Amplify

Front page
labelling interface for math questions (for one of our team member's students)

Inspiration

We're robotics engineers by background - and the volume of training data is the key limitation in scaling robotics to GPT level performance. We wanted to build out the infrastructure to enable us to pay people to label and collect data for us - but cut out the middle man and do it as cheaply as possible through using solana as a cross border payments method. Long term, collecting task I/O data in this structured format lends itself well to training models to first augment and then replace the completion of said tasks, producing an incredibly valuable public dataset.

What it does

Stores arbitrary task specifications on chain and matches labellers to the task. As it stands, urls to data are currently on bucket, but in future they could be on Arweave/filecoin. Unlike Mechanical turk (which hides individuals behind the interface and requires you to construct your own labeller quality database) or Scale (which has "talk to sales" as it's only pricing), we are constructing a transparent view of labeller quality and enabling direct payment in SOL. Did you know - until 2019 AWS paid turkers in gift cards?

How we built it

Anchor / Rust / React, team of friends in their spare time looking to make the jump to Web3.

Challenges we ran into

The space is young and documentation is poor - something we'd love to get involved with fixing. Conceptually / technically we feel it doesn't have to be as complex to start up in Solana as it currently is.

Accomplishments that we're proud of

Building a fully on chain backend, and building a wonderful community of 7000 AI x Crypto enthusiasts over at @ContrastiveNFT (the other thing which took up our time over the last month!). Both ammplify and contrastive's reality stone (generate your own) are live on devnet!