Local LLM computation
becomes mining work

First, what is a local LLM?

A local LLM is an AI model running on your own computer. Your PC performs the computation that summarizes text, answers questions, or writes code. MISAKA Studio provides an entry point for finding, downloading, managing, and running models locally.

MISAKA treats eligible model computation as a contribution to the network. Mining uses a model that meets the network’s rules, a specified computation, and evidence that lets others check its execution. Downloading a model, using it locally, and submitting mining work each have their own steps.

Five steps from computation to rewards

1. 01

   Choose a model eligible for mining

   Select an admitted model that qualifies for network work. Models with the same name may have different weights or versions, so the exact model must be identified.

   A job specifies which model to use, what to compute, and the limits of that computation.

2. 02

   Run the specified computation on your PC

   Execute the job with the selected model. For a summarization job, the model processes the specified text and generates a response. The execution also produces information needed for verification.

   A compatible execution environment packages the result and required evidence.

3. 03

   Submit the result and evidence

   Submit the response together with evidence binding it to the model, input, and computation. The submission carries your signature and fixes the result and evidence so they cannot be replaced later.

   The design supports operating your own node or delivering an authorized submission through a node or relay.

4. 04

   Assigned verifiers check the computation

   The network assigns participants to check the submitted work. They use the evidence to establish whether execution followed the rules. For a large model, verification is divided into scopes assigned to different participants.

   Required evidence must also be available. Disagreements are narrowed down and adjudicated.

5. 05

   Confirmed work connects to MSK rewards

   After verification and the required challenge period, work settles on the ledger. Rewards follow the protocol’s allocation and payment conditions. Submitting the same computation repeatedly cannot earn duplicate rewards.

   Submitting work, settling it, and releasing the reward are separate stages that proceed in order.

PALW is the mechanism that checks this work

PALW provides the rules for recognizing AI computation as network work. The essential requirement is knowing which model performed which computation and letting other participants check that execution. Compute providers, verifiers, and ledger participants follow common rules.

Verification checks whether the specified model computation executed correctly. Whether a response is useful or safe, and which model performs better, are questions for model evaluation. Execution verification and response evaluation each have their own role.

How can a large model be checked?

Having every participant rerun every large inference would require substantial resources from verifiers. MISAKA’s design checks computational relationships using evidence tied to the result and divides verification into assigned scopes.

For example, one verifier can check part of the earlier layers and another part of the later layers. Checks also bind the intermediate values together into one execution. If results disagree, the dispute narrows to a smaller scope for adjudication under the rules.

There is more than one way to participate

Misaka Network records and settles the work

Find models, run them, and follow the network