Since announcing a massive $20 million seed round last year, Unlikely AI founder William Tunstall Pedor has kept a tight rein on the up-and-coming UK-based model maker's efforts. But TechCrunch exclusively revealed that the company is taking a “neurosymbolic” approach to AI. Additionally, Unlikely AI has announced two senior executive hires, including former Unlikely AI CTO Tom Mason.
Neurosymbolic AI, as the name suggests, is a type of artificial intelligence that brings together both modern neural network approaches used in large-scale language models (LLMs) such as OpenAI’s GPT, and older symbolic AI architectures, addressing the weaknesses of each.
Tunstall-Pedoe first made waves in the UK tech world when Amazon acquired his voice assistant startup, Evi, in 2012. Two years later, Amazon launched the Echo and Alexa, which incorporated much of Evi's technology. With Unlikely AI, Tunstall-Pedoe is aiming to return to the spotlight by showcasing the technology he and his team have been working on since the startup was founded in 2019.
Meanwhile at Stability AI, Mason managed the development of key foundational models across a range of disciplines and helped the AI company raise over $170M. Now, as CTO at Unlikely AI, he oversees the company's “symbolic/algorithmic” approach.
Additionally, Fred Becker will join the company as chief administrative officer, having previously held senior executive roles at companies such as Skype and Symphony. At Unlikely, he will manage 60 full-time staff, primarily based in Cambridge (UK) and London.
The AI startup claims that its approach to foundational AI models aims to avoid the risks of bias, illusion (i.e. fabrication), accuracy, and reliability that we have quickly become familiar with. It also claims that the approach will be less energy intensive in order to reduce Big AI's environmental impact.
“We've been working privately for a number of years, but we're very excited about our two new senior executive hires,” Tunstall Pedor told TechCrunch over the phone.
Elaborating on the team's approach, he continued, “We are building a 'trusted' AI platform designed to address almost all of the major issues with AI today, including hallucinations and accuracy. We combine the capabilities of generative AI, statistical AI and symbolic algorithmic techniques to [and] We went with traditional software approaches for scalability and reliability.”
He explained that the platform is “horizontal” in the sense that it “integrates different types of applications.”
He was coy about specific applications, but kept emphasizing the phrase “trustworthy AI.”
Meanwhile, Mason said that during his time at Stability AI, the company built “some amazing models” and “an incredible ecosystem around the models and technology.” The company was also notable for the sudden departure of founder Emad Mostakeh, followed by other notable team members. While Mason wished his former colleagues “the best of luck,” he said he is “really excited” to be joining Unlikely AI.
Digging into the startup's tech, Tunstall Pedor said the platform is made up of two components: “The word neuro and the word symbolic. Neuro means deep learning, solving problems that machines haven't been able to solve for decades. Symbolic refers to the type of software that powers spreadsheets and other applications.”
“One of the weaknesses of neuro is that it gets it wrong sometimes. As you train the model and feed it data, it gets better and better. But it's never 100%. So, like, 80% of the time it's right, that means 20% of the time it's wrong.”
He said this is “very damaging to trust” because “neurocomputation is opaque.” In fact, there is an entire field of research trying to understand what goes on inside these giant LLMs.
Instead, he said, Unlikely plans to combine the certainty of traditional software, such as spreadsheets whose calculations are 100% accurate, with the “neuro” approach of generative AI.
“What we're doing is combining the best of both worlds,” Tunstall Pedor suggests. “We're taking the capabilities of the JD program, all the advances in deep learning, and combining it with the reliability, scalability, and other benefits (such as cost and environmental impact) of non-statistical machine learning. Our vision of AI is to deliver all these capabilities in a completely trustworthy way.”
He argues that compared with current LLMs, the combined approach also offers cost and environmental benefits: “These models are incredibly expensive. [to run] Not only is it environmentally unfriendly, but giving the wrong answer also has a cost in terms of credibility.”
Why didn't other foundational models follow a similar route?
“I think that's happening,” Mason replied. “We sometimes call this 'composite architectures.' We've seen things like RAG emerge, which is a type of composite architecture. This is exactly the same vein, but building all of this with the benefits of symbolic reasoning, allowing for completely accurate reasoning.”
In this regard, Mason believes Unlikely AI is “ahead of the wave,” he said.
Another question is whether Unlikely AI will create a more complete foundational model like OpenAI, or adopt a mixed approach like Mistral, providing both a foundational model and an open source model.
Tunstall Pedor said the company hasn't yet decided on a direction: “We haven't made that decision yet. That's part of our internal discussions. But we're building the platform and the rest is up in the air. We'll make that decision in the near future.”
But one thing is certain: it will be built in London and Cambridge: “Obviously, we have a much smaller population than the US or China, but London is a fantastic place to build an innovative AI startup. There's a lot of talent here and a lot of innovation.”
While the timeline for model release is unclear, Unlikely AI is confident about the strength of its ambitions. With AI being a top strategic priority for any trillion-dollar company, Tunstall-Pedoe said the company is aiming for mass adoption. “We want to be hugely successful and have a big impact, and we're looking at different ways to do that,” he added.