Years ago when gpu power was basically free through google colab I used that to my advantage by racking up sheepit credits rendering other people's projects on google's dime. I still have about 10 million credits right now and I don't think I'll be able to use them all.
I haven’t done render farm stuff since the 3DSmax 5 era in high school. Do they let you do stuff like ask to see a frame, kind of like when someone walks up to a production line to spot check the product?
How’s the security? If I render other people’s stuff, how do I know that render code doesn’t install a keylogger, steal my auth cookies, and send spam email from my machine?
Does a system exist where we can train open source LLM models in this way? Maybe distribute it across a cluster of computers with credit points that can be later used for inference.
Not for training, but execution in a SheepIt-similar way can be done with AI Horde, which not only does LLM inference but also Stable Diffusion. https://stablehorde.net/
I like this. I'd like it even more if it'd have some sort of a credit system. When I'm rendering for others I'd accumulate credits. When I need rendering I could use those credits for priority.
I seems fair but might be tricky to implement to prevent fraud/abuse?
Then make it a feature so that the person receiving the renders can accept/reject them. If they are rejected, then the renderer is not credited. If the person receiving racks up too many rejects, then they cannot submit new renders. The reject process would be hairy in that you'd need to explain why. It has potential for getting nasty, but any 2-party system can.
Seriously though, any platform is going to have its share of liars, cheaters, grifters,
con artists, and outright scams. What the world relies on is that enough people aren't. The trick is attracting enough honest people while driving away enough of the dishonest people. it's a constant game of cat and mouse.