Implement P2P Network Layer

[iblameandrew]

Implement a P2P player to connect client instances in peer networks. Figure out a summarization and broadcasting strategy for the reflection and action signals.

[musoles]

@andres-ulloa-de-la-torre I’d love to collaborate on this. Some context: I am the creator of Kalavai, a platform designed to aggregate computing resources from any device (desktops, cloud VMs, raspberry pi…). I feel this project is the perfect match as we make distributed deployments easy, handling communication, reliability and workload at any scale with over-the-internet networks.

As an example of what Kalavai can do, we have built CoGen AI a community owned, distributed AI provider for genAI models.

If there’s enough community interest I’d love to support this research effort.

[iblameandrew]

@musoles

Hey pleasure to meet you, i would love to collaborate - this is pretty much set from the AI pov so i dont expect to add on top of the original algorithm and graph architecture - last days getting this to work have been rough. Right now im testing with light workloads hunting for bugs - if any changes happen to the graph structure it would be over minor fixes. The network can be partitioned MPI style. Some parameters shall be added to the GUI like IPs of the collaborators and then how many agents in the dense layer are assigned to each peer; the reduce step to prepare the reflection pass can be handled by some host node, but then it requires again distributed partitioning in the update step where system prompts are updated for each hidden agent. Very exciting stuff and unlike conventional ML algorithms, LLM inference is so slow that the network latency would be meaningless.

I have updated the p2p branch with the latest additions. I can no longer recommend Gemini or any LLM for that matter to write changes to the app because they are starting to hallucinate pretty bad.

Feel free to engage with autonomy - whatever question about the algorithm you have, im available all day GMT-6.