DeepMind CEO: Why AGI Remains No Match for Human Reasoning

Artificial general intelligence may be advancing fast, but it isn’t yet thinking on the same level as humans.

This was the message delivered by Demis Hassabis, CEO at Google DeepMind, during the AI Impact Summit in Delhi. 

Demis stated that, while AI excels when it comes to narrow, task-specific problems, AGI – general-purpose intelligence that can reason and learn across unfamiliar domains – remains out of reach.

“I don’t think we are there yet,” he said, pointing to three stubborn gaps: continual learning, long-horizon planning and consistency across tasks.

On continual learning, he contrasted human adaptability with today’s largely static models: "What you’d like is for those systems to continually learn online from experience, to learn from the context they’re in, maybe personalise to the situation and the tasks that you have for them."

AI lacks long-term memory

Long-term planning is another fault line. Models can chart short sequences of actions, but struggle to sustain strategies over months or years, according to Demis. 

“They can plan over the short term, but over the long term, the way that we can play over years, they don’t really have that capability at the moment," he added.

Then there is consistency. Even as frontier systems solve Olympiad-level problems, they can still fumble elementary questions when phrased differently.

“A true general intelligence system shouldn’t have that kind of jaggedness,” Demis said.

To illustrate the creative and conceptual gulf, Demis offered a historical stress test: train a foundation model with a knowledge cut-off in 1911 and ask it to independently discover general relativity by 1915.

“That would be a good test for AGI and I think today’s systems clearly would not be capable of doing that.”

AI as a co-researcher

If fully-general intelligence remains out of reach, Demis sees immense near-term value in AI as a co‑researcher.

He forecast a “new golden era for scientific discovery” as human experts team up with generalist and specialist models.

One practical pattern, he suggested, is orchestration: a broad system such as Google Gemini delegating to highly-capable domain tools like AlphaFold when protein structure understanding is required.

“It would be better for it to call AlphaFold as a tool than put all that protein information into the main system,” he said.

Risks and rules

Demis also underscored technological risks. Beyond potential misuse of AI by individuals or states, he warned of technical and societal challenges and suspects the latter could prove harder.

Near‑term priorities include bolstering cyber and biosecurity and building international standards for safe deployment.

Summarising his current mindset, he continued: “I would say cautious optimism – if the best minds work towards that, I think we’ll solve the technical risks.”

While AGI has not yet reached human-level intelligence, Demis' remarks underscore both the promise of AI in augmenting human capability and the prudence needed to manage its risks responsibly.

He also highlighted the global potential of AI: “The generation that grows up native with that technology will end up doing some sort of incredible things that we can only dream of right now.”