Better machine learning, curing a crypto crisis
Tromero wants to take on the cloud kings by rethinking wasteful mining rigs
One type of tech we’ve not covered at all in PreSeed Now yet is anything on a blockchain.
I remain to be convinced that 99% of what I see in the crypto/web3/blockchain space has any longterm viability outside of a small but obsessive audience.
The field is full of interesting early-stage tech with little in the way of product-market fit, and often no sign of it ever being found.
But today’s startup managed to change my mind… at least a little. This one goes beyond an interesting technical challenge.
Tromero sits at the intersection of machine learning, blockchain, and concern about the environmental damage ‘Proof of Work’ mining is doing to the world.
They’re not the first to come up with the idea of ‘Proof of Useful Work’, but they’re making good progress and it’s well worth a dive into.
Can Tromero take on the A.I. cloud kings… with the help of wasteful crypto mining rigs?
It’s fair to say that outside of a small community, cryptocurrency mining has a bad reputation. All that computing work and energy consumption is harming the planet for the simple goal of making a few people potentially wealthier.
But what if we used all that computational power for something more broadly useful? Why not use all those energy-intensive mining rigs as a resource for machine learning? That’s what Tromero is working on.
“Everyone knows machine learning is expensive. Training models is a long and difficult process and it is a real pain point for a number of startups,” says co-founder Charles Higgins. “At the same time, cryptocurrency mining is chugging away in the background as the world's largest decentralised supercomputer.”
Tromero wants to launch a system where organisations that want to process machine learning work would be able to get that work conducted by crypto miners. Those miners would in the process mine Tromero’s own cryptocurrency and also get paid an amount in fiat cash.
“What this means is they'll get paid more than they would normally by mining any other cryptocurrency, because they get the external reward,” says Higgins’ fellow Tromero co-founder Sophia Kalanovska.
“But it also means that the problem providers can pay less than what they pay on AWS, because it's partially subsidised by the mining process. So by combining those two processes into one, we make it more efficient.”
Why complicate things with both a cryptocurrency and fiat currency payment for miners? Higgins says it’s essential to have tokens to mine for a system like this to work.
Kalanovska adds that they’ll be focusing on mid- to large-scale transactions to “assure better stability” of the cryptocurrency, with the external cash reward helping to attract miners even if the value of the currency did take a dip. Let’s face it - cryptocurrencies aren’t exactly well known for price stability.
But why go down the crypto and blockchain route in the first place? Couldn’t all this computing power be used without the added complication of a currency attached to it?
“It could be done. However, the savings we could offer–the value for money–would go away,” says Higgins. “We fundamentally offer value because we're solving two problems with the same process. A cryptocurrency needs to be mined, and that requires solving a problem - solving problems is machine learning, so more people benefit from the same process.
“If more people benefit from the same process, effectively the same number of compute cycles, the same amount of energy is expended. But it benefits both the people within a network and people outside of the network as well. And so because of this process becoming more valuable, we can offer it for cheaper than in a traditional way. The cryptocurrency is necessary in that respect to make it good value for money.”
“They do computation on the blockchain… they sit on top of already existing cryptocurrencies.
“With them you'll benefit from decreased prices because of marketplace competition, but you don't benefit from decreased prices from a mining perspective, because they don't mine any currency. They don't unite to process into one like we do.”
Rethinking cloud computation
It’s worth noting that this solution is still very much a work in progress (more on that below), but Tromero is working with Conception X to build their research and white paper work into a fully-fledged commercial operation.
And Higgins argues that some fresh thinking in processing machine learning–dominated by cloud computation through Amazon’s AWS, Microsoft’s Azure, and Google Cloud–is much needed.
“We can compete at a price point with AWS, Google, and Azure, meaning that we actually think we can capture a large share of marketplaces when we can fundamentally democratise and distribute machine learning training.”
The blockchain Tromero wants to build to underpin its offering will eventually be Turing complete so it can operate as a platform, like Ethereum does.
“We're building the new building blocks of a fundamental ecosystem rather than just a currency,” Higgins says. “But it is an incremental process…. the idea is it can become the first clean and truly decentralised Proof of Work chain, which goes back to the idea of a truly decentralised internet.”
But isn’t Proof of Work killing the planet?
Much has been made of Ethereum’s shift to a Proof of Stake system that removes the controversially massive energy consumption of the Proof of Work approach used by cryptocurrencies like Bitcoin.
Tromero’s approach is to make Proof of Work more useful rather than getting rid of it, so the energy used is put to better use.
And Kalanovska argues that if Tromero can move some of the machine learning work being done on the big cloud platforms onto mining rigs that would have been running anyway, they could reduce overall energy usage.
Tromero isn’t alone in the goal of making ‘Proof of Useful Work’ a success (see Flux, for example), but coming from an A.I. background, Higgins and Kalanovska come across as approaching the problem from a different angle to some others.
Making ‘Proof of Useful Work’ work
Higgins and Kalanovska are both final year PhD students focused on A.I. at King’s College London. They identified the environmental impact of blockchain technology as a problem worth solving, and their expertise in A.I. and frustration that they couldn’t train machine learning models faster, provided the perfect inspiration.
The pair began work on the project a year ago, and incorporated as a startup two months later.
Their first take on the idea was harnessing cryptocurrency mining to hack passwords, but after realising the customers of such a business would have “questionable morality” (to say the least!), they decided to switch to the much more generally wholesome world of machine learning.
Having written a white paper on this approach, they’re focused on building the tech with a launch date target in two years from now. So far, Higgins says they have developed a demo that shows their ‘proof of useful work’ concept works.
“If this were just a simple blockchain, we would be able to launch it far faster and probably without much investment. We're at a stage where we have a blockchain that works.
“The stage that we need to go further is we need to build a container where people's data are protected, which is how we can incorporate machine learning and data into our blockchain.”
Kalanovska adds that the other important part to be built is the infrastructure that means data can be transferred securely, too.
Investment, vision, and challenges
Essential startup information for paying PreSeed Now members:
Keep reading with a 7-day free trial