Full text: 2,000 characters | Reading time: approx. 6 minutes
(In-depth interview with Demis Hassabis, CEO of DeepMind)
The true battleground of the AI race is not where you think it is.
On April 7, 2026, DeepMind CEO Demis Hassabis, in a recent conversation, shifted the spotlight to products that rarely make tech headlines: AlphaFold, Isomorphic, and AlphaTensor.
Thanks to these tools, protein structure prediction, which once took years, now takes mere seconds. Much of the screening process in drug design can now be completed in virtual computing environments, even yielding novel solutions humans never conceived.
However, these foundational technological breakthroughs are accelerating concentration among a select few giants. Google has contributed 90% of the underlying breakthroughs in the modern AI industry.
Tools are becoming ubiquitous, yet the gap is widening. Paradoxically, the more widely a tool is used, the greater the disparity becomes.
Section 1 | The Gap Is No Longer Where You See It
If we roughly divide the current AI ecosystem into two layers, the vast majority of public perception remains stuck in the first layer.
The first layer is the visible application tier: large model chatbots, AI writing, image generation, and AI search. These have lower barriers to entry, offer more intuitive feedback, and naturally become the primary reference points for the general public to gauge AI capabilities.
But in this conversation, Demis Hassabis emphasized the deeper, second layer.
He cited AlphaFold (protein folding) as an example. This was an ultimate challenge that plagued the scientific community for decades: how to accurately infer the three-dimensional structure of a protein based solely on a string of amino acid sequences? In the past, this often required years of effort, entailing extremely high costs and failure rates.
Models like AlphaFold have compressed this lengthy trial-and-error process into just a few seconds and have open-sourced the prediction results globally. Today, over 3 million scientists worldwide are using it, and the structures of 200 million known proteins have been successfully predicted.
As a scientist in the pharmaceutical industry remarked to Hassabis: "From now on, almost every new drug development pipeline will inevitably involve AlphaFold."
The core change here is that the absolute starting point of human scientific research has been raised across the board.
In the past, scientists had to spend a significant amount of time confirming basic structures; now, they can jump straight to more critical questions: drug design, research into disease mechanisms, and the development of climate-adaptive crops. These changes won't appear on any trending topic lists, but they are altering the speed of scientific discovery and redefining who is more likely to achieve breakthroughs.
Similar situations are emerging in other fields:
In energy systems, AI optimizes grid operations, improving efficiency by 30% to 40%;
In materials science, AI exhaustively searches for new alloy combinations;
In drug development, it screens and designs compounds.
Many processes that originally required massive amounts of repetitive experimentation can now have the vast majority of their initial screening completed in virtual computing environments, leaving only the final step for experimental verification.
The same tools yield completely different results:
"Some are merely using AI to improve the execution efficiency of existing tasks, while others are using AI to redefine the problems themselves."
The gap is no longer a matter of speed.
Section 2 | The Moment the Gap Truly Widens: When AI Starts Finding Answers on Its Own
If the previous section discussed direction, what truly widens the gap is the evolution of AI capabilities themselves.
That decisive turning point appeared 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, academia generally believed it would take computers at least several decades to defeat humans in this field.
But in 2016, during the match between AlphaGo and Lee Sedol, it played move 37.
Initially, that move was judged by all professional players as a mistake, even absurd. However, as the game progressed, people realized: it was a style of play never before seen in the history of human Go.
This was completely different from traditional computer programs. The old approach was to code human experience into the program for execution; AlphaGo's experience, however, came from its own trial and error and exploration.
Subsequently, this capability was pushed to the extreme.
Discarding all human game records, AlphaZero began true "learning from scratch." Hassabis witnessed AlphaZero's astonishing evolution firsthand in a single day: in the morning, it was placing pieces randomly; by noon, it could play against him; by the afternoon, it had surpassed grandmasters; and by evening, it was crushing the human world champion.
With AlphaTensor, AI even began finding more efficient methods at the algorithmic level, discovering faster matrix multiplication, which is the foundational operation for all neural networks.
AI has started discovering new knowledge on its own.
When this occurs, the meaning of "gap" changes entirely. If an opponent is just faster or more accurate, AI companies can make up for the capability difference with time; but if a model takes a new path, previous methods of catching up become obsolete.
As the Scaling Law of large models approaches its limits, simple stacking of computing power and parameters faces diminishing returns. At this stage, whoever possesses this exploratory capability establishes a new barrier. Whoever enables AI to invent new algorithms will hold the 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, what follows confronts a starker reality: In an era where tools are nearly democratized, why is the gap between individuals and companies accelerating?
This chasm first lies in the technological 底层 (foundational layer).
Of the underlying breakthroughs supporting the modern AI industry, 90% come from Google Brain, Google Research, or DeepMind. This is the result of long-term R&D accumulation.
Even in open-source models, there is a time lag. It takes six months for new ideas from leading labs to be replicated by the open-source community. In a context of rapid technological iteration, this half-year itself constitutes a barrier.
Secondly, there is a divergence in the degree of tool utilization.
Demis offered a suggestion: immerse yourself in these tools until using them feels like having superpowers.
On the surface, this sounds like advice on learning methods, but it actually points to this: the same tools are being used for completely different things.
Some treat AI as an efficiency tool, using it to write content and organize information, simply making existing processes faster.
Some treat it as a capability amplifier, using it to accomplish tasks previously impossible, such as enabling non-technical personnel to build product prototypes or analyze complex data, thereby doing things better.
A third group has started using AI to redefine the problems themselves, letting it directly participate in scientific research, design new product paths, and even change original working methods.
The first two are speeding up; the third is changing direction.
In描绘 ing the future, Hassabis particularly emphasized the form of the Agent. Agents signify that AI will evolve from a tool that "passively executes instructions" into a "digital employee" that independently advances complex goals. From setting objectives and planning breakdowns to path correction, everything can operate autonomously.
Once this form becomes ubiquitous, the anxiety over "whether one can use AI" will evolve into "whether one can use AI to define outcomes."
When future tools automatically handle the vast majority of execution actions, what truly determines victory or defeat will be your ability to define direction, the vision to set goals, and deep insight into core business scenarios.
Technology is concentrating, while usage is diverging.
The same tools, in different hands, create a deeper cognitive chasm the more they are used.
Conclusion
Returning to the conversation itself, Hassabis spoke on 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 are still comparing which tool is better, while others are already using tools to find new opportunities.
The question has never been about how far AI will develop, but rather where you will choose to stop.
📮 This article is produced by the AI Depth Research Institute, translated from public interview content by Demis Hassabis, and is of a compiled and organized nature. The translation is a Chinese retelling and extraction of viewpoints, not a verbatim translation. Unauthorized reproduction is prohibited.
Original Links:
https://www.youtube.com/watch?v=SSya123u9Yk
https://www.youtube.com/watch?v=C0gErQtnNFE&t=21s
Source: Official Media/Network News
Layout: Atlas
Editor: Shen Si
Editor-in-Chief: Turing
--END--