I got a pretty nice bounty from @prlnet ($PRL) last week. Also received bounties from @gonka_ai, @Conflux_Network and @VerusCoin recently. Thank you all, and sorry if i missed anyone. The reward is nice, obviously. But so is a team that doesn't decide you're their enemy for reporting a problem. Some projects get this. Others still act like the bug wouldn't exist if you hadn't told them about it. Paying people to report problems before someone exploits them should be normal. Somehow it still isn't. Anyway, spending so much time around pearl also got me interested in their next upgrade. Specifically fp8. Right now pearl mining uses its own integer format, while a lot of the ai inference you'd actually want to serve runs in fp8. the hardware can do both, but that doesn't automatically mean both can use the same computation. there's still a mismatch. @togethercompute ( Together ai ) already showed why this matters: pearl rewards brought the price of serving a model below raw compute cost. so there is something here beyond “ai + crypto” on a website. Pearl's fp8 work is public, and it's purpose is to remove that mismatch. Do the inference in the format the model already uses, and use that same work for the mining proof. The user gets their output, pearl gets its proof. This is what made me think about @crynuxio (base:0x9557dd9e241bc9636732623b672b4090af519396 ) and @gonka_ai ( couldn't find the ticker ) . Crynux is already serving real jobs on home and edge gpus. Gonka is already running fp8 models on its network. So my question is pretty straightforward: could a gpu doing inference on either of these networks also earn pearl from that same work? Not splitting the card between two jobs. Getting more out of the computation it's already doing. To be clear, these stacks haven't been connected yet. This isn't an integration announcement, and your crynux node isn't suddenly going to start earning pearl tomorrow. There's work to do before we know whether this actually makes sense. But if it does, the node gets another source of revenue, mining rewards can cover part of the inference cost, and pearl gets secured by real ai work. That would also give crynux / gonka operators another reason to keep their nodes online. That's interesting enough for me to test. I want to see what it earns, what the overhead is, and whether it's actually worth running. not just whether we can make a demo work. I'm getting a setup ready to measure it under real operating conditions. Other pearl miners who want to try the same thing, let's compare results. I'll post the numbers either way. Including if they're disappointing. ps: another blackwell just landed. getting ready for fp8 :)