The surge of money flooding into artificial intelligence has resulted in some crypto-like hype that is obscuring the incredible scientific progress in the field, according to Demis Hassabis, co-founder of DeepMind.
The chief executive of Google’s AI research division told the Financial Times that the billions of dollars being poured into generative AI start-ups and products “brings with it a whole attendant bunch of hype and maybe some grifting and some other things that you see in other hyped-up areas, crypto or whatever.
“Some of that has now spilled over into AI, which I think is a bit unfortunate. And it clouds the science and the research, which is phenomenal,” he added. “In a way, AI’s not hyped enough but in some senses it’s too hyped. We’re talking about all sorts of things that are just not real.”
The launch of OpenAI’s ChatGPT chatbot in November 2022 sparked an investor frenzy as start-ups raced to develop and deploy generative AI and attract venture capital funding.
VC groups invested $42.5bn in 2,500 AI start-up equity rounds last year, according to market analysts CB Insights.
Public market investors have also rushed into the so-called Magnificent Seven technology companies, including Microsoft, Alphabet and Nvidia, that are spearheading the AI revolution. Their rise has helped to propel global stock markets to their strongest first-quarter performance in five years.
But regulators are already scrutinising companies for making false AI-related claims. “One shouldn’t greenwash and one shouldn’t AI wash,” said Gary Gensler, chair of the US Securities and Exchange Commission, in December.
In spite of some of the misleading hype about AI, Hassabis, who last week received a knighthood for services to science, said he remained convinced that the technology was one of the most transformative inventions in human history.
“I think we’re only scratching the surface of what I believe is going to be possible over the next decade-plus,” he said. “We’re at the beginning, maybe, of a new golden era of scientific discovery, a new Renaissance.”
The best proof of concept for how AI could accelerate scientific research, he said, was DeepMind’s AlphaFold model, released in 2021.
AlphaFold had helped predict the structures of 200mn proteins and was now being used by more than 1mn biologists around the world. DeepMind is also using AI to explore other areas of biology and accelerate research into drug discovery and delivery, material science, mathematics, weather prediction and nuclear fusion technology. Hassabis said his goal had always been to use AI as the “ultimate tool for science”.
DeepMind was founded in London in 2010 with the mission to achieve “artificial general intelligence” that matches all human cognitive capabilities. Some researchers have suggested that AGI may still be decades away, if attainable at all.
Hassabis said that one or two more critical breakthroughs were needed before AGI was reached. But he added: “I wouldn’t be surprised if it happened in the next decade. I’m not saying it’s definitely going to happen but I wouldn’t be surprised. You could say about a 50 per cent chance. And that timeline hasn’t changed much since the start of DeepMind.”
Given the potential power of AGI, Hassabis said it was better to pursue this mission through the scientific method rather than the hacker approach favoured by Silicon Valley. “I think we should take a more scientific approach to building AGI because of its significance,” he said.
The DeepMind founder advised the British government about the first global AI Safety Summit held at Bletchley Park last year. Hassabis welcomed the continuing international dialogue on the subject, with subsequent summits due to be held by South Korea and France, and the creation of UK and US AI safety institutes.
“I think these are important first steps,” he said. “But we’ve got a lot more to do and we need to hurry because the technology is exponentially improving.”
Last week, DeepMind researchers released a paper outlining a new methodology, called SAFE, for reducing the factual errors, known as hallucinations, generated by large language models such as OpenAI’s GPT and Google’s Gemini. The unreliability of these models has led to lawyers making submissions with fictitious citations and deterred many companies from using them commercially.
Hassabis said DeepMind was exploring different ways of fact checking and grounding its models by cross-checking responses against Google Search or Google Scholar, for example.
He compared this approach to the way that its AlphaGo model had mastered the ancient game of Go by double-checking its output. A large language model could also verify whether a response made sense and make adjustments. “It’s a little bit like AlphaGo when it’s making a move. You don’t just spit out the first move that the network thinks about. It has some thinking time and does some planning,” he said.
When challenged with authenticating 16,000 individual facts, SAFE agreed with crowdsourced human annotators 72 per cent of the time — but was 20 times cheaper.