Demis Hassabis: Agents, AGI & The Next Big Scientific Breakthrough

Predictions and Expectations:

• Demis Hassabis expects "AGI appearing in the middle of" a deep tech journey that starts today, noting his personal AGI timeline is "like 2030 or something like this."
• He believes "continual learning, long-term reasoning, some aspects of memory... are still unsolved" and that "all of these are gonna be required for AGI."
• He states there is "a lot of room for innovation in areas like memory" despite large context windows.
• He predicts "one or two big ideas left that need to be cracked" for AGI, with his betting on this being "about 50-50."
• He expects that "a year later after one of our leading pro models or frontier models goes out, half a year, a year after, you'll have them in the really tiny, almost edge models" via distillation.
• He foresees that "agents are that path" to AGI and that "we're just getting going" in this area.
• He expects to see "the really valuable places" for agents to emerge "in the next six to 12 months."
• He believes systems currently lack the ability to "invent go" from a high-level description and that this capability is missing.
• He predicts that "inference will ever be essentially free" is unlikely due to physical bottlenecks like chip creation, even if energy costs drop.
• He expects that "we'll have some big announcements" from Isomorphic Labs regarding drug discovery "very soon."
• He predicts we are "about 10 years away probably from something like a virtual cell like a full virtual cell."
• He believes AI will become helpful in "no area of science or engineering" in the foreseeable future.
• He expects "AI will be the ultimate toolbox for science" to advance understanding in medicine and the universe.
• He believes that deep tech areas involving the "world of atoms" are "pretty safe from just getting swarmed by whatever the next update is to the foundation models."
• He predicts that "we'll have a lot to talk about in the next couple years on all those areas" including materials, mathematics, and science.
• He expects that eventually systems will be able to "come up with a new set of of millennium Price problems" and pass an "Einstein test" to generate novel hypotheses like special relativity.
• He believes that general systems like Gemini or Claude will make use of "AlphaFold-like specialized systems as tools" rather than having everything in one giant brain.

Timelines and Milestones:

• Demis Hassabis places his AGI timeline "like 2030 or something like this."
• He states a deep tech journey is "usually... a 10-year journey for true deep tech."
• He predicts the "next six to 12 months" will see the full value of agent technology delivering results like hit games.
• He estimates it may be "a couple of years out" from systems solving the Riemann hypothesis or similar millennium prize problems.
• He notes the AlphaFold project is celebrating its "10-year anniversary" following the AlphaGo breakthrough in Seoul.
• He predicts "Gemma4" models will be "really amazing power for their sizes."
• He foresees "millions of agents, swarms of agents working together" as a future usage pattern.
• He believes "energy costs will be essentially zero" if fusion, superconductors, or optimal batteries are solved "in the next few decades."

Technology and Product Direction:

• Demis Hassabis indicates that "agents are that path" to AGI and DeepMind works on "bunch of agentic systems."
• He describes the current state of continual learning as "cobbling it together with duct tape" and expects innovation to move beyond "shove it all in the context window."
• He plans to continue building on multimodal foundation models for "robotics" and "digital assistants that come with you into the real world."
• He states that "edge models" for Android, glasses, and robotics will be "open models" and fully open at the "nano size level."
• He expects the development of a "virtual nucleus" as a starting point before modeling a "full working simulation of a cell."
• He predicts that "Gemini Robotics" will be built on multimodal foundation models, leveraging a competitive advantage in understanding "intuitive physics and the physical context."
• He outlines a future where "general purpose tool usage models" train specific specialized tools in separate systems.
• He mentions "Genie" as a system built on top of Gemini for "world model building."

Market and Industry Outlook:

• Demis Hassabis believes there is a huge need for "small and even smaller models" to serve "billions of users, a dozen, more than a dozen billion user products" efficiently.
• He notes that Chinese models are "currently leading in open source" but aims to create "Western stacks."
• He predicts that almost every drug discovered from now on "will have used AlphaFOLD at some point in the drug discovery process."
• He expects "AI" to be the "ultimate toolbox for science" impacting all research fields.
• He believes that "interdisciplinary teams" combining AI with deep tech areas like materials or medicine represent the "sweet spot" for startups.
• He foresees a future where "millions of agents" will utilize available inference resources.
• He suggests that "inference costs" will be rationed for "the next few decades" due to physical chip bottlenecks.

Company Plans:

• Demis Hassabis states DeepMind and Google DeepMind "work on both" scaling existing techniques and cracking one or two big new ideas.
• He confirms the company is building "Gemini" and "pushing toward the same goal" of artificial general intelligence.
• He mentions Isomorphic Labs is "trying to build out... not just alpha fold" for adjacent biochemistry and chemistry.
• He details the plan to serve "the biggest probably AI surfaces" with "flash" and smaller models for speed and efficiency.
• He plans to "unify" the decision to make edge models fully open at the "nano size level."
• He indicates the company will "publish in... the big journals" and continue publishing science work.
• He notes "Gemma" has reached "40 million downloads" in "two and a half weeks."

Financial Guidance:

• (None reported)

Risks and Caveats:

• Demis Hassabis warns that "it's not bad necessarily, but you have take that [AGI appearing in the middle of a journey] into account" regarding deep tech timelines.
• He notes that "looking it up and finding the right thing that's actually relevant... is non-trivial, that cost" even with massive memory.
• He cautions that "I don't think we'd see autonomous first" and that "people in this room operating a thousand X" will come first.
• He observes that "we haven't seen a AAA game that tops the App Store charts that was sort vibe coded yet."
• He points out that "we don't have enough data yet" to model live cells at nanometer resolution without destruction.
• He cautions about the "parable of Prometheus" and the need to be "careful with how we use that and what we use it for and also the misuse that can happen."
• He notes that "nobody has enough spare compute" to make two frontier models at maximum size with different attributes.
• He suggests that systems currently "might be capable" of deep creativity if used by a "brilliant enough creative person" but cannot do it alone yet.

Confidence and Disagreement:

• Demis Hassabis is pretty sure that "the components that you just mentioned... will be part of the final architecture for AGI" and "can't see a world in which we will sort of realize in a couple of years, this was a dead end."
• He states "I don't think it's more than one or two" big ideas missing for AGI "if there are out there."
• He says "I don't think we've got to any kind of... limit" regarding distilling power into smaller models "for now."
• He is not sure "inference will ever be essentially free."
• He expresses uncertainty about whether a system will be able to invent a game like Go from scratch, stating "clearly today systems... can't do that."
• He states "I haven't seen anything yet" that constitutes "true genuine... massive discovery" in scientific reasoning beyond pattern matching.
• He believes there is "huge scope for combining where AI is going with some other deep technology area."
• He maintains a long-term conviction that AI is "the most consequential thing" he could work on, stating "I would still be working on AI today" regardless of past struggles.