Feature Request: Validators Share LLM Models for Builder AI Agents

[phoenixit99]

Description:
Current UOMI Node AI validator hardware requirements (GPU, 64GB+ RAM) are high.

I propose allowing validators to share their LLM models as a public provider for builder AI agents. This would:

• Optimize Resources: Validators with powerful hardware can offer inference.
• Increase Decentralization: Distribute AI processing across the network.
• Benefit All: New revenue for validators, lower barrier for builders.

This enhances the UOMI ecosystem by fostering resource sharing and collaboration.

[TheGhoul21]

Thanks for the proposal — it’s an interesting idea.
Before we can evaluate it properly, could you clarify a few points? For example:

• When you say “validators share their LLM models as a public provider,” do you mean exposing an inference API to other nodes/builders, or actually distributing the model weights?
• Would these providers handle only inference requests, or also participate in verification?
• How do you envision trust and performance being managed between validators and these shared providers?

For our current architecture, there are a few important constraints:

1. Output logits are required – The OPOC verification process depends on full output logits from inference, not just the final text. Any provider must expose these.
2. Verification capability – A provider must also be able to perform verification given the logits of an execution to check.
3. Direct validator–provider link – Validators would need to connect directly to the provider. If a provider misbehaves (intentionally or not), all validators using it could be slashed, so we need to avoid indirect trust dependencies.

Once we better understand your intended setup, we can explore whether it fits within these requirements or if we’d need adjustments to the protocol.

Please read these materials for a better understanding:

Consensus
Deterministic Indeterminism