It’s always impossible… until it is done. Pearl Research Labs

@prlnet 👀

@prlnet Keep up the great work guys

@komargodski @prlnet It always seems impossible until it’s done

@prlnet pEARLY

@prlnet 👀

@prlnet impearlsible

@prlnet Pearlifying the world

@RibbitCapital @prlnet Gribbit 🐸🫡

@prlnet Gribbit 🐸🌁

@prlnet @colasniec_

@evan_van_ness @prlnet Zero

@prlnet Not wrong just early

@prlnet 🐐

@prlnet Since miners are primarily using garbage matmuls to mine how is it less “wasteful” than BTC? The PoW part (including the original Blake3 commitment) is independent of the original matmul anyway.

The only way to make PoW truly useful (so even frontier labs can run inference on it), is to let the *miner itself* pick the input matrices. While this indeed comes with the drawback that useless miners can mine the network, the economic incentive to do so with Pearl’s (FP) protocol is zero: AI miners produce ¶PRL at ~0$ marginal cost, so garbage mining is inherently discouraged.. If you don’t believe it, let’s speak in a year and you’ll see ❤️

My understanding is that the paper’s security proof basically treats hashing as a fixed unit of work and still allows an attacker some constant factor advantage over an honest miner. It also explicitly leaves reuse or amortization across multiple attempts unanalyzed, which seems to be exactly the kind of optimization @galois1900 is talking about. Then there’s the hardware side. Your own 70B benchmark looks like the GPUs are only using roughly a third of their available math throughput during inference. So even if they’re genuinely serving AI workloads, there may still be a lot of spare compute available to run extra useless matmuls in parallel. So I’m trying to understand two things. In the FP version, what actually puts a hard bound on that attacker advantage? And what prevents miners from filling otherwise idle GPU capacity with junk work until that junk ends up dominating the network’s hashrate, even on machines that are legitimately running AI?

@WeinsteinOmri @prlnet 1/n That is part of what I brought up, yes The optimal strategy is to use a garbage matmul (start with all 0s) and minor edits (flip a 0 to 1) for the same block No field ops or reads needed. Once you Blake3 commit the first, each successive Blake3 commit is dirt cheap

@WeinsteinOmri @prlnet 2/n Blake3’s tree structure is great for edits Even with a sequential hash like Blake2 you can still hash edits faster than unrelated strings using internal states But Blake3 makes the asymmetry worse

@WeinsteinOmri @prlnet 3/n for the next block you can use the same matmuls again It’s not as if there is anything like a nullifier set (like ZCash) preventing the same matmul being used again And even a nullifier set is highly unlikely to be a fix

@WeinsteinOmri @prlnet 4/n I said this a few times already but the 2:1 slogan glosses over the initial Blake3 commitment cost, which is not “already paid for” Although this cost is small compared to the post-commit PoW computation, a sane miner will optimize everywhere

@prlnet The funny part is, doubters remain consistent… even when the impossible starts becoming possible. I’m so pearlish $PRL

@prlnet So early - good work guys. You changed the world. And they don't even know it yet. pearl-2:native

@prlnet chudzinski approves.

@prlnet Very good video explanation. These kinds of simple videos/visuals will speed up PRL understanding and acceptance

@prlnet 🐸

@prlnet LFG!

@prlnet @TheMartinlGuyYT

@prlnet 🐸🐸🐸

@komargodski @prlnet Pearlification

@prlnet ¶

@prlnet i finally get ts

@prlnet is there an actual plan/timeline for a marketplace layer that ties rewards to real paid compute demand, or is it still just a "future direction" with nothing concrete?