『Re:0』

Most engineering and science problems boil down to solving optimization problems. One of the most important optimization problems being worked on today is the task of finding optimal weights and architectures to solve a problem.

If an institution wants to find the best possible solution to an AI optimization problem, they are currently limited to their internal research team.

Academia is also plagues by problems like biased peer reviews, IP theft via AI platforms (centralized contests) and requirement of blind trust required for codeless papers. These are problems we have personally faced as researchers.

A possible solution to this is incentivized open contests accessible to researchers worldwide. However, there's a catch : solutions need to be revealed publicly and the contests are hosted on a centralized server. E.g. Kaggle, TIG, IExec RLC etc.

We created a system which

1. Does not make the weights or architecture public
2. Trustlessly proves optimal weights
3. Assures guaranteed payouts to winners

We created a platform and mechanism to crowdsource the calculation of optimal weights/architecture through a contest powered by ZKPs without having to write complex ZK circuits.

The organizer creates a contest for their problem and the payout is locked in a master smart contract. We have created a template repo for creating contests: https://github.com/levihackerman-102/proof-of-optima-template

The competitors compute their solutions and generate proofs off-chain using RISC0. These proofs are then submitted to the master contract and verified on chain using the deployed RISC0 verifiers.

During the competition, we maintain a true leaderboard of the weights/architectures ranked by accuracy at all times which ensures competition and guarantees anonymity at all times.

At the end of the contest, the winner sends their winning solution to the organizer using public key cryptography and the master contract automatically releases the funds to the winners after verifying that it's indeed the winning solution. The organizer can decrypt the solution to get access to the optimal weights.

I think using GPUs for parallelizing proof generation on RISC0 was a good move and really "optimized" our process making it upto 4x faster.