Google DeepMind’s Demis Hassabis and the paradox of AI progress

If Google’s deal with Apple and its recent Gmail integration are any indication, the search giant is pulling ahead of the AI pack. It’s planning an unprecedented buildout of accelerated-compute data centers to run and train AI models, and it has reoriented itself around the generative AI boom kicked off by OpenAI’s ChatGPT. London-based DeepMind, a research lab that for years operated without pressure to contribute to Google’s bottom line, is now central to transforming the company’s core products.

But the AI boom is reshaping many of the dynamics that produced past research breakthroughs. The massive wave of investment in data centers is no longer just about inventing new technology and generating new breakthroughs, but about running the AI we have for ever-growing numbers of users. “We’re now also in the era of full commercialization of these systems, so you’ve also got to balance serving with training,” Hassabis said.

The global buildout of AI data centers means more entities are now competing for a limited number of resources, including semiconductors, memory, and energy. “Right now, it’s memory chips, but it’ll probably be something else tomorrow,” he said.

Adoption in areas that create real-world economic impact are slow and uneven. Companies have changed drastically since the ChatGPT moment, but it’s hard to get a read on any measurable return on investment.

Another sand in the gears is that companies are now opting to keep research private that they might otherwise have shared publicly, slowing down the cross-pollination of ideas that once helped AI go from a research backwater to a hotbed of activity. “There’s so much commercial pressure, so some of those things can’t be shared quite as openly anymore, which is a shame on the one hand, but it’s understandable,” Hassabis said.

One area that could impede AI’s progress is a growing populist disdain for the technology that spans the political divide. Some political candidates have begun to campaign on anti-AI platforms, accusing tech companies of causing spikes in energy prices. In towns across the country, citizens are organizing grassroots opposition to data-center construction. And climate activists oppose the technology because of emissions used to power the data centers.

One way to combat the opposition, Hassabis said, is to use AI in science, to create new breakthroughs in health care, materials science, and energy.

The biotech industry is using AI to accelerate drug discovery, while other AI tools are helping identify new materials for applications ranging from carbon capture to better batteries that can store energy produced by renewable sources like solar and wind. And DeepMind has used AI to tackle nuclear fusion. “One of the only ways to tackle climate in today’s fragmented political world is to come up with some new technologies,” Hassabis said. “I actually think AI as an industry needs to be doing more of that.”

It’s hard to imagine a time when the sharpest minds in AI research, many of whom were on the payroll of big tech companies, once gave away their greatest discoveries for free.

I’ve asked people who work for top AI labs how the dynamic of secrecy affects AI progress, and the response tends to be that it doesn’t. A common refrain is that talent moves so quickly between companies, the research might as well be published.

That’s why I thought it was telling that Hassabis thinks otherwise. While he didn’t quantify the effect, it made me think more broadly about other ways the AI arms race actually makes breakthroughs less likely.

It will be difficult to replicate the golden era of tech, when top AI researchers tapped private-sector funding and the freedom of academia. There were so many huge advances during that time that all fed off one another.

“There were only a few dozen researchers working on AI properly and we all knew each other. We were, in most cases, really good friends,” Hassabis said.

But the paradox of AI progress is that all the commercial success of generative AI might actually stretch out the timeline and delay whatever comes after generative AI. And that might not be so bad.

Change that happens too quickly could be destabilizing in an already unstable world.