Real-world agentic AI is 'complex' cautions Google's Demis Hassabis
The DeepMind CEO warned of the long road ahead
CEOs and tech vendors are pressing for more AI, but DeepMind CEO sounds a warning note.
It will be no surprise to anyone that Google is going big on artificial intelligence. But it’s not just about consumer-facing systems like Gemini; Google Cloud CEO Thomas Kurian says the company is investing massively in models for science and business.
Speaking at a press event in London today (see all the news here), Kurian said Google’s “general vision” is to offer “a broad range of models” – including those that can generate text, images, video and scientific formulae, as well as increasing business efficiency.
Advertising giant WPP is already using Google’s AI for that latter goal, on its own WPP Open platform. More than 40,000 people in the company use it every day, a number that is up 7% this year already.
“I don’t think this is overhyped,” said WPP CEO Mark Read. “It’s not about replacing people, it’s about augmenting them… I don’t think there’s a part of our business that won’t eventually be touched by AI.”
Demis Hassabis, CEO of Google DeepMind, said the next era for AI will be in creative talent like advertisers embracing these opportunities.
“[The time saving from AI] will allow a lot more creativity to happen.”
While WPP is in the creative game, the telecoms business is very different – and so, BT is using AI very differently, said CEO Allison Kirkby.
“Not many people know that we create as many AI patents every year as DeepMind. We’re investing heavily in AI research.”
BT is using AI in customer service (where ‘Aimee’ has a net promoter score higher than human operatives), but also in the backend to make networks safer and more efficient. Even end-user customers are seeing the benefits; Google Gemini, available on the most expensive EE tariff, is able to optimise a customer’s 5G network experience.
Agent of change

The next big step in AI technology will be agentic AI, but Demis cautioned that this is a complex area. Gaming, where Google has already had some success with agentic AI, is very different from the real world. Building world models – AI that can understand the world around us – and combining them with planning algorithms is very hard.
“If your world model has just a 1% error rate, if you build over 50 or 100 steps, that 1% compounds. By the time you’ve done those 50 or 100 steps, you’re in potentially a random place.”
That isn’t stopping companies from experimenting with AI agents – WPP has built more than 28,000 to date, and Mark said he had even trialled a CEO agent.
“Those agents are getting smarter. They can test ideas against audiences or against experts. That is really going to allow us to optimise and automate more and more of the workflow.”
An existing restriction on agentic AI is inter-agent communication. Currently, humans have to do this; any communication between agents is programmatic.
Demis talked about advances in multi-agent models (Google uses approach this in gaming), with “a society or league of agents” that “could be competing or could be cooperating.”
Thomas said Google is currently working on a way for AI agents to describe themselves to other agents (“What are your skills? What tools do you use? How much do you cost to use?”), which will form a new interface for agentic AI.
And after agentic models, the goal is to build AGI, or artificial general intelligence. Demis believes the opening moves will be here within the next decade.
“Today’s systems are impressive, but there’s a lot they can’t do. Within ten years we’ll start moving towards AGI, which has always been the end goal.”
There is even a step after AGI known as ASI, or artificial super intelligence: self-improving systems that can build and refine their capabilities. While this has some exciting possibilities, particularly in fields with defined rules like coding, Hassabis said it was still speculative for now.
The elephant in the room

No AI discussion – at least one where journalists are allowed to ask questions – would be complete without some mention of IP and ethics.
In response to our question about the use of copyrighted IP for AI training, Demis said this is “an important area” that “needs clarification” (precedent is just starting to be established).
“The complication is that these models are kind of global, they’re used everywhere [but] trained in one place… We need to set an international standard for this.”
He added that the UK must be a leader in this area, and stressed the importance of events like the Bletchley Summit and AI Action Summit in Paris.
While Demis sounded several cautionary notes throughout the event, it’s clear that the pace of change is not going to slow down – and neither tech vendors nor CEOs, at least, want it to.
“As a country we have a productivity challenge, and over time this will be a big unlocker of that challenge,” said Mark. He added, “Part of my job as CEO is to do public demos, to show people that if I can use it, they can use it.”