The head of DeepMind reveals that AGI is coming in five years, with the computing power demand surging tenfold and reasoning calculations consuming everything.

Hassabis is definitely one of the smartest and most interesting minds in the world today.

In the latest podcast, he said, "Any pattern that can be discovered in nature can be efficiently learned and imitated by machine learning algorithms."

AlphaGo and AlphaFold are building a model for complex problems with an unimaginable number of possibilities. Proteins in our bodies can complete folding in just a few milliseconds.

Natural systems are structured. As long as something can evolve, it can be efficiently understood and imitated.

It's like nature is playing a game. The most amazing thing is that the things it creates in the game happen to be efficiently understandable through models.

Veo can simulate liquids and the reflections of various materials with surprisingly good results.

Hassabis used to work on physics engines in a game company and knows how torturous it is to write a program from scratch to achieve this.

AI seems to have figured out the laws of physics just by watching YouTube videos.

Rendering, lighting, etc. in video generation are the most core and fundamental aspects of physics. It is revealing some fundamental truths about the structure of the universe to us.

If we can understand how physics achieves this and then model this process, it should be feasible.

Building True AGI

He used to think that, as many neuroscience theories suggest, one needs to act in the world to truly perceive it deeply.

But now it seems that it can be understood through passive observation.

The next step might be to make these videos interactive, allowing people to truly step into the video and move around in it. That would be extremely shocking.

Hassabis believes that this is starting to approach a "world model," including the mechanics, physical laws of the world, and everything in it.

This is exactly what a true AGI system needs.

AGI "Video Games"

He often fantasizes about what kind of things could be created if the AI systems of today were available in the 1990s.

Hassabis says that you could definitely create mind - blowing and amazing games.

In the game, there is a simulated world with AI characters. Then the player interacts with this simulated world, and the world adjusts and adapts according to the player's gameplay.

He believes that it would be the most fun game, and everyone's gaming experience would be unique.

We set the parameters and initial conditions, and then the player immerses themselves in it and co - creates their story with this simulated world.

However, programming an open - world game is of course very difficult.

No matter which direction the player goes, it has the ability to create corresponding and engaging content.

Now, we may be on the cusp of a new era.

In the next few years, perhaps five to ten years, we will have AI systems that can truly create based on your imagination.

They can dynamically change the plot. No matter what you choose, they can tell a dramatic story around your choice.

Hassabis thinks this may be within reach.

Imagine how great the interactive version of Veo would be.

What players really want is that anything can happen in that game environment.

Video games may become a place where people seek meaning.

You can create extremely rich and meaningful experiences in the game, and even more diverse lifestyles.

On the other hand, it is also very important to enjoy and experience the physical world.

We will have to face this question again: What is the nature of reality?

What exactly is the difference between these increasingly realistic, multiplayer, emergent simulated worlds and what we do in the real world?

ASI Super Researchers

There are many interesting ideas about the steps towards ASI, including "super programmers" and "super AI researchers."

He mentioned a very interesting term, "research taste."

Whether AI can help outstanding human scientists judge which directions are likely to produce truly novel ideas seems to be an extremely important part of doing top - notch scientific research.

Hassabis believes that imitating or modeling "taste" and judgment will be one of the most difficult things.

This is exactly the key difference between great scientists and good scientists.

Proposing a truly good conjecture is much more difficult than proving it.

At the International Mathematical Olympiad, AlphaProof won a silver medal last year. Maybe we can finally solve a Millennium Prize Problem.

As long as the problem is posed correctly and the experiment is designed correctly, failure itself is extremely valuable.

The weather system is notoriously difficult to model, but Google DeepMind has also made progress.

DeepMind has created the world's best weather forecasting system, which is better than traditional systems based on fluid dynamics.

Traditional systems usually take several days to run on huge supercomputers to get the results.

Even though these dynamics are very complex and in some cases almost chaotic systems, they can still be modeled with the neural network WeatherNet.

In programming, AlphaEvolve makes recursive self - improvement possible.

The path to AGI may not be a straight line but a process of gradual improvement over time.

Scale Up the Computing Power!

How crucial is scaling up the computing power for building AGI?

Hassabis believes it is very crucial.

For training, the required computing resources are usually concentrated in one place. Even the bandwidth limitations between data centers can have an impact.

Since AI systems are now integrated into products and used by billions of people around the world, a huge amount of inference computing is required.

On top of that, there is a new paradigm that emerged in the past year. The longer the inference time it is given, the smarter it becomes during testing.

As AI systems get better, they become more useful, and the demand for them will also increase.

The computing power required for training is actually just a part of it, and it may even become the smaller part of the total computing volume required.

As Veo becomes more and more incredible, the servers are "breaking a sweat."

DeepMind has many interesting hardware innovations.

It has its own TPU product line and is also researching pure inference chips and how to make them more efficient.

They are also very interested in building AI to help solve energy problems, such as improving the efficiency of data center cooling systems and optimizing the power grid.

Ultimately, it can also help with plasma confinement in nuclear fusion reactors.

Then there is materials design, which he believes is one of the most exciting new fields, such as new solar materials, room - temperature superconductors, and optimal batteries.

The solution to any of these problems will bring an absolute revolution to climate and energy use.

DeepMind as a "Startup"

Any large company has many management levels and such things. This is the nature of its operation.

Hassabis can still, and has always, run DeepMind in the mode of a startup. Although it is not small in scale, it is still a startup.

DeepMind works with the determination and vitality of the best small organizations.

They are trying to have the best of both worlds: having an incredible product platform that reaches billions of users and empowering them with AI.

There are few places in the world where you can do this: do world - class and incredible research one day and connect it to products to improve the lives of billions of people the next day.

He once talked with Irving Finkel, an expert on cuneiform script at the British Museum. Finkel didn't know about ChatGPT or Gemini.

Finkel's first encounter with AI was the "AI mode" on Google Search. Many people in the world still don't know about AI.

When designing AI products, you can't just look at what the technology can do today, but what it can do in a year.

So, you have to be a product person with very deep technical skills, have a good intuition and feeling, and be able to "intercept" the future direction of this rapidly developing technology.

Hassabis believes that we will start to use other devices, such as smart glasses, audio earbuds, and even some kind of neural device, to increase the input and output bandwidth to, for example, a hundred times that of today.

Hassabis believes that we will enter an era of AI - generated interfaces.

These interfaces are likely to be customized for you personally, so they will fit your aesthetics, your feelings, and the way your brain works.

Reference materials:

https://www.youtube.com/watch?v=-HzgcbRXUK8

This article is from the WeChat official account "New Intelligence Yuan". Author: New Intelligence Yuan, Editor: Ying Zhi. Republished by 36Kr with permission.