Andrew Chi-Chih Yao's Latest Speech: It Is Precisely a Good Thing That AI Has Boundaries

In 2026, AI seems to be capable of anything: playing Go, predicting protein structures, writing code, conducting scientific research, and even starting to "do mathematics on its own."

Yet a more fundamental question has rarely been addressed head-on: Is there a theoretical boundary to AI's capabilities? And beyond that boundary, what comes as AI's "next level"?

In July 2026, at the 2026 World Artificial Intelligence Conference (WAIC), Andrew Chi-Chih Yao, Turing Award laureate, Academician of the Chinese Academy of Sciences, Dean of the Institute for Interdisciplinary Information Sciences at Tsinghua University, Dean of the School of Artificial Intelligence, and Dean of the Shanghai Qizhi Institute, delivered a keynote speech titled The Power and Limitations of Artificial Intelligence: A Perspective from Theoretical Computer Science.

Drawing from the foundations of theoretical computer science, he put forward three assertions: The essence of AI is "Turing machine + data", which means there are problems it is destined to be unable to solve; The limitations of AI are not bad news—they are precisely the theoretical cornerstone of AI safety; The next level of AI does not lie in "AI for Science", but in "Science for AI", with Quantum AI now emerging.

The following is the edited and abridged content of Andrew Chi-Chih Yao's speech.

1. The Essence of AI: Machine Learning Equips Algorithms with "Data" as a Weapon

Let us first reflect on the essence of AI. AI relies on an extremely powerful tool: machine learning. This algorithmic paradigm has existed for many years, but it was long overlooked by the mainstream computer science community, even as its quality of problem-solving continued to improve.

The basic machine learning framework works as follows: Given a task and an algorithm template that depends on a parameter θ. In other words, this is an entire class of algorithms, and we do not know in advance which one to use. There are two core issues here: The first is the "representation problem"—does a good algorithm actually exist within this template, meaning does there exist a set of parameters that can complete the task? The second is the "learning problem"—if such a set of parameters exists, can it be learned from data within reasonable time and space constraints?

This is very different from the classic computer science paradigm established by Turing. In classic computer science, these two issues are merged into one, solved through researchers' mathematical analysis without any need for external data. The "secret weapon" of machine learning is that it adds "learning from data" to the algorithm's arsenal. It is precisely because of this that we can witness today's astonishing achievements: AI can play championship-level Go and predict protein folding, feats that people have even grown accustomed to.

2. The Boundaries of AI: There Are Problems It Is Destined to Be Unable to Solve

Many people today worry about AI's capabilities, fearing safety risks and loss of control. So, is there a limit to AI's power? This question is not only intellectually intriguing but also of practical value: If we understand where AI's limitations lie, we can use these limitations to design safer AI systems.

If you are a computer scientist, you might immediately give an answer, or even a proof: The famous Halting Problem remains unsolvable by AI. The reason is simple—at its core, AI is a Turing machine, albeit one enhanced with data. Even with the addition of data, it remains a Turing machine, so the Halting Problem still lies beyond its capabilities.

But a more interesting question is: What practical problems are impossible for AI to solve? AI can play Go and predict protein structures—tasks that once sounded like "mission impossible"—leading people to believe that anything demanded in the real world will eventually be achievable by AI. This is not the case. Let me give you two examples.

3. Two "Impenetrable" Fortresses: Cryptography and Quantum Communication

The first example concerns data security. We know that the latest and most powerful AI models can already be used to launch cyberattacks. People naturally worry: Could such AI crack all ciphers? That would be a catastrophe, because our bank accounts and personal privacy all depend on the protection of cryptographic systems.

But the answer is: There exist encryption methods that are secure against all attackers, regardless of whether the attacker uses classic algorithms or AI systems. Let us look at a specific attack model called the "Chosen-Plaintext Attack" (CPA). Suppose I have an encryption box, and a friend uses it to encrypt plaintext into ciphertext for transmission to me over a public network. Only I in the entire world have the decryption box that can restore the content. Now suppose a malicious person gets hold of the encryption box. They can arbitrarily encrypt content of their own choosing, such as "The weather is nice today" or "I have a cat", doing this ten thousand times to accumulate a large number of plaintext-ciphertext pairs. The question is: Is it possible for them to reverse-engineer my decryption key from this?

This model is not abstract—it describes the widely used public-key cryptosystem: The encryption key can be posted on a website for anyone to access, while the decryption key is held exclusively by its owner. CPA security is the lifeblood of public key infrastructure. Cryptosystems like ElGamal, under certain widely accepted mathematical assumptions, can be proven to resist Chosen-Plaintext Attacks. Cryptography is an extremely deep and rich field within theoretical computer science.

Theoretical computer science started from scratch in the 1960s and has grown organically in the decades that followed. The most wonderful part of this field is that new ideas keep emerging, often coming from outside the discipline: Mathematicians, biologists, and electronic engineers continuously bring in fresh perspectives, and almost every decade witnesses the birth of exciting and profound theories. It intersects with physics, biology, and economics, constantly creating new interdisciplinary fields.

The second example, also concerning secure communication, is called Quantum Key Distribution (QKD). It allows two strangers to converse only through a public channel and eventually agree on a random key known exclusively to the two of them. Even if an eavesdropper listens to the entire conversation, they cannot obtain the key. This sounds like a miracle, but it is indeed achievable—and this path goes beyond computer science itself: It uses the laws of physics to detect eavesdropping. Once a hacker attempts to measure the communication between the two parties, they will be detected, and that segment of communication will be immediately discarded. Precisely because of this, secret communication becomes possible.

This method was proposed by Charles Bennett and Gilles Brassard in 1984, more than 40 years ago. Just last March, ACM awarded the 2025 Turing Prize to the two of them, in recognition of their foundational contributions to quantum information science. This is the first time in the history of the Turing Prize that it has been awarded for quantum information research.

Quantum Key Distribution uses quantum technology, and China is perhaps a leader in this field: In 2016, exactly ten years ago, China launched the world's first quantum science experiment satellite, Micius, which enabled Quantum Key Distribution across distances of over a thousand kilometers. China has also built a nationwide long-distance Quantum Key Distribution network connecting overseas regions, with a total length exceeding 10,000 kilometers, serving sectors such as e-government, cross-border finance, and power grids.

This is something AI is powerless against: No method using only computers can break it. Even more remarkably, although it was originally designed to defend against hackers, because it is based on the laws of physics, in principle no force in the universe can crack it—unless someone overturns quantum mechanics.

4. The Most Exciting Direction: AI for Science, Where Machines Begin to "Do Theory"

So, what is the most important direction for future AI research? There are many answers to this question—building incredibly intelligent robots and training larger models both hold great potential. But from a theoretical perspective, I believe "AI for Science" is the most interesting and promising direction in the next three to five years. Let me give you two examples.

The first example is in astronomy. A few months ago, the journal Science published a paper completed through collaboration between the team of Academician Dai Qionghai from the Department of Automation at Tsinghua University and the team of Associate Professor Cai Zheng from the Department of Astronomy. Several of the co-first authors are very young researchers. The astronomy community has long been eager to understand the history of the universe from the Big Bang to the present: How do galaxies evolve? The key lies in observing galaxies in the early universe.

These galaxies are extremely distant, with extremely faint signals. People have expended enormous efforts to decode and reconstruct the early universe from telescope data. This team designed an AI model named ASTERISK, essentially an AI-enhanced processor dedicated to reconstructing those dim galaxies from data. Notably, it does not require collecting new data—it simply analyzes the old data from the Webb Space Telescope better than anyone ever has before, increasing the detection depth by one magnitude and instantly discovering more than 160 new candidate galaxies from 200 million to 500 million years after the Big Bang. This number is three times the total of similar international discoveries previously made. For comparison: The Hubble Space Telescope can see some galaxies, JWST (the Webb Space Telescope) can see more, and JWST equipped with an AI detector has identified a large number of new galaxies—those orange dots in the image are all newly discovered galaxies.

The application of AI in experimental science is already widespread. You can also view protein structure prediction as "conducting experiments from data." But what happened next truly took me by surprise. I used to never worry that AI would replace researchers, thinking it was just a tool to help us analyze data and extract information. But recently, I have truly started to worry about my own job (laughs). Of course, after worrying for a while, I felt happy instead: AI has begun to make theoretical breakthroughs, which was unimaginable before.

The second example is the "cosmic string" problem in cosmology: There is a theory that strings left over from the early universe are still producing gravitational radiation today, and an unresolved problem is to determine the power spectrum of this radiation. Just last March, using Gemini Deep Think, Google Research solved the key integral problem in the gravitational radiation power spectrum of cosmic strings, providing an exact analytical formula and putting an end to this 40-year-old problem. This is no longer just "calculating mathematics"—it is truly "doing mathematics."

There is also a genuine mathematical problem, the Unit Distance Problem: Among n points on a plane, what is the maximum number of point pairs that are exactly 1 unit apart? Erdős once conjectured that this number would not substantially exceed a linear bound. This is an 80-year-old open problem that was falsified by an OpenAI reasoning model last May, with the proof employing profound algebraic number theory. This is true autonomy: You give the problem to the machine, then sit back and wait for two days to see what results it produces.

5. The Next Level: When AI Meets Quantum

I promised everyone that I would talk about the "next level" of AI—not about what AI can do now, but fundamentally, whether there exists a more powerful way for humans to acquire information and create knowledge than existing methods (including AI) in the human knowledge base. To discuss this, we must talk about physics and the story of AI and quantum.

AI and quantum are the two most dazzling technologies in today's world, and almost every tech report lists them side by side. AI is the biggest current hotspot, mature enough to continuously demonstrate various "impossible" feats. Quantum science started very early, almost 100 years ago, but humans did not know how to harness it until four or five decades ago. In the 40 years since, scientists have been working hard to apply this knowledge to computing. I want to talk about how these two technologies can mutually empower each other.

Let me start with a known fact: AI can boost quantum computing. In 1981, physicist Richard Feynman proposed the concept of a quantum computer, using quantum bits to replace the classic 0 and 1 bits for computation. This paper ignited the physics community, and later the computer science community as well: The idea of breaking the monopoly of the Turing machine is irresistible, both intellectually and in terms of its future prospects.

In the 1990s, computer scientist Peter Shor proved that quantum computers can solve some truly difficult problems. For example, mathematicians have long believed that factoring large integers is beyond the capabilities of classic computers, but Shor proved that as long as the kind of quantum computer Feynman envisioned is built, it can crack many cryptographic systems. This paper is extremely important—it truly aroused the enthusiasm of the scientific community and funding agencies for quantum computers.

A few years later, Shor did something that I consider even more remarkable. Initially, physicists were very pessimistic: Quantum bits are too fragile and too susceptible to interference, and the noise problem seemed unsolvable. Shor dispelled their doubts. He proved that as long as the error rate of basic quantum logic gates can be suppressed below a certain threshold (such as 1%), theoretically, a quantum computer can be scaled to any size while remaining reliable. This is entirely parallel to John von Neumann's solution back then for "how to build a reliable classic computer from unreliable components." This means that in principle, physicists no longer have any reason to object: Quantum computers can be built and put to use.

Quantum error correction has therefore been the core challenge for 40 years, until AI recently came to the rescue. A year and a half ago, Google published a paper in Nature introducing a quantum chip called Willow: Using about 100 quantum bits, it created a stable qubit that can theoretically "survive forever", bringing the error rate below the 1% threshold. This is the first time in history, and it is incredibly exciting news for scientists working to build quantum computers.

What role did neural networks play in this? One of the core technologies was designing a decoder for quantum error correction. This decoder could not be written through pure thinking, but it can be learned: A neural network is trained on a large number of samples, and ultimately this extremely valuable decoder is obtained. At the moment, scientists around the world are racing along this path, including some Chinese teams that are performing very well.

Now, conversely: Can quantum enhance AI? Our answer is: Replace the Turing machine in the machine learning framework with a quantum machine as the basic architecture for learning, or even learn directly from quantum data. Under this more general model, is it possible to do things that even AI cannot achieve? People began exploring this direction ten years ago, and in recent years, we have truly seen the light. You can call it "Quantum AI", and it is now emerging. I personally believe that in the next five to ten years, we will see tremendous progress. When the novelty of AI fades and it no longer makes headlines, Quantum AI will take over the baton and allow the marathon of science to keep burning.

6. Epilogue: Understanding Boundaries Is the Prerequisite for Discussing AI Safety

Andrew Chi-Chih Yao: To summarize, AI has demonstrated enormous power in countless fields, but AI algorithms always operate within the boundaries defined by the laws of physics and theoretical mathematics. This understanding provides the foundation for an important task: making AI systems safe and controllable. What this year's WAIC focuses on is precisely the power and safety of AI. Cryptographic technologies are carefully designed by scientists specifically to resist attacks from classic computers; similarly, with a solid mathematical foundation, we can also build systems that can resist AI attacks.

Understanding this boundary is crucial to work in the field of AI safety. And there are so many exciting explorations ahead: AI for Science, such as Quantum AI; reliable large-scale AI systems, which require "Mathematics for AI"; AI safety; and "AI for AI", using AI to improve AI. All of this will lead us to a deeper understanding of the essence of "intelligence".

(This article is compiled based on the recording of Andrew Chi-Chih Yao's speech at the 2026 World Artificial Intelligence Conference and has not been reviewed by Mr. Andrew Chi-Chih Yao. This article was completed with the assistance of Kimi.)

This article is from the WeChat public account "Big Data Digest", author: Digest Bacteria, and is republished by 36Kr with authorization.