After announcing a major $20 million investment last year, Unlikely AI founder William Tunstall-Pedoe has kept the emerging UK foundation model maker’s vision close. So far: TechCrunch can exclusively reveal it’s taking a “neuro-symbolic” approach to its AI. In an additional development, it’s announcing two senior appointments — including Stability AI’s former CTO, Tom Mason.
Neuro-symbolic AI is a type of artificial intelligence that, as the name suggests, integrates modern neural network approaches to AI – as used by large language models (LLMs) such as OpenAI’s GPT – and earlier symbolic AI architectures to overcome the weaknesses of each.
Tunstall-Pedoe first hit the UK tech scene in 2012 when Amazon acquired his voice assistant startup, Evi. Two years later Amazon launched Echo and Alexa, which incorporated much of Evi’s technology. With Unlikely AI, Tunstall-Pedoe is aiming to put himself back in the spotlight as he unveils the technology he and his team have been working on since 2019, when the startup was founded.
Meanwhile, at Stability AI, Mason managed the development of key foundational models across a variety of fields and helped the AI company raise over $170 million. He is now the CTO of Unlikely AI, where he will oversee its ‘symbolic/algorithmic’ approach.
In addition, Fred Baker is also joining as Chief Administrative Officer. He has previously held senior positions at companies such as Skype and Symphony. His role at Unlikely will now be to guide 60 full-time employees – primarily based between Cambridge (UK) and London.
The AI startup claims that its approach to foundational AI models will seek to avoid the risks we have become too familiar with – such as bias, ‘confusion’ (aka fabrication), accuracy and trust. It also claims that its approach will use less energy to reduce the environmental impact of Big AI.
“We’ve been operating privately for a number of years, and we’re very excited about our two new senior hires,” Tunstall-Pedoe told TechCrunch.
Explaining the team’s approach, he said: “We are building a ‘trustworthy’ AI platform designed to address all the major issues with AI at the moment, as it relates to confusion and accuracy. We are combining the capabilities of generative AI, statistical AI, symbolic algorithmic methods, (and) traditional software methods to achieve extensibility and reliability.”
He described the platform as “horizontal,” saying it would “combine many different types of applications.”
As for the exact applications, he was coy – but continued to emphasize the phrase “trustworthy AI.”
For his part, Mason said his time at Stability AI saw the company build “some amazing models” and “an incredible ecosystem around the models and the technology,” as he put it. This also included the sudden exit of founder Imad Mostek, followed by several other high-profile team departures. While Mason wishes his former colleagues “all the best,” he said he is “extremely excited” to join Unlikely AI.
Explaining the startup’s technology, Tunstall-Pedoe said the platform is made up of two things: “The word neuro and the word symbolic. Neuro means deep learning, i.e. solving problems that machines haven’t been able to solve for decades… Symbolic refers to the kind of software that powers your spreadsheet or other application.”
“One of the weaknesses of ‘neuro’ is that it is sometimes wrong. When you train a model, you feed it data, it gets better and better. But it never reaches 100%. For example, it is right 80% of the time, which means it is wrong 20% of the time.”
He said this is “incredibly damaging to faith” because “neuro computation is unclear.” In fact, there’s an entire field of research dedicated to understanding what happens inside these giant LLMs.
Instead, he said Unlikely plans to combine the certainties of traditional software, such as spreadsheets, where calculations are 100% accurate, with a “neuro” approach to generative AI.
“What we’re doing is combining the best of both worlds,” suggested Tunstall-Pedoe. “We’re taking the capabilities of AI, all the advances in deep learning, and we’re combining it with the reliability and scalability and other benefits of non-statistical machine learning – including things like cost and environmental impact… Our vision of AI is all of those capabilities, but in a way that’s completely trustless.”
He argues that a combined approach would also bring cost and environmental benefits compared to today’s LLMs: “These models are incredibly expensive (to run) and bad for the environment, but they are also costly in terms of trust because they give the wrong answer.”
Why didn’t the other basic models take a similar path?
“I think that’s happening,” Mason responded. “Sometimes we talk about it as ‘compound architecture.’ We’ve seen the rise of things like RAG. That’s a kind of compound architecture. It’s very much the same thing, but it’s building on all of that with the benefits of symbolic logic, making it possible to do completely precise reasoning.”
In this regard, he said Mason believes Unlikely AI is “ahead of the wave.”
Another question is whether Unlikely AI will create a full baseline model like OpenAI – or take a mixed approach like Mistral, offering both baseline and open source models?
Tunstall-Pedoe said the company had yet to decide its direction of travel: “We haven’t made any such decision yet. It’s part of the internal discussions. But we are building a platform and everything else is yet to be decided… It’s a decision we’re going to make in the near future.”
One thing is for sure, though: it will be built in London and Cambridge: “Obviously we have a much smaller population than the US and China. But London is a great place to build an innovative AI startup. There’s a lot of talent here. There’s a lot of innovation.”
While the timeline for a model release is unclear, Unlikely AI is certain about the strength of its ambition. Noting that AI is the number one strategic priority of every trillion-dollar market cap company, Tunstall-Pedoe said he is shooting for major adoption. “We want to be massively successful, we want to have a big impact. We’re definitely open to different ways of achieving that,” he said.
