Decentralized Autonomous SAAS with Proof of Steak
Exploring a SAAS business that runs itself using Proof of Steak.
What follows is a compilation of ideas and technology around AI DAOs and curation markets, applied to the problem of building a business that self-organizes and runs itself.
📷 Example: Let’s Put the “AI” in ClarifAI
I find that thinking of a concrete implementation while discussing abstract ideas helps quite a bit. So read on with the concept of “decentralized, autonomous Clarifai” in your head.
Clarifai is an API for image recognition. Tagging, classification, that sort of thing. We want to build a version of Clarifai that runs itself. It makes money by providing a service and uses that money to make its service better. (which isn’t a new concept, mind you; that’s what normal business do, kind of by definition. What’s interesting here is the “decentralized and autonomous” part).
📎 Token Curated Registries and Data Markets
Simon de la Rouviere has a lot of brilliant things to say on the topic, so see below for more reading. But I’ll do my best to highly summarize and highlight my favorite ideas around Token Curated Registries (TCRs).
A Token Curated Registry is a set of “things” (a registry) that is curated by the crowd. TCRs work due to economic equity; because the registry is more valuable than the sum of its parts, there’s rational, economic incentive to make it as high quality as possible.
Let’s go concrete: adChain is using a TCR to curate a list of reputable domain names. You could use a TCR to curate playlists. You could use a TCR to curate adblocker blocklists. You can use a TCR to curate images of steaks.
And you could use a TCR to curate data.
Data curation is a surprisingly great application of TCRs; well-curated, high quality datasets are worth way more than the individual data within.
🎈 The Philosophical Part of Open Data Curation
The technology isn’t 100% of what makes a good machine learning model perform well. The effectiveness of a machine learning model depends hugely on the scale and quality of the data that it’s trained on. Big companies publish tons of AI/ML tooling and research because they know the data is half the battle.
The big companies have tons of data–everyone else, not so much. If everyone joined together, the resulting dataset might match the value of the locked datalakes of FaceGoogAppSoft and friends.
A few different projects, namely Ocean, are working on a data marketplace (both paid and open datasets), but I’m particularly interested in open data marketplaces.
💸 The Economic Part of Open Data Curation
Participants that curate data in a specific market (say, labelled images) have equity in the registry now; they’re invested in its quality because they use it for training their own models. They benefit from additions to the registry and are invested in keeping it high quality.
👷 Decentralized MTurk
Now we have a registry that’s being decentrally curated, but the only economic motivations to contribute to the registry is the speculative value of its quality. That is, if you think your contribution will make the list more valuable, you purchase tokens, contribute information, and can then either sell the tokens for a profit or, if you have equity in the registry, join the curator community.
By adding a “bounty” on submissions that make it into the registry, it’s possible to motivate participants that don’t have intrinsic equity in the value of the registry to contribute; utility value. You can combine this with “decentralized MTurk” (which Gems is tackling) to collectively put bounties on tasks that contribute to the registry.
To make this concrete, say I have equity in that registry of labelled images, but I need more of them. If I believe that the value of the registry (and the service I provide based on that registry) will go up, it makes sense to pay someone to contribute to it. This can be as simple as “submit labelled images” or something more like “I have a bunch of unlabelled images, label them and submit them”. If the content is accepted into the registry, it was of high-enough quality and the expense is objectively worth it.
Running the SAAS
We’re interested in the market of labelled images (seeded with ImageNet, for example). We can use this open dataset for free, contribute our own data for equity, or pay outside participants to contribute to the registry. Let’s use it to train a model to label additional images by delegating this computation to TrueBit and then store the model on IPFS.
Now we need to provide a service. At the basic level, the Ethereum Virtual Machine is effectively a decentralized Functions-As-A-Service platform; you make a request that’s processed on demand, and billed incrementally. When there are no requests, the code doesn’t run. Ethereum is a bit different from the traditional understanding of FAAS, but we can make it more approachable by exposing it as a traditional HTTP API with a gateway.
So, as a minimum viable service, we can write a smart contract that accepts requests to label an image, delegates the actual processing to TrueBit, and then returns the resulting tags, keeping the difference as revenue.
Making a trustless/decentralized HTTP gateway for Ethereum is an open question (TrueBit could probably verify this!). Until then, we can provide an HTTP gateway binary that anyone can set up and run on their own infrastructure.
Finally we can take that revenue from providing a service and invest it into a bounty that increases the scale and quality of the registry that we draw value from.
DAOs, Governance, and zeppelin_os
As much as we’d like to believe that the code that governs this economic cycle will be perfect, it will almost definitely need to be upgraded, either as bug fixes or to implement a better model.
To make this possible, we can throw in some DAO governance inspired by how zeppelin_os will eventually be governed, which itself is similar to a TCR; participants vouch for the different versions of code that are running.
While zeppelin_os doesn’t provide equity in the sense that token holders receive dividends, I expect this would be a logical step for a self-sufficient SAAS business; the contributors that put work into making the service more efficient or valuable receive equity for their contributions and receive dividends on its performance.
You can read more about how zeppelin_os handles code upgrades in their recently-updated whitepaper.
Proof of Steak
Proof of Steak is the protocol responsible for creating a TCR and incentivizing its use. It combines the TrueBit Classic protocol with PLCR voting to create an open curated registry. It includes traditional transactional economics for motivating participants that don’t have equity in the quality of the registry.
And it includes a joke implementation called the Steak Network.
Further Reading
Continuous inspiration from Simon de la Rouviere and Trent McConaghy.
