you don't need mixture of agents, you can just replace it with internal dialogues with itself inside the same prompt, in fact you can fit most context of the searching algorithm and will still work, that's what i did with my search script
No code available. I've been working on my own KG, already have coding compile loop and agents. Now I'm looking for open source KG approach hopefully with Monte Carlo
this feels like the right steps forward. only thing it's missing is a search over a hypergraph or knowledge graph for injection of real time / domain specific knowledge. i hadn't yet heard about this. thanks for the video!
@sgttomas I think I found an open source project that @1littlecoder mentioned. Monte Carlo tree search for Math Olympiad. Plus one I had on my list: Domain-Specific Retrieval-Augmented Generation Using Vector Stores, Knowledge Graphs, and Tensor Factorization
Not sure if this is particularly helpful. The main objective at least on the surface seems to be to improve a much smaller model. What's the point of running a smaller model and then connecting to an API to improve? Why not just use the bigger model in the first place? I just don't get it.
![⭐️( ลอยกระทง 2024 ) Ver. แดนซ์ Remix BY [ ดีเจกิต รีมิกซ์ ]](http://i.ytimg.com/vi/rMWOLhgXUHc/mqdefault.jpg)
you don't need mixture of agents, you can just replace it with internal dialogues with itself inside the same prompt, in fact you can fit most context of the searching algorithm and will still work, that's what i did with my search script
I tend to agree. Thinking longer (chain of whatever in one go) not role playing/multiturn works for me. Context limits are a thing tho.
No code available. I've been working on my own KG, already have coding compile loop and agents. Now I'm looking for open source KG approach hopefully with Monte Carlo
this feels like the right steps forward. only thing it's missing is a search over a hypergraph or knowledge graph for injection of real time / domain specific knowledge.
i hadn't yet heard about this. thanks for the video!
I believe the Monte Carlo search uses a KG... Checking now
@@KevinKreger would love to hear what you find!
@sgttomas I think I found an open source project that @1littlecoder mentioned. Monte Carlo tree search for Math Olympiad. Plus one I had on my list: Domain-Specific Retrieval-Augmented Generation Using Vector Stores, Knowledge Graphs, and Tensor Factorization
Not sure if this is particularly helpful. The main objective at least on the surface seems to be to improve a much smaller model. What's the point of running a smaller model and then connecting to an API to improve? Why not just use the bigger model in the first place? I just don't get it.
Some of us have to use small open source models in our work. It's very relevant.
price, privacy, on device, speed, customization, intellectual property
@@KevinKreger And what exactly is open source about the API?
@@sgttomas Isn't this API in the cloud?
@@marcfruchtman9473 ah, yeah good point. eliminates most of those points made :\
The ARC paper you mentioned is doing test time tuning BTW. I thought it would have a KG to reason through the DSL tho.
They should add it to an inference SPACE on huggingface.
interesting approach!
Great stuff
❤