @colinfrisch I'm very sorry, my morning schedule got overthrown. Do you have time for a meeting between now and 14:30?

Edit: As discussed on Matrix, we meet 12:30 CEST.

[tpike3]

Agggh, I am sorry I missed this?? How is is going?

[tpike3]

Posting in both discussions and this may be too late... but is there any thought to 1 project integrating LLMs into models and another building a custom Mesa LLM (available via the docs?) that addresses the Mesa 2 issue? Or maybe even an agent AI that integrates the chatbot with tests so at some interval it checks and improves its responses or users have the option to have it run tests when they want more detailed responses?

[colinfrisch]

Both @sanika-n and I have started working simultaneously on different parts of mesa-llm (link to the repo here). But because we both code on it, the project is moving fast and, even if I don't want to get ahead of myself, we might finish it early : if there are other interesting things to do with LLMs in mesa, we would surely be very interested to work on them after this !

We're doing a mesa-llm meetup every Tuesdays and talk on the element mesa-llm chat (which I see you joined). If you have time, you could join the next one and share your ideas !

You can also check out this drive folder where we took a few notes from our previous meetings. If you need editing access, just send me your mail and I'll add you