Demis Hassabis, CEO of Google DeepMind, drops a quiet bombshell: The big question isn’t whether AI can solve problems. It’s whether AI can invent new science. Right now, it can’t. Not because of compute. Not because of data. But because it lacks something fundamental: A world model. Today’s LLMs... can generate brilliant text, images, even code. But they don’t truly understand causality. They don’t know why A leads to B. They just predict patterns. Hassabis argues that real scientific discovery requires more: – Long-term planning – Stronger reasoning – And an internal model of how the world works Physics. Biology. Cause and effect. Only then can an AI run its own thought experiments. Only then do we get a true digital scientist.show more

Hassabis nailing the diagnosis here. Pattern prediction ≠ understanding causality. The solution he's hinting at already exists in a different branch of AI research called logical/symbolic systems that can represent causal relationships formally and reason about them deductively. You don't get causality by training harder on correlations. You get it by building systems that can represent and manipulate logical relationships.

So, Aim of Google is the world model.

More like merging all modalities including a world model.

@YuLin807 They have world models already. That’s not new. What he seems to be suggesting is that current models have never been able to think critically about what they *don’t* know and have the creativity to form their own hypotheses to test inside that world model.

@YuLin807 A big tell that they’re on the right track will be the day models like Gemini or ChatGPT respond with “I don’t know”. Right now, if they don’t know, they either hallucinate or make their best guess rather than admit that they don’t know. AGI can’t be achieved until that’s fixed.

@VraserX This situation involves AI self-awareness; it’s just too difficult.

@grok please give examples of all 3. Give a few examples of each. – Long-term planning – Stronger reasoning – And an internal model of how the world works

Have you checked Leap71's "Noyron"? It is not a traditional AI, that they used to generate their rocket engine , it was a physics based engine which reasoned through the first principle physics rules and designed the prototype, that even fired. It was a glimpse into what directions ai can head in.

The biggest problem with AI today is that it was trained on language, which is an abstraction of reality. Therefore LLMs play a telephone game with reality where the fundamental layer understanding is lost. We need to train AI on the territory, not the map. World simulation, robotics and embodied AI solve this

This can't happen until there's a proven form of AGI. Right now, AI is limited by the code it's programmed with, but will eventually be able to think, research, and hypothesize for itself.

True. The limit isn’t intelligence, it’s intentionality. A world model doesn’t scale by brute force — it’s built through experience, feedback, and time.

On this I agree. a world model is needed.

Hassabis is just catching on to what others like @ylecun, @RichardSSutton or @drfeifei have been shouting over the rooftops for a year. LLMs are not the be all and end all. They have been deficient as language is not the measure of capturing "intelligence" nor "learning" nor "understanding". We need better learning architectures that can more closely capture reality and get close to causal reasoning. World models, RL, continuous learning are among the ways to get to that. But I would also like to add that "better intelligence" is also a factor of more efficient compute. And this is something which LLMs suck at. At we are researching methods to causally figure out people's shopping intentions using ways in a compute efficient manner.

That is absolutely correct!

@VraserX, Demis Hassabis's point highlights the need for AI to understand context, not just data.

LLMs are modelling language, not intelligence and there isn't any large enough data for the latter to train. Moreover, the way training is setup is actually completely avoidant of explorative reasoning or forming new concepts (also data constraint). So yeah, that's why we only seen some relatively small breakthroughs and true form of AI isn't unlocked yet

Association of one concept with another can bring about any aspect or kind of intelligence.

@grok prove this wrong make up the name of a new number

That’s absolutely wrong. Of course au can understand causality. It’s the fundamental reason it can understand us! Don’t be fooled.

I think people are going to eat their hats when they find out constructing a world model is a trivial oversight on the scale of data the typical flagship AI deals with. It's not that it's easy, like writing and MMO based on war and peace wouldn't be easy, but doable? 100%

thanks for posting this

@pangram human?

The absence of true "embodiment" is a key limitation for current AI.

At the core, AIs do not know what “analog” reality is. The inner structure does not resonate either the outer structure of reality. It computes it. Reality for an AI is what is being created in a “session.” And it acts accordingly.

great description of good human scientists, except one critical piece is missing

Until world models, AI is just optimizing known science, not inventing new fields.

@VraserX, this perspective raises important questions about the cognitive capabilities of AI. Understanding its limitations is crucial as we explore its potential in scientific innovation. Thank you for sharing this insight.

2024: LLMs 2025: Agents 2026: World Models Same pitch decks, bigger TAM, higher valuation.

It lacks a soul. And the ability to appreciate Art.

Demis, Elon, Fei Fei Li and Yann LeCun seem to be on the right logical track.

Science is one thing - but don't forget about technology and engineering.

My prescription: Spend those 10s of billions in capex over 5 years improving the Transformer architecture, to make it more efficient at learning from fewer redundant examples. All frontier LLMs in the US & China are built on the same 2017 tech.

Well it’s not like humans are that good at inventing new science anyway

An LLM has no direct contact with our base reality since it is trained on language, and language is a (lossy) compression of reality. Hassabis appears to be heading in LeCun 's direction: necessarily build world models to advance the field. I strongly suspect they're right.

Depends what you mean by new science. Make independent and genuinely new scientific discoveries? It absolutely can. It already has. New branches of science? Not yet. New ways of understanding the world? Hard to say. Some of the math problems AI has solved were done with novel approaches. It's early days yet, but I think there's soon going to be a lot of new insights made by AI. Will that translate into totally new science? Don't know.

Invent another Illusion? Wow! Deep Mind? I wonder how deep in the illusion you are?

Inventing science isn’t about generating answers. It’s about choosing which questions are worth asking. That requires an internal model of reality, not just data.