Amazon Alexa inventor says most companies are getting AI wrong

William Tunstall-Pedoe, the inventor behind Amazon’s voice assistant Alexa, has said that many companies that are trying to embed AI in their operations are getting it wrong.

“Pretty much every medium and large business out there is trying to bring AI into their business,” he told City AM. “But a lot of these projects are failing, and the reason why they’re failing is because the [machine learning] ML piece doesn’t work well enough.

“It isn’t trustworthy enough, goes wrong every now and then in a way that potentially costs money or is outside of regulatory requirements, or is brand damaging,” he added.

After selling his voice assistant startup, Evi Technologies, to Amazon in 2012, he worked on the creation of Alexa, but now he is working on a new venture, Unlikely AI.

Although the deep tech startup has mostly been kept in the shadows so far, Tunstall-Pedoe said it is designed to tackle the issues that are tripping up many companies using large language models (LLMs), such as bias, hallucination, accuracy and trustworthiness.

“We are trying to create a platform that addresses all these problems, so in industries that are particularly high risk, and that includes finance, that includes health, anything to do with people, we’re basically creating an AI platform that’s trustworthy and explainable, so that businesses can can apply AI in a very confident way,” he explained.

London-based Unlikely AI has been quietly developing its technology supported by a $20m (£15.4m) seed round led by Amadeus Capital and Octopus Ventures.

In the meantime, it has been beefing up its team, recently appointing Fred Becker, who helped scale up Skype, as chief administrative officer, and Tom Mason as chief technology officer. Mason joined from Stability AI, where he led the development of its core platform.

Mason said Unlikely AI’s platform will allow businesses to solve complex problems. For example, it could automate claims processing on behalf of a customer against company policies, “in a way that’s predictable”, unlike traditional language models.

City AM understands the company plans to launch commercially in the second quarter of next year.