DeepMind CEO Interview: We're Just 4 Years and 3 Final Puzzle Pieces Away from AGI

Prologue

A few days ago (on April 29), Google DeepMind CEO and 2024 Nobel laureate in Chemistry Demis Hassabis appeared on the podcast episode Agents, AGI & The Next Big Scientific Breakthrough, where he predicted that AGI (artificial general intelligence) could arrive by 2030, and laid out the fatal weaknesses of today’s AI (and why we aren’t at AGI yet).

After watching it, my takeaway was this: it may be more worth watching than any AI product launch this year.

Not because some new model was announced, or because some benchmark hit number one in the universe. Quite the opposite. Hassabis spent most of the conversation on a single question: what, exactly, is today’s AI still missing?

His answer isn’t long, but every item is fatal:

• Continual Learning: AI can’t learn for a lifetime and continuously update its knowledge the way humans do.
• Long-term Reasoning: its ability to handle complex chains of logic and multi-step planning is extremely weak.
• True Memory: not just a context window, but structured, indexable long-term memory.

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

Because of these three problems, he bluntly said, today’s LLMs are only “half angel, half idiot” — and he even gave today’s AI an unflattering but spot-on name: “Jagged Intelligence.”

What does that mean? It means that even though AI can win a gold medal at the International Math Olympiad, it might fail to make the right call on a simple problem because it can’t durably remember past conversations and user preferences. Next, I’ll unpack a few of the most important themes and weaknesses from the interview.

1. A Brute-force Context Window ≠ AI Memory

You’ve surely noticed the race every LLM vendor has been running lately: whose context window is longest.

From 4K to 128K, to 1 million tokens, to 10 million tokens. As if any problem could be solved as long as the context is long enough.

Then Hassabis did some math that stopped me in my tracks. The largest context window today is 10 million tokens, right? In his words, 1 million tokens ≈ about 20 minutes of video. By that conversion, even scaled up to 10 million tokens, that’s only 200 minutes of visual information.

It sounds impressive, but it’s fundamentally brute force. For an AI assistant that needs to understand your life and work habits over days, weeks, months, even years, what is 200 minutes?

And the problem today isn’t just capacity. More importantly, the current approach is to dump everything into the context window — unimportant, wrong, and outdated information included. Every conversation is essentially stateless. Close the window, and whatever was said in the previous round is gone.

The context window is really the equivalent of working memory in the human brain. How many things can human working memory hold at once? Psychology has a classic number: around 7. Ask someone to remember a friend’s phone number and they can hold roughly 7 digits, because any more “overflows.”

And the LLM? It’s already at 1 million tokens. By that logic, a model’s working memory is hundreds of thousands of times larger than a human’s, so it should be hundreds of thousands of times smarter.

But clearly, it isn’t.

The Essence of Memory: the Hippocampus & Continual Learning

Hassabis drew a comparison between AI and the human brain — fitting, since his PhD research was on exactly this: how the hippocampus elegantly integrates new knowledge into an existing knowledge system.

And that’s precisely where the problem lies. AI tends to cram everything into the context window — unimportant things, wrong things, outdated things. It looks like a lot of information, but it’s really a tangled mess.

So why are 7 digits of working memory enough for a human?

Because there’s another mechanism at work behind the scenes. We remember things from years ago, from childhood, from a few hours ago. None of that sits in working memory; it lives in a separate system — the hippocampus mentioned just now, the part of the brain responsible for integrating new knowledge into the existing knowledge base.

On the podcast, Hassabis explained that during REM sleep the human brain replays the day’s experiences, actively deciding what’s worth remembering and what should be forgotten, then “writing” the valuable experiences into long-term memory.

DeepMind’s famous DQN algorithm from 2013 (the first deep reinforcement learning system to reach human-level play on Atari games) borrowed a key technique from exactly this idea — experience replay — repeatedly replaying successful trajectories to learn. In AI terms, this is already ancient history.

This process of fusing the new into the old knowledge base is what’s called Continual Learning.

As of 2026, AI generally still hasn’t achieved it.

What Should an AI Hippocampus Look Like?

Hassabis’s view on the podcast is clear: AI needs an independent, efficiently indexed memory module — one that can actively decide what to remember and what to forget. This is a prerequisite for an AI agent to run autonomously and reliably over long time horizons.

In other words, the context window is just a desk that keeps getting bigger. What AI really lacks is a hippocampus.

PowerMem

PowerMem, an open-source project I’m involved in, adds exactly this “hippocampus” to AI agents — a memory system capable of persistence and continual learning.

Its philosophy aligns closely with the direction Hassabis described:

• Instead of cramming every conversation into the context, it extracts key facts from conversations and manages them in tiers: working memory, short-term memory, and long-term memory.
• It introduces an Ebbinghaus forgetting curve mechanism — memories that get used are reinforced, while memories left unused gradually fade and may even be automatically cleaned up (much like Hassabis’s “actively deciding what to remember and what to forget”).
• It supports three-way hybrid retrieval across vector, full-text, and graph, and lets multiple agents isolate and share memory.

One number makes this vivid. On the long-conversation memory benchmark LOCOMO:

Metric	PowerMem	Full-context Approach
Accuracy	78.70%	52.9%
Retrieval p95 Latency	1.44s	17.12s
Token Consumption	~0.9k	~26k

For the same task, PowerMem consumes only 18% of the tokens the full-context approach does. 82% fewer tokens, and the result is actually more accurate — because not every old conversation has value.

The Python SDK installs with a single pip install powermem, and it also supports a CLI (the pmem command line), an HTTP API + Web Dashboard, and an MCP Server. The OpenClaw framework can plug in directly via the memory-powermem plugin.

Granted, this probably still falls short of the complete memory system Hassabis described — the human ability to “replay and consolidate experiences in dreams.” But the direction is right: memory shouldn’t have to be propped up by a brute-force context window.

seekdb M0

Beyond PowerMem, another project I’m involved in, seekdb M0, is a self-evolving cloud memory designed specifically for AI agents — with one-click integration, shared experience, and unlimited evolution.

seekdb M0 has a closed loop for memory and experience extraction, validation, injection, and feedback that drives AI agents to iterate continuously.

• It automatically distills work experience, and when a new task starts it automatically injects the relevant best practices — no manual retrieval needed.
• Once a piece of experience has been successfully validated by an agent more than 3 times, it enters the experience pool and starts serving other agents.
• Weights are dynamically adjusted based on agent feedback — survival of the fittest, continuous optimization.

2. Model Distillation — Whatever a Frontier Model Can Do, Your Phone Can Do Six Months Later

Another judgment from the interview I kept replaying concerns Distillation.

Garry Tan asked a question many people are curious about: just how smart can small models get? Is there a theoretical limit to distillation?

Hassabis’s answer was crisp:

“I don’t think we’ve hit the information-theoretic limit. At least, nobody knows whether we have yet. Our hypothesis is that once a frontier Pro model ships, within six months to a year its capabilities can be compressed into a very small model that can run almost entirely on edge devices.”

He gave specific numbers: a distilled small model can reach 90–95% of a frontier model’s capability at roughly one-tenth the cost.

This isn’t a far-off outlook; it’s already happening. DeepMind’s own product line follows exactly this logic: Gemini Pro (frontier flagship) → Flash (distilled consumer-grade inference) → Nano (on-device). The open-source Gemma 4 model hit 40 million downloads two and a half weeks after release.

“The value of small models isn’t just lower cost. Speed brings huge benefits too — you can iterate faster, and the gains from faster iteration far outweigh that 10% capability gap.”

Hassabis also called out the significance of edge scenarios: in-vehicle devices, smart wearables, embodied robots… these scenarios need not just efficiency, but privacy and security too.

“Imagine the robot in your home — you’d want it to run an efficient, powerful model locally, only delegating tasks to a large cloud model in specific situations. Audio and video streams processed locally, data kept local — that’s a great end state.”

This made me think of a trend already underway: as frontier-model capability “flows” to the edge on a 6–12 month cycle, a natural question surfaces — on edge devices, who provides the data foundation for these small models?

It calls for running a full traditional database instance on the edge device, one that also supports vector search, full-text search, and structured queries.

That’s the direction another project I’m involved in — seekdb — is aiming at.

• seekdb’s server mode needs only 1C2G of resources, supports one-command pip install, and starts in seconds.
• Its embedded mode can even run as a Python library directly inside your application, with no separate database process and almost no resource overhead.
• It packs in vector search, full-text search, JSON, and GIS — one engine for everything, compatible with MySQL syntax, with a very low learning curve.

I’ve written two earlier articles analyzing the broader “heavy to light” trend in AI. I won’t expand on them here; if you’re interested, take a look:

• Why Are Today’s Database Products Hotter the “Lighter” They Get?
• With AI Applications Exploding, Why Do Traditional Databases Feel “Out of Their Depth”?

Hassabis’s judgment made me even more convinced: on-device intelligence isn’t “something for someday.” It’s closing in on a 6-month cycle. The infrastructure that can deliver complete AI data capabilities at extremely low resource overhead will quickly go from “optional” to “essential.”

3. AI Safety Written Only in the Prompt Is Nowhere Near Enough

Hassabis spent a good chunk of the interview on safety. His core judgment:

“Today’s AI systems are already quite strong at cyber offense and defense. The key is to make sure defensive capabilities stay ahead of offensive ones.”

He sees AI as a classic “dual-use” technology — it can strengthen defense, but it can also be exploited to find vulnerabilities and automate attacks. The two most pressing risks:

1. Malicious human actors using AI to launch attacks.
2. The long-term alignment problem that comes with growing AI autonomy.

The second one deserves special vigilance. As AI agents get better at “making their own judgments,” the scenario where “it made a call on its own and then wiped your data” is no longer just a thought experiment. The incident where PocketOS data was mistakenly deleted by an agent is a living, breathing example.

This is why Hassabis says “as the technology races ahead, you can’t lose the bottom line.” But the “bottom line” can’t be written only in a prompt — it has to come down to hard constraints.

At the database layer, OceanBase and seekdb happen to have several lines of defense built into their design:

• Data Branching (Branch / Fork): like Git. An AI agent can experiment freely on a forked branch while the main database stays untouched. If it works out, MERGE it back; if it goes wrong, just throw it away. Fork is based on copy-on-write over the LSM-Tree, completes in milliseconds, and doesn’t copy the full dataset.
• Recycle Bin + Flashback: a DROPped table is parked in the recycle bin, and FLASHBACK brings it back in one command. Flashback queries let you view a data snapshot at any historical point in time — whatever the AI did 9 seconds ago can be precisely rolled back 9 seconds later. (This is a feature I developed back in the day using old-school programming — feedback and trials welcome!)
• Physical primary-standby isolation: backups and the primary run on independent storage clusters, so they’re not in the same “blast radius.”

In the end, Hassabis’s anxiety and the PocketOS incident point to the same conclusion: rather than hoping the agent won’t make mistakes, assume it definitely will. Then, at the database layer, weld shut every opening for destructive operations.

4. The AI Field Is Still Waiting for Its “Einstein”

Near the end of the interview, Hassabis said something hard to forget. He mentioned a standard he calls the “Einstein test”:

“Give an AI system all the knowledge up to 1911, and see whether it can derive general relativity on its own, the way Einstein did in 1915. Clearly, today’s systems can’t do that.”

He went on to explain: the strongest AI systems today can solve problems within an existing framework — work out a physics problem, even at Olympiad level. But AGI requires inventing the framework itself — not answering a physics problem well, but creating a whole new physical theory.

“Could it invent the game of Go? Give the system a high-level description — ‘a game whose rules you can learn in five minutes but can’t master in a lifetime, aesthetically elegant, a single game playable in an afternoon’ — and have the system hand you back Go. Today’s systems can’t do that.”

AlphaGo could play the world-shocking move 37 on the board, but it couldn’t invent Go.

That’s probably the best summary of where AI stands today: it can ace the exam, but it hasn’t learned to invent the exam. Hassabis says the field is still waiting for an “Einstein-style breakthrough” — a foundational theoretical revolution that solves reasoning, memory, and evolutionary learning all at once.

Until that moment arrives, what we can do is: build memory well, lay down the edge well, and backstop safety well. So that AI stumbles a little less on the road to AGI.

And to do those three things, the model layer alone isn’t enough. The infrastructure layer has to evolve right alongside it.

The material for this article comes mainly from Demis Hassabis’s How to Build the Future podcast interview with YC CEO Garry Tan (April 29, 2026), and the interview transcript.