Demis Hassabis: Why LLMs Will Not Commoditize & Why We Have Not Hit Scaling Laws

Why does it matter? Because the AGI timeline Demis Hassabis set in 2010 is still on track — and almost nobody believed it then.

Hassabis has been running the same calculation since DeepMind's founding, when AI was a dead end and virtually no one was paying attention. The conclusion then: AGI in roughly 20 years. The conclusion now: we're on schedule. That's an unusual kind of credibility — a prediction made before the hype cycle, surviving through it.

• Scaling laws haven't plateaued — returns are still substantial, just no longer exponential, and that distinction reshapes the competitive race
• Current AI systems are "jagged intelligences" — superhuman one moment, failing elementary tasks when the question is rephrased slightly
• AGI is likely within 5 years, a timeline unchanged since 2010
• AI is simultaneously overhyped in the next 12 months and still massively underestimated on a 10-year horizon — both are true right now

Scaling laws aren't dead — but the next phase of the race belongs to labs that can invent, not just spend

The widespread narrative that AI progress is hitting a wall is wrong, according to Hassabis — just imprecise. "No, I don't think so. I think it's a bit more nuanced than that." The early days of large language model development produced near-doubling performance gains with each generation. That exponential can't continue forever. But slowing from exponential doesn't mean diminishing returns.

"The returns are kind of still very substantial, although they're a bit less than they were obviously at the start of all of this scaling."

What changes in the next phase is where advantage accumulates. Squeezing gains from existing architectural ideas gets progressively harder. "Those labs that have capability to invent new algorithmic ideas are going to start having bigger advantage over the next few years as the last set of ideas are sort of all the juices being rung out of them."

This matters structurally. Compute has been the great equalizer — more resources, more performance. That dynamic fades as existing ideas reach saturation. Research depth starts to compound in ways that capital alone can't replicate. Hassabis estimates 90% of the breakthroughs underpinning the modern AI industry came from Google Brain, Google Research, or DeepMind. His bet is that the labs with the deepest research benches pull further ahead from here, not the ones with the biggest clusters.

"Jagged intelligence" is not a quirk — it's the core unsolved problem standing between today's AI and AGI

Here's the failure mode anyone who has used AI agents has hit: configure a workflow carefully, then watch it collapse when a single input changes slightly. Hassabis has a name for it. "I sometimes call these systems jagged intelligences because they're really amazing at certain things when you pose the question in a certain way but if you pose a question in a slightly different way they can actually still fail at quite elementary things."

The interviewer finished his sentence: "That's a disaster." Hassabis agreed without hedging.

"A general intelligence shouldn't be that sort of jagged."

This isn't a UX problem or a prompting problem. It's a structural gap between what current systems are and what AGI would require. True general intelligence — modeled on the brain, the only existence proof of general intelligence we have — doesn't have holes that open up when a question is rephrased. Every leading lab is sitting with this unsolved. Builders deploying AI agents for autonomous tasks should treat this as the binding constraint: these are brittle tools that require human oversight not as a temporary precaution, but because the consistency problem is genuinely open at the frontier.

AGI within 5 years: a forecast made in 2010 that nobody believed and is now tracking accurately

In 2010, DeepMind co-founder Shane Legg was publishing blog posts predicting AGI timelines. AI was considered a dead end. Almost no one was working on it seriously. The prediction: roughly 20 years.

"We predicted around 20 years it would take from when we started out and I think we're pretty much on track."

That puts the window at 2030, roughly aligned with Hassabis's current probability distribution: "There's a very good chance of it being within the next 5 years. So that's not long at all."

The credibility here is unusual. Most AGI timeline estimates are made in the middle of hype cycles, shaped by whatever the current moment feels like. This one was set before transformers, before GPT, before anyone outside a small research community cared. It has survived the full arc of the field's development without revision. Hassabis frames the magnitude in stark terms: "10 times the industrial revolution at 10 times the speed. So unfolding over a decade instead of a century." The industrial revolution eliminated child mortality of 40% and created modern medicine — and it took a hundred years. The compressed version arrives within a single career phase.

AI is overhyped for 2025 and still underestimated for 2035 — both simultaneously

"Literally today as of today and in the next year things are a bit overhyped in AI." Coming from the CEO of Google DeepMind, that's worth sitting with.

But Hassabis holds both things at once. The short-term expectations are inflated — the gap between what current systems can reliably do and what the hype implies is real. At the same time: "I still think it's still very underappreciated how revolutionary this is going to be in the sort of time scale of about 10 years."

"So there's still that dichotomy even today with AI."

The practical read: anyone calibrating plans on 12-month AI capability expectations is likely to be disappointed. Anyone calibrating on 10-year expectations is still thinking too small. These aren't competing claims — they describe two different time horizons with different dynamics. Short-term, the system is in hype overshoot. Long-term, the transformation is still not being taken seriously enough by most institutions.

The missing ingredients for AGI aren't more compute — they're memory, continual learning, and long-horizon planning

Hassabis is specific about the gaps. Current context windows are "a bit brute force — you just put everything in them." Real memory architecture, the kind the brain uses, is more selective and more durable. There's significant architectural work left to invent there.

Continual learning is another open problem. Today's systems don't learn after training ends. The brain handles this through sleep — memory consolidation, where experiences from the day are replayed and selectively incorporated into existing knowledge. "Perhaps we need something like that to incorporate new information along with the existing information base."

Long-horizon planning is the third gap. "These systems are not very good at planning at long time horizons, you know, many years into the future — which we as you know with our minds we can do."

Taken together, these aren't incremental improvements — they're the research directions that define what comes after the current generation of models. Watch for neurologically-inspired architectures in memory and learning systems specifically. That's where Hassabis is pointing, and it signals where genuine algorithmic invention, not compute scaling, will determine the next leap.

LLMs won't be replaced — AGI will be built on top of them

Yann LeCun has been vocal about the limits of large language models as a path to general intelligence. Hassabis disagrees, but carefully. "I think there might be a 50/50 chance there's some things maybe missing that we still need to make breakthroughs in — perhaps world models."

But the frame he rejects is replacement. "I don't think it's going to get replaced. I think it's going to get built on top of these foundation models just like the way we do with our world models."

"The only question really is when you think about a future AGI system is — is an LLM foundation model going to be the key component only, or is it the total system?"

The distinction matters for how to think about bets in the space. LLMs as a platform are not going away. What gets added above them — world models, richer memory systems, long-horizon planning components — is where the meaningful architectural questions live. The post-LLM debate is being framed as replacement versus survival. Hassabis says that's the wrong frame entirely.

AI will likely solve its own energy problem — by inventing the breakthroughs needed to power itself

The energy concern around AI assumes current consumption patterns persist indefinitely. Hassabis's counter: AI will generate the scientific progress that eliminates the constraint. Grid optimization alone could yield 30–40% efficiency gains. Beyond that, AI-accelerated fusion research, new battery chemistries, and superconductor breakthroughs could fundamentally change the energy situation. "AI will in the medium to long run more than pay for itself in terms of energy costs." The near-term consumption problem may be self-correcting — not through restraint, but through the breakthroughs AI makes possible along the way.

The questions nobody is asking yet are philosophical, not economic

Every serious AGI conversation eventually lands on jobs, wealth concentration, and regulation. Hassabis thinks those are the easy problems. The harder ones haven't really started yet: what is meaning when machines can do most cognitive work? What is purpose? What does it mean to be human?

"I worry a lot about the philosophical questions around it... we'll find out what consciousness is. What does it mean to be human?"

The technical and economic challenges are hard. But a civilization that solves both and skips the philosophical reckoning arrives somewhere it isn't prepared for. The field needs great philosophers as much as it needs great engineers. That work hasn't begun in earnest — and AGI may not wait for it.

---

Topics: AGI, scaling laws, deep learning, AI safety, drug discovery, energy, labor displacement, LLMs, DeepMind, Google, research, regulation