Hacker News Reader

Comment by randomdrake | Hacker News Reader

I’ve been toying with the idea of a SETI@Home successor that uses leftover subscription credits from the (currently) subsidized AI providers to help contribute to a collective goal like this.

I wonder if there will be any opportunities to combine resources collectively like that again in the future.

I built a small prototype that would let people in less-developed countries get AI responses from it for free. But it’d be super neat to work together and put our laptops and subscriptions to work on something like this.

View on HN

Sounds like a great way to get your account suspended… My Claude account was once mysteriously suspended (I wasn’t doing any reselling/sharing and not asking anything sensitive) then mysteriously restored three days later, so I’d say actually violating ToS is likely more risky, given that they seem to not mind suspending false positives at all, and you have zero recourse once you’re suspended. Their customer support is an AI chat bot and even that chat bot page just redirects to the suspended page lol.

You’re right, of course. I never made it out of PoC. But I also implemented something with locally running models to do the same thing. Request comes in, gets forwarded to a machine not currently busy, it serves.

It would be great if we could use the credits we pay for and don’t use on our subsidized subscriptions as well. But, yes, a matter of “when” not “if” for your account getting shut down.

After a bit of research, it looks like BOINC[0] and AI Horde[1] are along the lines of what I was building. It appears NVIDIA PAIR[2] has been working on distributed inference for local networks as well.

[0] - https://boinc.berkeley.edu/

[1] - https://aihorde.net/

[2] - https://www.nvidia.com/en-us/ai-on-rtx/personal-ai-router/

I feel like it would be better to create open synthetic training data for all to benefit from,or something along those lines, as opposed to giving it to those claiming to be in need, as such a system would be exploited and abused in a matter of days.

There used to be folding@home for exactly this; pre-Alphafold days, where you could run protein folding on your machine.

Yes! I used to love watching the visualizations for this. You’ve got the right idea. If it truly is AI processing to help us eventually create this virtual biology, why not our idle machines with models running on them?

To avoid suspension I was thinking of something similar driven by Github issues, pr and a skill people would load and run.

They’ll easily detect that. Remember when they routed subscription users’ sessions to extra billing when a certain word appears in the context? https://github.com/anthropics/claude-code/issues/53262 https://news.ycombinator.com/item?id=47952722

Claude Code is known to snoop on other aspects of its operating environment too, not just the texts sent to them: https://news.ycombinator.com/item?id=48734373

Reply on news.ycombinator.com