William Tunstall-Pedoe, who sold his voice-activated search startup Evi to Amazon in 2012 and paved the way for Alexa, is no stranger to building AI-powered tech products that can change the world. But, with his latest venture, there are no plans on selling to an American Big Tech business.
Since 2022, Tunstall-Pedoe has been building London-based startup Unlikely AI — and he says that he wants to rectify the fact that the US has six trillion-dollar tech companies, while Europe has none.
“That’s what we’re trying to build,” he tells Sifted. “We're trying to build a very significant UK standalone business.”
Unlikely AI thinks it can get there by building an enterprise AI platform that combines many of the capabilities of machine learning techniques that power large language models (LLMs), with the advantages of more traditional software that can be relied on to not make mistakes.
“In the case of LLMs, wrong answers are often extremely plausible. They're often as plausible as the correct answers. So the world is becoming actually a much less trusted place,” says Tunstall-Pedoe.
“But with spreadsheets, databases — that kind of software produces answers that are always right. If it doesn't, it's considered to be a bug. If your spreadsheet ever miscalculated the value of a cell, that would be considered to be a very serious bug, potentially suing Microsoft for it.”
Unlikely AI is competing with a number of other startups trying to crack AI’s trust problem, as enterprises around the world puzzle over how to bring the technology into their businesses. The company has the backing of Amadeus Capital, Skype cofounder Jaan Tallinn’s Metaplanet and Octopus Ventures — and announced a $20m oversubscribed seed round in September 2022.
Like many of its competitors, the startup is in pre-launch phase and is working with an early design partner as it refines its product in the race to build a powerful and dependable AI system for the modern business, which it says will hit the market “soon.”
The best of both worlds?
Today’s most powerful generative AI models — like OpenAI’s o1 — are designed in a way that means occasional errors in their output are practically impossible to fully eliminate, making them unsuitable for many business use cases. They also are, by design, “black boxes," as the vast amounts of data they crunch make it impossible for a human or the system to explain how an output was reached.
But, as Tunstall-Pedoe points out, the machine learning and deep learning technology that powers these models also creates huge value.
“For things like understanding lots of images, processing of complex data, understanding natural language, keyword searches over documents — to solve those kinds of problems you need machine learning,” he says.
“The world we want to get to is where you have all the capabilities of deep learning, and you combine that with the trust, explainability and accuracy of the conventional software world — and we have technology that bridges those two.”
Unlikely AI is not giving away too much about how its system has cracked this, but does say that it uses an approach that is more dependable and explainable than an LLM, according to CTO Tom Mason. This is partly because it draws on well-sourced information, rather than the huge swathes of often-unreliable written material from the wider internet that models like ChatGPT are trained on, and it can cite exactly where its knowledge from.
Selling a product
The Unlikely team is also keeping its cards close to its chest in terms of how its product will work and what it will be useful for, but does say that it will be helpful to organisations working in the legal, finance, healthcare and insurance sectors.
“Sectors where reliance upon AI systems has zero [error] tolerance,” says Mason, adding that the system will be useful for “reliably cross referencing data in different places and making inferences and coming up with decision-based intelligence across lots of different information sources in a reliable way.”
He says the team is in the latter stages of developing a system it can sell to customers, and is currently focusing on one of the aforementioned four sectors, but wouldn’t say which.
Mason, who previously worked at generative AI startup Stability AI — which operated a “models-as-a-service” business where it sold generative models to business clients — says he’s confident that Unlikely is selling a product that makes sense to enterprises.
“Throughout the last few years, I’ve been talking to a lot of companies who were going: ‘Wow, this technology is amazing but how do we use it?’ — and never really having a completely reliable answer for them,” he says. “That nagging feeling has always been quite a difficult thing to deal with.”
Mason says this experience has led him to focus on building “something tangible that users can buy into,” and that Unlikely AI’s system has “quite a recognisable pattern of what you're buying and using” across different types of business.
“All of the chief AI officers of these huge banks and other companies are looking for ways to make this work, and they've hit a lot of barriers and brick walls trying to implement the current wave of LLMs. So actually, they're really very open for new solutions and new ideas,” he adds.
The cost of cutting edge AI
Unlikely AI has one advantage over companies trying to build a viable business model by selling generative AI technology: it’s not building a large model that requires terabytes of data and huge amounts of computational power from cloud providers like AWS and Microsoft Azure.
“We're not training a massive model. We're not going to be burning through hundreds of millions of dollars of AWS costs,” says Tunstall-Pedoe, reflecting on the question of whether it makes economic sense for startups to train their own massive AI systems and try to accrue value from them.
“I think LLMs have largely become commoditised. I think there's not a lot of differentiation, and it's incredibly expensive to build a big foundation model. So I'm glad we're not in that business.”
Unlikely AI’s biggest cost, therefore, is talent, as companies around Europe try to snap up the brightest machine learning and data science minds — but Tunstall-Pedoe believes that his experience of selling his last company to Amazon is helping him recruit.
“I think that we have a reputation as being a very exciting place to work. I think the story of my last startup helps quite a bit,” he says.
Tunstall-Pedoe is competing with other European startups like Aligned AI, Unmai and Conjecture in developing more reliable AI architecture for enterprises. But he thinks that his team — now around 60-people strong, with experience at companies including Google, Amazon, Microsoft and IBM — is getting close to a system it can take to market.
As it does, Unlikely AI will continue its hiring efforts for the best technical talent, as well as building out a go-to-market strategy. And while Tunstall-Pedoe says that for many companies, being acquired by a Big Tech giant can make sense, his new startup won’t be looking for a buyer any time soon.
“An acquisition isn't necessarily a negative thing, it can be a very positive thing. It results in skills being developed and more money being pumped into the UK economy,” he says. “But that’s not what we’re planning for with Unlikely AI.”