Google DeepMind CEO: Is AGI just 1 - 2 breakthroughs away?

On December 4, 2025, the Axios AI+ Summit.

In the last conversation, across from Mike Allen, the co-founder of Axios, sat Demis Hassabis, the CEO of Google DeepMind and the winner of last year's Nobel Prize in Chemistry.

When asked how far we are from AGI, he didn't shy away:

We're only one or two AlphaGo-level technological breakthroughs away from AGI.

This statement shook the entire AI community because it was so specific. But this isn't a wild guess; it's an inference based on the current technological progress.

What's crucial is that he also laid bare the underlying logic of the AI competition: it's not about who burns the most money or has the most GPUs, but who can integrate research, engineering, and product development.

Section 1 | Is AGI near? What makes him say so?

Hassabis gave a time frame of 5 to 10 years to achieve AGI. This isn't inferred from the scale of model parameters but based on several very specific advancements:

1. The model has evolved from a text expert to a multimodal understanding system

What surprised him most about Gemini wasn't its ability to write code or generate text, but its ability to understand videos and discern the intentions behind actions.

When he tested Gemini, he played a clip from the movie "Fight Club," where the protagonist took off his ring before a fight. He casually asked:

What's the significance of this action?

Gemini didn't just describe the surface action but provided a profound interpretation: it's a symbol of discarding one's identity and breaking free from rules, representing the character's transformation from reality to extremism.

Hassabis's evaluation was: shocked.

Because this has gone beyond pattern recognition and started to exhibit insight.

2. The model has its own judgment and no longer just caters

Hassabis specifically mentioned that Gemini achieved a small but significant feat. When you say, "You're wrong," it won't blindly cater but will gently refute you.

This isn't just about judging right or wrong. While the model understands the context and corrects biases, it can also control its tone and balance its expression.

Hassabis said he likes Gemini's personality: concise, calm, confident, and not overly eager to please.

This means the model is starting to break away from the positioning of a chat tool and move closer to a system with a stable personality.

3. From generating content to building products

Gemini can now accomplish two things:

Generate playable mini-games with one click: Previously, it took weeks to develop, but now a complete prototype can be generated in a few hours.

Generate front-end web pages: It doesn't just write code but also knows how to make it look good and be user-friendly.

Behind this is the model's overall understanding of code structure, design logic, and interaction experience. It's not just about generating text but creating things that can be directly used.

Current models have started to possess understanding, judgment, and creative abilities.

So he said: We're very close; we're just one or two AlphaGo-level technological breakthroughs away.

Section 2 | But it's not AGI yet: What's the gap?

While these capabilities are making breakthroughs, he also admitted: We haven't reached AGI yet, and current models still have obvious shortcomings.

If the previous section talked about what it can do, this section is about what it can't do.

1. Lack of continuous learning ability

Current large models can't learn continuously. They can only learn once with the data from training.

This means: It won't grow during use, won't become smarter through interaction with users, and can't correct mistakes with experience like humans.

Online learning and long-term memory systems are one of the key breakthroughs towards AGI.

2. Inability to execute long-term plans

Although models are getting stronger in single-round reasoning and code generation, they can't execute long-chain reasoning or formulate and execute long-term goals.

This isn't due to a lack of ability but because the underlying structure isn't designed as a multi-step decision-making system. AGI needs to be able to complete scientific research tasks spanning several days, adjust strategies in a timely manner, and achieve goals in stages.

Current models are more like instant response systems rather than goal-driven systems.

3. The agent system is still unstable

When talking about Gemini's goal of being a general assistant, he emphasized: We still can't entrust an entire task to them and be sure they'll complete it.

Current agents can't reliably execute multi-step tasks in complex environments.

Whether agents can operate stably determines whether AGI can be implemented in real scenarios.

4. Lack of stable cross-dialogue memory

Although Gemini has a "personality," it only remains consistent within a single conversation and isn't a true agent with personalized memory.

True AGI should be able to:

Maintain a consistent stance (no self-contradiction)

Remember users' long-term preferences (coordinate memory and decision-making)

Adjust behavior based on context rather than starting over

This requires more fundamental architectural design, such as long-term memory networks and goal retention mechanisms, rather than just optimizing prompts.

So, where will these breakthroughs come from?

Section 3 | Where will the key breakthroughs come from? He pointed out two directions

Hassabis didn't directly list what those one or two AlphaGo-level breakthroughs specifically are, but he pointed out two clear paths in the conversation.

This is also the direction he's currently pushing forward.

1. World models: Enable AI to understand the operating laws of the physical world

DeepMind has a world model called Genie. You can generate a video and then move around in it like in a game, maintaining coherence for a minute.

The significance of Genie isn't just about generating videos but about building a virtual world with physical consistency and spatio-temporal coherence. It doesn't rely on data stacking but simulates the operating logic of the world.

This means AI is:

Jumping from understanding pictures to understanding physical laws,

Crossing from processing text to inferring the real world.

A multimodal world model that can construct and understand simulated environments is the foundation of AGI.

2. Agent systems: From answering questions to completing tasks

Although current models can chat and write code, we still can't entrust an entire task to them with confidence and be sure they'll complete it from start to finish.

Gemini's vision of being a general assistant aims to solve this problem, which is the next-generation agent:

Embedded in glasses, daily life scenarios, and workflows,

Capable of memory, reasoning, and task delegation,

Not just answering questions but achieving goals.

If a stable agent system that can autonomously plan, execute tasks, and adjust based on results can be built, AGI won't be far away.

This is why he emphasized that simply expanding the scale of LLMs won't bring about AGI. The real breakthroughs come from technological innovations in world modeling and agent systems.

Section 4 | Variables: Risks and competition

However, the 5 - 10 year prediction has a premise: AGI won't be achieved just because the time is up; there are many variables along the way.

The biggest uncertainty comes from technological risks and geopolitical competition.

1. Technological risks and protective measures

The host asked a series of pointed questions: Will malicious actors use AI to synthesize pathogens? Attack infrastructure? Will agents get out of control?

Hassabis's answer was:

"These risks exist, and the probability isn't zero, but no one knows the exact figure. The only thing we can do is take them seriously and prepare in advance."

He outlined three main risks:

Malicious use: Bad actors use AI to do bad things, especially hackers and organized crime need to be guarded against.

Agents going astray: The more complex AI is, the more likely it is to deviate from instructions, but it hasn't reached the point of complete loss of control.

Failure of security mechanisms: AI may gradually deviate from the security boundary during continuous learning.

But he also pointed out that the market itself will form protective measures: Enterprises won't buy unsafe agents. Enterprise customers will require you to prove the reliability of AI, and they'll switch suppliers once there's a problem. Business logic will reward more responsible AI companies.

The core of AI security competition is trust. Whoever can make customers trust them will survive.

2. China - US competition: The leading window is only a few months

When talking about the geopolitical technology landscape, Hassabis said:

The West currently still leads in terms of algorithms and innovation, but China isn't far behind.

He specifically mentioned China's Qwen and DeepSeek models. They have strong technical capabilities and are progressing rapidly, especially keeping up closely in terms of execution, iteration speed, and model scale.

The gap is a few months, not years.

The competition window for AGI is shrinking. China and the US aren't first and second but are advancing side by side on two parallel tracks.

Security is the premise, and speed is the key. Whoever can run faster and do more comprehensively while ensuring security will lead in the competition.

Section 5 | Competitive advantage: The real moat is the scientific method

Technological progress can be caught up with, GPU resources can be bought, and engineering teams can be assembled.

But for Hassabis, what determines who can truly reach the end of AGI isn't what you can do but how you do it.

1. The scientific method is a tool at the civilization level

The host asked him why he always emphasizes that he's a scientist first.

He said:

"The scientific method may be the most powerful tool in human history. It gave birth to the Enlightenment and modern civilization."

The same goes for developing AI. It's not just simple technological evolution but using the scientific method to gradually approach the essence of human intelligence.

2. Don't blindly believe in a single route; conduct comprehensive trial - and - error

DeepMind didn't blindly believe in LLMs in the early days. They simultaneously explored reinforcement learning (AlphaGo, AlphaZero), cognitive architectures, and neuroscience modeling and discovered the widely used Chinchilla scaling law.

When the outside world was still arguing about whether reinforcement learning or Transformer was better, DeepMind's strategy was:

Comprehensively explore,

Strictly track data performance,

Keep multiple possibilities and continuously verify.

This is DeepMind's greatest non - technical asset: scientific decision - making ability.

3. The real advantage is turning unknown problems into usable products

When asked about DeepMind's advantage, Hassabis summarized it like this: We integrate world - class research, engineering, and infrastructure, which is our unique ability.

What he emphasized isn't the scale of resources but the cooperation of the three:

Research has original capabilities and can be verified in practice

Engineering isn't just about writing code but constructing complete solutions

Infrastructure isn't about burning money but monetizing technology in the right way

Overall, DeepMind's real moat is how to transform a never - solved problem into a truly usable product.

Conclusion | The time window is closing

Hassabis gave a timeline: 5 - 10 years, 1 - 2 breakthroughs.

This isn't a prediction but an engineering judgment based on technological progress. The competition window is closing rapidly, and there isn't much time left for everyone.

Every choice you make today will determine your future position.

📮 Original link:

https://www.youtube.com/watch?v=tDSDR7QILLg&t=19s

https://www.axios.com/2025/12/05/ai-deepmind-hassabis-gemini

This article is from the WeChat official account "AI Deep Researcher", author: AI Deep Researcher, published by 36Kr with authorization.