It may be take up to 10 years for a lot of the capabilities of artificial intelligence (AI) come to the fore, according to the founder and CEO of DeepMind, Demis Hassabis.
Speaking at an event to mark the availability of audio generation model Chirp 3 on the Google Vertex AI platform, Hassabis said that within that time, AI will have evolved to artificial general intelligence, whereby the AI system exhibits “the cognitive capabilities” of humans. “That’ll be a moment when we have finally arrived with a kind of general intelligence, which is the original aim of the whole field of AI,” he added.
The Chirp 3 audio generation model is generally available in 31 languages, offering 248 distinct voices with eight speaker options. According to Google, Chirp 3 on Vertex AI delivers detailed speech functionality that captures the nuances of human intonation, making conversations more engaging and immersive. It can be applied to voice annotation, real-time meeting transcription, audiobooks, and sentiment collection from customer calls.
Google Cloud also announced that it is expanding its UK data residency commitment to include Google Agentspace, with availability coming in the second quarter of 2025. Agentspace provides AI agents that combine Gemini’s reasoning with Google Search and enterprise data, regardless of where the data is hosted.
Google Cloud said Agentspace includes a single, company-branded, multimodal search agent that employees can use to answer complex questions and take specific actions based on an organisation’s proprietary information, including unstructured data such as documents and information stored in third-party applications. It also includes a feature called NotebookLM Enterprise, which Google Cloud said can help employees quickly synthesise large quantities of information to uncover new insights.
The age of agents
Looking at how AI is evolving in the short to medium term, Hassabis said: “We’ve been improving our ‘thinking portion’ for the models, so inference time, compute is going to be a big thing this year as the systems become more agent-like.” This involves the AI agents spending more time thinking and planning before they act.
Analyst firm Forrester regards agentic AI as AI systems that can plan, decide and act autonomously, orchestrating complex workflows with minimal human intervention. In a blog post, Forrester analysts said agentic AI systems are not only poised to become the backbone of the knowledge economy, but they will completely redefine how organisations operate and compete.
Hassabis said the effects of agent-based AI are beginning to be felt. “They’re not just passive question-answering systems, but they can break down a problem into sub goals and then choose those goals.”
DeepMind has previously used this type of problem-solving in game play, such as with AlphaGo, the AI that beat Go world champion Lee Sedol in 2016. However, Hassabis said: “Games are very limited and they’re quite easy. The rules are prescribed information, so they’re relatively easy setups compared to the real [world].”
The challenge for DeepMind and other AI developers is how quickly they can generalise AI systems to exhibit behaviours required for planning and reasoning that work reliably in the real world.
While Hassabis said there has been good progress in AI world models over the past few years, the challenge is how best to combine these with planning algorithms. As an example of why this is very difficult, Hassabis said: “If your AI model has a 1% error rate and you plan over 5,000 steps, that 1% compounds like compound interest.”
By the time the 5,000 steps have been worked through, the compounded error, according to Hassabis, means the possibility of the answer being correct is random. “For a games model, you have the rules of Chess or Go,” he said, which aids the planning algorithm in making the correct decision. “In the real world, you don’t have perfect information. There’s hidden information that we don’t know about, so we need AI models that are able to understand the world around us.”
For Hassabis, one of the interesting developments expected to appear over the next few years is the deployment of multiple AI agents that work together to solve a problem.
“We’ve done a lot of work with things like StarCraft II [the real-time strategy game], and in the past, where you have a society of agents or a league of agents, they could be competing or cooperating,” he said.
If agents can help people complete tasks, Hassabis believes it makes sense to have a whole set of them with complementary skills. “How those agents should cooperate or compete with each other is a very interesting branch of research,” he added.
He anticipates that such multi-agent AI systems will start becoming useful in the next couple of years. “You’ve got this notion of a general agent, something like Gemini, but then it can call specialised agents for mathematics or programming, for example,” he says.
Even though these are AI systems themselves, Hassabis said AI agents could be among the tools that a general AI system uses to solve problems.
An early example of how bringing different types of AI techniques together can be seen in the AI Co-scientist tool released a couple of weeks ago. This, according to Hassabis, is a hybrid system with Gemini under the hood, but it also uses specialised AI systems to look up scientific papers and try to make connections between different pieces of research.
Co-scientist is designed to mirror the reasoning process underpinning the scientific method and, according to Google, is intended to uncover new, original knowledge, and to formulate demonstrably novel research hypotheses and proposals, building on prior evidence and tailored to specific research objectives.
