Demis Hassabis: The gap in AI competition will widen.

The real battlefield of the AI competition lies beyond what you can see.

On April 7, 2026, Demis Hassabis, the CEO of DeepMind, turned his attention to products like AlphaFold, Isomorphic, and AlphaTensor in a recent conversation. These products rarely make it to the top of tech headlines.

Thanks to these products, protein structure prediction, which used to take years, can now be done in just a few seconds. Most of the screening work in drug design can be completed in a virtual computing environment, and even new solutions that humans have never imagined are emerging.

However, these fundamental technological breakthroughs are accelerating towards a few giants. Google has contributed 90% of the fundamental breakthroughs in the modern AI industry.

As tools become more widespread, the gap is widening. The more people use the same tools, the greater the gap becomes.

Section 1 | The Gap Is No Longer Where You Can See

If the current AI ecosystem is roughly divided into two layers, the majority of people's understanding remains at the first layer.

The first layer is the visible application layer: large - model conversations, AI writing, image generation, and AI search. They have lower thresholds, more intuitive feedback, and are more likely to become topics of public discussion. Naturally, they form the main reference for people to measure the strength of AI.

But in this conversation, Demis Hassabis emphasized the deeper second layer.

He gave the example of AlphaFold (protein folding). This was an ultimate problem that had puzzled the scientific community for decades: how to accurately infer the three - dimensional structure of a protein based on just a string of amino acid sequences? In the past, this often took several years, with high costs and a high failure rate.

Models like AlphaFold have compressed the long trial - and - error process into just a few seconds and have directly open - sourced the prediction results to global researchers. Now, more than 3 million scientists around the world are using it, and the structures of 200 million known proteins have been successfully predicted.

As a scientist in the pharmaceutical industry sighed to Hassabis: From now on, AlphaFold will surely be involved in the R & D pipeline of almost every new drug.

The core change here is that the absolute starting point of human scientific research has been raised as a whole.

In the past, scientists had to spend a lot of time confirming basic structures. Now, they can directly move on to more critical issues: drug design, disease mechanism research, and improvement of climate - adaptable crops. These changes won't appear on any hot - search lists, but they are changing the speed of scientific research and re - determining who is more likely to make breakthroughs.

Similar situations are also emerging in other fields:

In the energy system, AI optimizes the operation of the power grid, increasing efficiency by 30% to 40%.

In materials science, AI exhaustively searches for new alloy combinations.

In drug R & D, it screens and designs compounds.

Many processes that originally required a large number of repeated experiments can now be mostly pre - screened in a virtual computing environment, leaving only the final step for experimental verification.

The same tools produce completely different effects:

"Some people are just using AI to improve the execution efficiency of existing tasks, while others are using AI to re - define the problems themselves."

The gap is no longer just about speed.

Section 2 | The Moment When the Gap Truly Widens

If the previous section talked about the direction, what truly widens the gap is the change in AI capabilities themselves.

The decisive turning point occurred on the Go board.

The number of legal Go games is as high as 10 to the power of 170, more than the total number of atoms in the known universe. In the past, the academic community generally believed that it would take at least several decades for computers to defeat humans in this field.

But in 2016, AlphaGo played the 37th move against Lee Sedol.

At first, that move was judged as wrong, even absurd, by all professional Go players. But as the game progressed, people realized that it was a playing style that had never appeared in the history of human Go.

This is completely different from traditional computer programs. In the past, human experience was written into the program for execution, while AlphaGo's experience comes from its own trial - and - error and exploration.

Subsequently, this ability was pushed to the extreme.

AlphaZero, which abandoned all human Go records, started true "learning from scratch." Hassabis witnessed AlphaZero's amazing evolution in one day: in the morning, it was making random moves; by noon, it could play against him; in the afternoon, it surpassed grandmasters; and by evening, it had crushed the world champion.

With AlphaTensor, AI even began to look for more efficient methods at the algorithm level. It found a faster matrix multiplication, which is the basic operation of all neural networks.

AI has started to discover new knowledge on its own.

When this happens, the meaning of "gap" is completely different. If the opponent is just faster and more accurate, AI companies can still make up for the ability gap with time. But if the model takes a new path, the past ways of catching up won't work.

As the Scaling Law of large models approaches its limit, simply piling up computing power and parameters faces diminishing returns. At this time, whoever has this exploration ability builds a new barrier. Whoever can make AI invent new algorithms will gain an advantage in the next round of competition.

Because the dividends of the previous round have been fully exploited.

Section 3 | Why the Gap Will Continue to Widen

If the first two sections clarified the direction and capabilities of AI evolution, then what we need to face next is a more harsh reality: in an era when tools are almost equally accessible, why is the gap between individuals and companies accelerating?

This gap first lies at the technical foundation.

90% of the fundamental breakthroughs supporting the modern AI industry come from Google Brain, Google Research, or DeepMind. This is the result of long - term R & D accumulation.

Even in the open - source model, there is a time lag. It takes six months for new ideas from leading laboratories to be replicated in the open - source community. In a rapidly iterating technological environment, this half - year itself is a barrier.

Secondly, there is a differentiation in the degree of tool use.

Demis gave a suggestion: Immerse yourself in these tools until you feel like you have superpowers.

On the surface, this statement is about the learning method, but in fact, it points to the fact that the same tools are being used for completely different purposes.

Some people use AI as an efficiency tool to write content and organize information, just making the original processes faster.

Some people use it as an ability amplifier to complete tasks that were previously impossible, such as enabling non - technical personnel to build product prototypes and analyze complex data, making things better.

Some people start to use AI to re - define the problems themselves, allowing it to directly participate in scientific research, design new product paths, and even change the original work methods.

The first two are speeding up, while the third is changing the direction.

When Hassabis described the future, he particularly emphasized the form of Agent. Agent means that AI will evolve from a tool that "passively executes instructions" to a digital employee that "independently promotes complex goals." From setting goals, planning and decomposition to path correction, everything can operate on its own.

Once this form is fully popularized, the past anxiety about "whether you can use AI" will evolve into "whether you can use AI to define the results."

When tools automatically handle most of the execution actions in the future, what truly determines victory or defeat will be your ability to clarify the direction, the vision to set goals, and the in - depth insight into core business scenarios.

Technology is concentrating, and usage is differentiating.

The same tools, in the hands of different people, the more people use them, the deeper the cognitive gap becomes.

Conclusion

Back to this conversation, Hassabis talked about three things: the direction is changing, the capabilities are changing, and the usage is changing.

When these three things happen simultaneously, the gap will not stop widening.

Some people are still comparing which tool is better, while others are using tools to find new opportunities.

The question is never how far AI will develop, but where you will stop.

📮 Original Link:

https://www.youtube.com/watch?v=SSya123u9Yk

https://www.youtube.com/watch?v=C0gErQtnNFE&t=21s

This article is from the WeChat official account "AI Deep Researcher". Author: AI Deep Researcher. Republished by 36Kr with permission.