How Blockchain Could Upend Enterprise Development of AI Capabilities


Picture this: an enterprise operates an image classifier to help a machine identify different types of vegetables, but it is having a hard time distinguishing between Anaheim and poblano peppers. To address the problem, the company starts a competition on the Ethereum blockchain, and invites AI experts from around the world to design a pepper detector for a nominal fee.

Once the competition ends, the company gets some new AI capabilities, and the maker of the best-designed pepper spotter gets paid. It’s a future that could be coming to your business in the not-too-distant future, thanks to some recent developments.

Meet DanKu

Algorithmia, a Seattle-based AI startup, released a smart contract earlier this year that allows people to solicit custom machine learning algorithms using the Ethereum blockchain. The “DanKu” contracts set out the task, required minimum accuracy for the algorithms aiming to solve it, a set of training data for creating the algorithms and private test data for ensuring that the resulting algorithms produce accurate results. What makes this process different is that the contracts also verify the test results of the algorithms using the compute capacity of the Ethereum blockchain, so submitters and solicitors can be assured the systems work as advertised.

When the contest ends, the contract automatically pays out a pre-set reward in the Ether cryptocurrency for the most accurate algorithm that passed the minimum accuracy threshold.

As is the norm for blockchain applications, the structure of the contract means neither party needs to trust the other, since the code automatically handles compliance. When Algorithmia recently finished its first competition (which asked for predictions of voting patterns during the last election), all the company got in return was the Ethereum network address of the winner. That’s both good and bad for enterprise applications. Companies won’t necessarily know where their code is coming from, but they will be guaranteed to get an algorithm that passes their test and the person creating it won’t have to worry about if they will or won’t get paid.

This system is meant to provide a lightweight method for companies to tap into the capabilities independent AI practitioners have to offer without engaging a crowdsourcing platform or hiring expensive employees to solve small tasks. Algorithmia CEO Diego Oppenheimer told me that some AI experts are interested in remaining at least somewhat independent; putting out an open call for their expertise through the Ethereum blockchain could help enterprises tap a larger talent pool that would not ordinarily be available to them.

This competes, of course, with existing crowdsourcing platforms like Wipro-owned Topcoder, which offers a wide variety of services that let companies tap into a community of experienced technical talent. The big difference between the two is that there’s no intermediary with a DanKu contract.

Blockchain blockers

While solving enterprise problems with grassroots methods is exciting, these contracts aren’t necessarily ready for enterprise prime time. For one, the most accurate model submitted to the competition required more computing power than the network had available. Oppenheimer expects that problem will be solved in time, since other companies are relying on Ethereum to provide compute power for their applications as well. If the network’s compute capacity doesn’t scale up, that’s a problem for the whole application ecosystem built on top of the cryptocurrency.

What’s more, while trustless transactions are commonplace for cryptocurrency fans, there’s a big difference between buying a box of chocolate or digital asset and acquiring new code for a critical enterprise system. Companies will need the ability to verify the code that they’re given and be comfortable with the nature of the transaction before this can work. Training data for the contest is stored with the DanKu contract and publicly available, so it won’t be a good fit for personally identifiable data that cannot be safely or effectively anonymized and other information businesses want to keep private.

There’s also the matter of popularity – the contracts are still new to the market, and companies might not find the right AI experts to help them with their needs. But in the event nobody submits to a DanKu contract, or the submissions don’t meet a company’s accuracy threshold, all the ether is returned to the contract owner at the conclusion of the contest.

Ether’s price volatility poses another risk. The creator of a DanKu contract places the reward in escrow for the duration of the contest (so AI experts can ensure they’ll get paid), which means the currency could end up worth more or less than a company expected to initially pay out.

Finally, it’s worth considering that this isn’t a tool for AI novices. Conducting a competition requires a clear grasp of the problem at hand, a strong definition of what constitutes success, the ability to describe all of that in a smart contract and the ability to deploy and operate the resulting model. (Algorithmia builds software that lets companies deploy and operate AI models, which is why the company released these contracts to the world.)

In other words, using DanKu contracts can give companies a boost to their existing capabilities, but it isn’t a way to figure out how to integrate AI into a business or launch new AI capabilities.

Autonomous opportunities

In the not-too-distant future, it’s possible we’ll see these contracts used as a method of automated improvement for AI systems. For example, a machine learning-based system could notice that it’s underperforming on a problem and solicit a solution through a DanKu contract (or another similar mechanism). That contract could get answered by other autonomous systems watching for challenges that meet their skillsets.

But that sort of autonomous self improvement is still a pipe dream at this point. Still, it’s interesting to consider how the decentralized power of the blockchain may just solve the machine learning talent shortage in the near term.