Professor, CS at University of California, Berkeley. Deep learning, security, blockchain expert. Founder of Oasis Labs.
Research Scientist at Stanford. Program chair for PAPIs machine learning conference. Machine learning and quant finance at major banks, hedge funds.
Chairman of Danhua Capital.
Professor, Physics at Stanford & professor of IAS at Tsinghua.
Partner, DraperDragon Fund. Investor in blockchain and AI. Former CEO of QunZhong E-Commerce & founder of OLEA Network.
Program Director at Singularity University, faculty in Law and Blockchain. Principal at Crypto Lotus. Ex-Gunderson Dettmer. JD/MBA, Yale.
General Counsel at Crypto Lotus, a cryptocurrency hedge fund. Advisory practice on blockchain companies and funds. U. Chicago Law.
CEO, Ambisafe. Founder of Orderbook. Early adopter of blockchain tech. Launch of 20+ successful ICOs. Decentralized systems since 2008.
CTO, DigitalX. Expert on blockchain, cryptography and machine learning. Math and physics, UF.
The HUMAN Protocol is a broadly applicable approach to organizing, evaluating, and compensating human labor. It enables a new generation of machine intelligence to apply human labor to AI model self-improvement in order to achieve human parity in task performance.
Today this work is commissioned by machine learning practitioners. The protocol's immediate application is thus to improve the most labor intensive problems in machine learning: making datasets fit for training via annotation and validating model inference quality.
While the HUMAN Protocol supports and improves today's practices, it is engineered for the next evolution of human inputs to machine intelligence: letting machines ask people directly for the data they need to improve.
HUMAN Tokens ("HMTs") serve as the medium of exchange in the HUMAN Protocol. They are EIP20-compatible tokens, and the complete system forms a decentralized platform with an open protocol. Each component receives a fee for its role, and interactions are coordinated via smart bounties on the Ethereum blockchain.
Initial protocol design work & prototyping
Production-quality implementation begins
Private beta of first app: hCaptcha.com
Human Token contract live on testnet
hCaptcha public beta
Initial open source code releases
Open source reference implementation
"Proof of balance" live in production network
The recent success of deep models has led to use of increasingly large datasets. Creating these datasets via Mechanical Turk, etc is slow and expensive. Today both the defining of requirements and the labeling work are done by people.
More actors must be allowed to participate. Friction must also be reduced within the market to enable the next generation of systems for continuous improvement via human review.
A substantial portion of dataset value is today captured by Google at very low cost via reCAPTCHA. Creating economic incentives for website owners by providing a drop-in replacement for reCAPTCHA will democratize access to high volume human evaluation.
This system will test for bots at least as well as reCAPTCHA while at the same time paying website owners for their audience.
HUMAN Exchanges maintain an order book of job requests, matching labor with demand
One of the more interesting areas of research today is in factored cognition: decomposing more complicated work into its simplest cognitive components. Practical applications of this idea map very nicely onto the HUMAN Protocol.
The protocol defines standard job types that serve as building blocks for many tasks, and anyone can publish a new job type.
For example, an Exchange can offer a high level job type ("scan a page") and then factor it into smaller tasks ("type in the letters or numbers") that can be sent out to other Exchanges based on available capacity and current order book price.
Anyone can run an Exchange and publish a new job type, adding value on top of labor pools across the world.
Combining our new token-level Bulk API with the HUMAN Protocol smart bounty lifecycle design allows the protocol to scale into billions of tasks and users per day on the current Ethereum mainnet.
How does it work? The Bulk API extends the standard EIP20 token interface to enable efficient one-to-many payments, allowing ~1000x more efficient micropayments via message packing.
We plan to submit it as a formal standard for the entire Ethereum community to benefit from our work, and believe it is a novel and additive approach to scaling real-world applications.
In Proof of Stake each participant needs to buy tokens, which in turn increases cost and friction for new participants. We have designed a novel mechanism to address this problem for systems where every added participant increases network value.
Proof of Balance benefits more invested participants and increases attack resilience without discouraging new entrants from participation in the network. (Please see the technical whitepaper for more details.)
Whether you're a developer, partner, or just interested, we'd like to chat.
Requesters of work launch new bounties onto the blockchain that specify a job: the question to ask and the set of tasks to ask it about.
Exchanges pick up jobs, manage bidding on job types, and serve tasks to agents doing the work.
Recording Oracles collect potential answers and provide a rolling evaluation of answer quality.
Reputation Oracles make a final evaluation of answer quality and reputation score per job, and finally pay out bounties.
Advantages today: allows “open books” to prove the system is fairly distributing bounties, enables efficient micro-payments, reduces required trust between protocol actors.
Even more advantages when blockchains are faster: verifiable reputation for every actor that opts in, oracle can compute earnings on-chain to further reduce required trust for interactions.
Ethereum has perhaps the most robust smart contract support of any popular blockchain, but is currently too slow and expensive for many applications without additional development.
Our Human Token contract thus implements a custom Bulk API that supports efficient micropayments via one-to-many bulk transfers. This enables new and interesting use cases while remaining EIP20-compatible.
We are open sourcing the audited contract with a library and comprehensive test suite to help other projects in the wider Ethereum developer community adopt this approach if it suits their needs as well.
Current systems: no blockchain, centralized authority, REST API. No compensation for use.
Challenge: no blockchain today has adequate performance for use as a full-scale distributed human review system. Plasma, Lightning, Hashgraph, etc are still orders of magnitude away from necessary cost/speed performance and not yet robust. Future improvements may eventually make this feasible, but still early days.
Hybrid model thus ideal: faster to build, easier to scale using robust, proven strategies. Blockchain is used primarily for settlement, rather than trying to put every bit of logic into an on-chain oracle and every bit of data directly on-chain.
Start more centralized, cheap and deterministic for performance.
Build in ability to decentralize each component as tech evolves.
What do we want on the blockchain?
What do we NOT want on the blockchain?