Alexa co-creator William Tunstall-Pedoe says AI lacks the intelligence to drive a singularity
William Tunstall-Pedoe says current AI can generate ideas but cannot reliably identify which ones are valuable, a distinction that underpins UnlikelyAI's neurosymbolic approach.
By Ryan Merket · Published
Why it matters
Tunstall-Pedoe is drawing a line between broad model competence and autonomous invention, a distinction that underpins UnlikelyAI's push into regulated enterprise decisions.
William Tunstall-Pedoe, whose voice-assistant technology helped produce Amazon Alexa, rejected Sam Altman's claim that humanity has entered an AI singularity. His argument also explains the technical and commercial bet behind his current startup, UnlikelyAI.
In an as-told-to essay published by Business Insider on August 7, 2026, Tunstall-Pedoe said current systems have become broadly competent without demonstrating what he calls "discovery intelligence": the ability to produce, evaluate and pursue genuinely valuable ideas without relying on human judgment to identify which ones matter.
Tunstall-Pedoe has spent much of his career trying to make machines understand and act on human language. He learned to program as a teenager at the High School of Dundee, then completed Cambridge's first three-year computer science degree, according to a Royal Society profile. He later founded True Knowledge, renamed Evi Technologies, to build a system that could answer spoken or typed questions.
Amazon acquired Evi in 2012. Tunstall-Pedoe joined the Alexa effort and remained at Amazon until 2016. He then returned to founding with UnlikelyAI, pursuing an architecture that combines neural models with symbolic or conventional software methods.
The capability he thinks AI is missing
Tunstall-Pedoe accepts that large language models outperform people in several dimensions. They can discuss a wider range of academic subjects than any individual and retrieve patterns from volumes of information no person could memorize. Computers had already surpassed humans at arithmetic and storage long before chatbots became mainstream products.
Those capabilities do not prove that intelligence is accelerating without a ceiling, he argues. Language models learn from large collections of human-produced data and become proficient at reproducing patterns within it. They can generate ideas, yet still depend on people or external systems to decide whether an idea is useful, feasible or important.
"AI would need to show stronger evidence of what I call discovery intelligence," Tunstall-Pedoe wrote in the Business Insider essay.
That distinction sets a higher bar than benchmark gains, lower inference costs or wider chatbot adoption. Under Tunstall-Pedoe's definition, a singularity would require AI systems that can improve AI itself through original invention, recognize which inventions are valuable and repeat the cycle fast enough to produce an unmistakable acceleration across society.
Business Insider presented Tunstall-Pedoe's essay as a response to Altman's statement on the Relentless podcast that "we are now, like, in the singularity." Tunstall-Pedoe separates the spread of powerful products from autonomous technological acceleration. Friends outside technology are contacting him about AI's effects on their work, he said, but rapid adoption still falls short of a system inventing and validating its own successors.
The critique is also UnlikelyAI's product thesis
UnlikelyAI has built its business around an architecture that combines neural models for language and pattern recognition with symbolic or conventional software methods. Those additional methods are intended to add structure, control and auditability to AI-assisted decisions.
UnlikelyAI's public materials position the startup toward enterprise and regulated settings, including finance and healthcare, where errors can carry material consequences. UnlikelyAI says its approach can improve explainability, accuracy and auditability; those performance claims remain its own, with no independent benchmark established in the supplied reporting.
UnlikelyAI's current website targets decisions in regulated industries, where a fluent answer has limited value unless a customer can trace and defend it.
UnlikelyAI announced a $20 million seed round in September 2022, co-led by Amadeus Capital Partners and Octopus Ventures. Cambridge Innovation Capital and Jaan Tallinn's Metaplanet participated, alongside angels including former Google finance chief Patrick Pichette and former Amazon UK executive Christopher North. UnlikelyAI had previously raised a 1.1 million pound angel round in October 2020.
The singularity debate arrives during a consequential period for that bet. Sifted reported in June 2026 that UnlikelyAI had experienced senior departures, widening losses and a strategic shakeup. The shift toward specific regulated workflows gives Tunstall-Pedoe a nearer-term market test: whether enterprises will pay for constrained, explainable systems instead of relying on general-purpose models with added guardrails.
A practical test for extraordinary claims
Tunstall-Pedoe's standard makes the singularity observable. A system capable of recursive invention would produce changes far beyond improved chatbot answers or faster software development. It would create valuable discoveries, evaluate them and use them to improve its own capabilities at a pace visible throughout science, industry and everyday life.
"We would recognize it if it were here," he said.
The threshold remains difficult to measure because useful invention is often recognized only after deployment, peer review or market adoption. Still, it gives founders and researchers a more concrete test than model intelligence as a single rising score. They can ask whether an AI system merely proposes possibilities or can establish, through evidence and execution, which possibility deserves to become a product, experiment or new technical direction.
Tunstall-Pedoe built Evi around one hard problem: allowing computers to understand questions expressed in ordinary language. With UnlikelyAI, he is pursuing the next constraint he sees in the stack, turning statistical fluency into decisions people can verify. His disagreement with Altman follows directly from that experience. Current AI can perform remarkable work. Tunstall-Pedoe argues that the self-directed discovery machine implied by the singularity still has to be built.