Demis Hassabis on What Still Stands Between Today and AGI

Classy aggregates publicly available material; we did not conduct a private interview.

Watch the public conversation

Google DeepMind CEO Demis Hassabis spent forty minutes with Y Combinator CEO Garry Tan on a live episode of How to Build the Future, revisiting the same mission he set when he co founded DeepMind in 2010: build artificial general intelligence and use it to accelerate science.

The conversation is not a product launch. It is a map of what Hassabis thinks is still missing from the current large model stack, and why he believes agents are the practical bridge.

What Hassabis says is still unsolved

When Tan asked how much of AGI architecture already exists in today pretraining, reinforcement learning, and chain of thought stacks, Hassabis answered with a short list of gaps: continual learning, long term reasoning, aspects of memory, and consistency across long horizons.

He said all of those will be required for AGI. His personal timeline remains roughly 2030, which means any deep tech founder starting today must plan for AGI arriving mid journey.

That framing matters because it separates capability spikes from durable systems. Hassabis argued that passive models alone will not reach AGI. You need active systems that can pursue goals, recover from failure, and update behavior over time. Agents, in his view, are that path, and the industry is only at the beginning.

Memory, context, and the million token illusion

One of the clearest technical passages concerns memory. Hassabis noted that a million token context window sounds enormous until you translate it into live video or multi hour workflows. Naive tokenization can burn through context budgets quickly, which is one reason he treats memory as an unsolved systems problem rather than a marketing metric.

He also discussed distillation and smaller models, highlighting Gemma adoption as evidence that open, efficient models can spread faster than many observers expected. Multimodal design was another theme: Gemini was built multimodal from the start because the real world is not text only.

Science, AlphaFold, and the next breakthrough pattern

Hassabis returned repeatedly to science as the highest value application. He walked through the AlphaFold breakthrough pattern: identify a grand challenge, combine the right team, and release tools broadly when the social return is larger than any single product margin.

He pointed toward virtual cell modeling and materials discovery as areas where AI could produce genuine scientific discoveries rather than incremental summaries. Founders, he advised, should build before AGI arrives but assume AGI will reshape every roadmap.

On scaling versus new ideas

Near the end, Hassabis offered a calibrated bet: roughly fifty fifty between scaling current techniques and needing one or two major new ideas. That is not hype in either direction. It is a research leader acknowledging both the power of present architectures and the historical record that several breakthroughs still came from new concepts.

Tan pressed Hassabis on whether agents are overhyped. Hassabis answered that hype cycles are normal, but the underlying need for active problem solving systems is real. He compared today agent stacks to early mobile apps before the iPhone moment fully matured.

Sources

Y Combinator, Demis Hassabis public talk on agents and AGI (April 29, 2026)<br />PodLexicon transcript summary of the same episode (April 29, 2026)