Enabling Distributed LLM Computing

[P Lawton]

Folks:

This is a proposal for BOINC 2.0, a new tool which would be enable Distributed LLM Computing for Public Benefit.

The rapid evolution of large language models (LLMs) has concentrated computational capability into the hands of a few organizations with access to massive GPU clusters. However, many valuable LLM tasks — including batch inference, alignment evaluation, and benchmarking across large datasets — do not require real-time performance or tightly coupled infrastructure. These tasks are ideal candidates for distributed, volunteer-based computing. BOINC has a proven track record in this domain, and is uniquely positioned to evolve into a next-generation framework (“BOINC 2.0”) that supports the decentralized operation of open-source LLMs.

Unlike traditional BOINC workloads, LLM-related tasks require modular execution, encrypted data handling, model caching, and redundancy verification. We propose the creation of a new task architecture within BOINC that allows for chunked inference jobs, federated evaluation, and optionally lightweight fine-tuning. With an updated coordination layer, model shard management, and a reputation-based result validation system, BOINC 2.0 could facilitate the open, auditable, and low-cost operation of community-driven LLMs. This would support academic research, transparency initiatives, and global participation in AI development without depending on centralized infrastructure.

We urge the BOINC development team to consider launching a dedicated initiative to build this new capability. This system would not attempt to replicate hyperscale training environments but instead carve out a distinct, highly valuable niche in public LLM utility: enabling scalable, distributed AI research at near-zero marginal cost. BOINC 2.0, tailored for distributed LLM workloads, could become a cornerstone of democratic AI infrastructure — resilient, inclusive, and globally accessible.

Of course, your team may already be well along the road toward such a tool, which would be great news. 

If not, and this idea sounds good to you, please feel free to run with it as far as you wish - and let me know, please, if this one has legs. 👍 

Thanks so much for your time and attention.

Peter Lawton

Tampa, FL