What if we built something like Folding@home, but for LLMs?
retained text
I've been thinking about this for a while, and I'm curious what other people think. We have billions of processors around the world sitting idle for a good portion of the day. Gaming PCs, servers, laptops, phones, tablets, and even smart appliances. All that processing power and memory just sitting there doing practically nothing. Why couldn't we build something similar to Folding@home, except instead of protein folding, we're collectively running and potentially training massive LLMs? I'm talking about a completely decentralized, peer-to-peer network. No central servers, no subscriptions, no cryptocurrency, no tokens, no corporate ownership. You install an application, dedicate whatever spare resources you're comfortable contributing, and in return you get access to the collective AI. The idea would be to distribute the model and its computational workload across participating devices. One computer might contribute a few gigabytes of RAM, another might have a powerful GPU, and another might just have some CPU cycles to spare. Not every device would need to do the same kind of work. And I'm not particularly concerned about speed. I'm not trying to get 200 tokens per second. If it takes a little longer to generate a response, who cares? The point is making enormous models accessible without requiring everyone to own thousands of dollars worth of hardware or depend on a handful of corporations. Anonymity would have to be a fundamental part of the architecture. No accounts, no IP logging, no centralized conversation histories. Just an open network of people voluntarily contributing resources and using the resulting compute. Obviously, making that genuinely private would be a technical challenge in itself. Think of it as one gigantic, voluntarily operated botnet dedicated to AI. I know projects like Petals have explored distributed inference, so the underlying concept isn't entirely new. I also understand that network latency, synchronization, malicious nodes, and verifying computations would present some pretty significant challenges. But are those actually insurmountable problems, or are they engineering problems that haven't received enough attention? We already have distributed computing projects demonstrating that millions of unrelated computers can cooperate on incredibly complex scientific workloads. I realize LLMs aren't quite the same thing, particularly when it comes to sequential inference and training, but it makes me wonder how much of our existing hardware we're leaving on the table. And imagine if the network eventually became large enough to collectively train and run models that no individual participant could possibly afford to operate. An open-source AI that isn't controlled by any single company, government, or organization, and that continues to function as long as people are willing to participate. I'm curious whether people working with distributed systems or LLM architectures see a fundamental technical limitation here that I'm overlooking. Is something like this realistically achievable at internet scale? And if so, what's actually stopping us from building it?
document fetched 2026-11-03 00:28 · canonical https://reddit.com/r/OpenAI/comments/1xm8zmn/what_if_we_built_something_like_foldinghome_but