Professor Shen Yang of Tsinghua University on the AI endgame: Will humans be left with only the right to sign off?

Will Humans Ultimately Be Left With Only the Right to Sign? Professor Shen Yang of Tsinghua on the Endgame of AI

New AI Era, October 25, 2026, 16:15:23, from Beijing

New AI Era report

118,511.

The team also had AI rank these conjectures and select the top 100. In AI's eyes, the Goldbach Conjecture, the one most familiar to Chinese people, ranks only 49th.

Today, New AI Era holds an in-depth conversation with Professor Shen Yang.

As an interdisciplinary scholar spanning the humanities, sciences, engineering, and medicine, he has served as a professor or postdoctoral mentor in fields including computer science, information management, journalism and communication, artificial intelligence, and clinical medicine, and his explorations in AI painting, AI music, and AI in the humanities and social sciences have also attracted widespread public attention.

In the interview, Professor Shen Yang systematically laid out four core lines of thought: self-evolving mathematics (the PCU framework), self-evolving algorithm optimization (the PAE framework), self-evolving software engineering (the ECX paradigm), and two endgame scenarios for the future structure of an AI society. This is not only a technology evolution roadmap, but also a memorandum of thought for humanity to reclaim the "helm" amid the flood of intelligence.

From "Solving Hard Problems" to "Self-Evolving Mathematics"

Recently, OpenAI's breakthrough in mathematics has sparked heated discussion. Public opinion is all asking: exactly how many hard problems has AI solved? Has it already surpassed human mathematicians?

But in my view, a more striking fact has been overlooked—the 722 manuscripts released this time are only the tip of the iceberg. If the system were willing, it could at any time throw out 7,000 or even 70,000 manuscripts for humans to verify.

[Core Turning Point] In the future, the speed at which machines produce mathematical knowledge will, over the long term, exceed the speed at which humans understand it. For the first time, humanity will on a large scale possess mathematical knowledge that it does not yet truly understand but must continuously verify and use. The focus of mathematical research is shifting from "conquering a single hard problem" to "creating new methods that solve a class of problems"; part of the brain function that once belonged to mathematical geniuses is beginning to migrate into continuously running intelligent systems.

Mathematics may thus become the first discipline to achieve large-scale "self-reproduction of knowledge": old theorems help discover new theorems, new theorems give rise to new tools, and new tools in turn unearth more mathematical structures, forming a continuously expanding research loop. This also forces reform in traditional mathematics education—AI is friendly to senior scholars with judgment, friendly to young people who dare to explore, and weakens only the academic barriers built purely on seniority.

1. The "Mathematical Interface" Hypothesis: Creating Human "Explanatory Interfaces" for Unfamiliar Mathematical Structures

When AI can easily mass-produce mathematical results, the most important work of future mathematicians will no longer be merely proving new theorems, but creating new concepts, new modes of expression, and new theoretical systems based on AI's self-evolution. This requires us to re-examine an underlying question: are the mathematical concepts used today the only way to express the mathematical world, or are they cognitive tools that humans built to understand the world?

Humans are accustomed to using symbols, formulas, and figures because they fit the thinking habits of the human brain. Humanity's existing mathematical system may be merely a "mathematical interface" adapted to the human brain—neither the only way to express all mathematical truths, nor necessarily the optimal one. AI is not subject to the physiological limits of the human brain, and in the future it will likely discover mathematical structures that are logically entirely valid yet difficult to translate directly into concepts familiar to humans. By then, the important task of mathematicians will be to create "explanatory interfaces" that humans can understand for these unfamiliar structures.

2. The PCU Research Framework: Mathematics Moves From "Scarcity of Output" to "Scarcity of Human Verification"

Based on this change, I propose a mathematics research framework for the AI era—PCU (Production of knowledge, Certification of results, Understanding):

This will also completely rewrite the evaluation criteria for mathematics. When theorems can be mass-produced at low cost, simply increasing the number of theorems is no longer valuable, and mathematics will move from "scarcity of output" to "scarcity of human verification." Without the ability to filter, organize, and integrate, massive AI outputs will only pile up into "knowledge garbage" that is hard to use. What is truly scarce are "hub results" that can be repeatedly invoked, can unify a large number of existing conclusions, and can give rise to new theories.

3. Self-Evolving Mathematics and Cross-Intelligence Mathematics: Co-Creating a New Path to Truth

Following this line of thought, I further propose "self-evolving mathematics." Traditional mathematics mainly studies: under existing axioms, definitions, and rules, what can we prove? Self-evolving mathematics asks: can we create entirely new concepts, modes of expression, and reasoning structures that make mathematical research that was previously impossible become possible?

It studies not only mathematical knowledge itself, but also how mathematics can leverage human-machine collaboration to automatically distill complex underlying derivations into new tools and continuously grow new mathematical capabilities. The most valuable breakthrough may not be solving the hardest problem, but creating a new method that makes thousands of problems easy to solve.

From this, we can also use AI to explore a more far-reaching "cross-intelligence mathematics." Human intelligence, AIs of different architectures, and even other intelligent systems of the future will form their own mathematical concepts, modes of representation, and proof preferences. Some problems that humans find extremely difficult can be easily solved by machines with a different representation; conversely, structures that humans can understand by intuition may not be approached by machines in the same way. Mathematical truth may not change with different intelligences, but the paths to truth are far richer than previously imagined.

Looking back at the history of mathematics, the breakthroughs that truly changed the face of mathematics were often the creation of new concepts, new modes of representation, and new theoretical systems. What is truly worth anticipating in the AI mathematics revolution is not how many famous conjectures machines have solved, but the establishment of a "self-evolving mathematics" system that lets AI and humans continuously co-create the future of mathematics.

The "PAE Framework" that crosses a ten-thousand-fold computing power gap

This research began with AI painting. Over the past three years, I have used AI to create a cumulative 140,000 paintings and held 13 AI art exhibitions around the world. Through massive creation, I developed an exploratory method: first set a general direction, then generate a large number of different options, and through comparison, screening, and repeated revision, continuously approach a better result.

This method can be fully transferred to algorithm research. The two share an underlying commonality: you do not need to know the optimal answer from the start; instead, first open up as many possibilities as possible, then seek breakthroughs through continuous feedback. The essential difference is that painting pursues aesthetic novelty and uniqueness, while algorithms must withstand cruel objective testing—if it cannot improve performance, reduce cost, or run stably in real environments, it cannot be called a success. Simply put, AI is responsible for creating more possibilities, and experiments are responsible for judging which possibilities are truly valuable.

As practice deepened, I began to think about a more core question: can every successful experience of algorithm optimization be transformed into the capability for the next round of algorithm optimization? This is the "self-evolution" I truly hope to achieve.

1. Seven hardcore cross-modal practices: the track record of a team spanning the arts and sciences

Along this route, our team simultaneously entered multiple hardcore cross-modal fields, including data compression, optimization solving, nuclear fusion simulation, image recognition, industrial voiceprint detection, CAD geometric solving, and text classification. This was inconceivable in the pre-AI era, and we have now achieved seven interim results:

Beyond these seven public results, we also have nearly fifty algorithm explorations in progress—some with remarkable results, some progressing slowly, and some that ultimately failed. But failed algorithm experiments are just as important as successful ones, because what we want to build is not a showcase that only displays success stories, but an evolutionary system that can continuously learn from both success and failure.

2. The PAE algorithm self-evolution framework: dual-layer loops and four-level breakthroughs

Looking back at these cross-disciplinary cases, we neither used the same algorithm nor applied a fixed model; what was truly the same was the research logic behind them: every time we enter a new field, we repeatedly ask three questions—where is the current algorithm stuck? Can the limitation be broken through? How can experiments verify the effectiveness of the breakthrough? On this basis, I distilled the general PAE algorithm self-evolution framework:

Penetration (see the bottleneck clearly): break down the entire algorithm operation process (data, structure, training, inference, and resource consumption) to find the root cause that truly limits performance; study not only algorithms, but also "the methods for studying algorithms."

Advancement (realize value): make optimization produce verifiable net gains in objective metrics and real scenarios.

Expansion (open new space): this is the part I value most. "Inward expansion" creates new algorithm representations and solving mechanisms, making previously unattainable algorithm spaces explorable; "outward expansion" greatly reduces computational cost and runtime, freeing up resources so that higher-frequency experiments and more complex real-time interaction become possible.

Penetration discovers the breakthrough point, Advancement realizes the gains, and Expansion creates the possibilities for the next round. To this end, we built a dual-layer loop mechanism: the inner loop continuously optimizes specific algorithms (making algorithms stronger and stronger), while the outer loop continuously optimizes the methods for discovering problems, designing experiments, and selecting solutions (making the system that creates algorithms stronger and stronger). At the same time, we divide optimization into four levels by breakthrough difficulty: weak optimization based on human algorithms, strong optimization that further improves AI optimization results, extreme optimization that challenges the boundaries of strong optimization, and impossible optimization that appears unsolvable under existing constraints.

3. A computing power gap does not equal an algorithm performance gap: a comeback across a ten-thousand-fold computing power gap

[The Law of Computing Power Breakthrough] A computing power gap does not necessarily equal an algorithm performance gap. In the past, people believed in "brute force creates miracles," but in specific algorithm problems, search paths, problem decomposition methods, and convergence efficiency are equally critical. If others need to blindly search ten thousand candidate solutions, while through deep "penetration" and algorithm construction we need to search only a few key solutions, then even with a ten-thousand-fold difference in computing resources, we can still overtake them on specific tasks.

When AI systems with powerful computing power have picked all the low-hanging fruit, humanity's opportunity lies precisely in the search paths that AI has not yet found—the human role will shift from "personally completing every step of computation" to discovering key constraints, creating new problem representations, judging exploration directions, and designing verification mechanisms.

English text is being prepared.

III. Software Engineering and ECX

From "Software Development" to "Software Self-Evolution"

Although AI has not eliminated engineering challenges, it has dramatically lowered the cost of turning ideas into prototypes. The center of gravity in software R&D is undergoing a fundamental shift—rather than "how to build the feature," what matters now is answering three questions: Is it worth implementing? What advantages does it have over other approaches? Once launched, why is it worth keeping for the long term?

1. The ECX Three-Dimensional Parallel R&D Paradigm: Building a Closed Loop of Software Evolution

Based on this thinking, and building on my previously proposed "Four-Stage Verification" theory (AI verification, human verification, social verification, and verification by time and natural laws), I further propose the ECX three-dimensional parallel R&D paradigm:

E (Evolution, performance evolution): AI automatically identifies system bottlenecks and generates optimization plans, continuously improving operational efficiency, stability, and resource utilization.

C (Candidate, feature candidates): Drawing on user feedback, competitor analysis, and AI-generated ideas, candidate solutions are generated in batches; through comparison, combination, and screening, decisions are made on which features to keep or drop and how they evolve.

X (eXperience, interaction validation): Putting people first, real users' interaction feedback and behavioral data are used to identify experience pain points and unmet needs, serving as signposts for the next round of evolution.

These three dimensions are not a serial pipeline but an evolutionary closed loop that runs in parallel and triggers one another. At present, we are testing the core proposition through two projects—"Xiaosenlin" and the "AI Scientist" led by Associate Professor He Jing of Beihang University: Can software, through continuous use, optimize its existing capabilities while simultaneously growing entirely new ones?

2. Case Study One: "Xiaosenlin"—A Two-Person Team Running a Complete AI-Native Self-Evolution Loop

"Xiaosenlin" is a typical AI-native software R&D practice: with a team of just two people, it acquired 60,000 users and completed an angel round of financing just two months after development began. In March of this year, the lead, Qingqing, who had no programming background whatsoever, shared her software idea on social media; her zero-follower account's very first post received 3,000 likes, completing the initial demand validation. She then used AI "Vibe Coding" to build a first-generation demo within a month and brought in a professional programmer, forming a minimalist two-person model of "non-technical lead + professional programmer" that fully ran through the ECX mechanism:

In the E (performance evolution) dimension: the software's "relationship network" feature had previously shown noticeable lag when rendering 60 characters and 180 relationships. With AI-assisted localization and optimization-path analysis, the team adopted strategies such as batch drawing, skipping invalid line segments, and cache reuse, compressing rendering time directly from 138–182 milliseconds to 18–19 milliseconds—a sharp reduction of 86%–90%. The software can run smoothly on a phone from six years ago.

In the C (feature candidates) dimension: Qingqing compiled a total of 300,000 characters of user comments and used AI to categorize needs and break down scenarios. AI generated multiple solutions for the same need, and after prototype comparison and cost evaluation, the best one was launched. Recently, every feature release has received over 10,000 likes, with the highest being a single feature that received 57,000 likes on Xiaohongshu.

In the X (interaction validation) dimension: through a highly engaged user community, real operating habits and aesthetic preferences are continuously collected, making users co-creators of product evolution.

3. Case Study Two: "AI Scientist"—Recombination and Emergence of Existing Capabilities

If "Xiaosenlin" explores the agile evolution of a lightweight application, then the research platform "AI Scientist (Making Research as Simple as Chatting)," which already has nearly 300,000 users, faces the evolution of a complex toolchain. In iterations aimed at large-scale research users, this project demonstrates another characteristic of software evolution: new capabilities come not only from developing new features but also from recombining existing capabilities. For example, in a recent update, the system connected the previously independent modules of retrieval, reasoning, code generation, and writing, giving rise to a composite workflow of "autonomously generating algorithms—verifying algorithm effectiveness—producing a paper manuscript."

4. Bidirectional Evolution and IPL Role Restructuring: Why Does OPC (One-Person Company) Work?

Growing outward: continuously expanding the boundaries of problem-solving, growing new capabilities that did not exist before through adding modules or recombining toolchains.

When both inward optimization and outward growth can be accelerated by AI, the easier generation becomes, the more important verification becomes. Therefore, in the ECX system, humans have not exited R&D; rather, they have shifted from repetitive coding to goal setting, value judgment, and risk control.

Breaking down the R&D process of AI-native software, there are in fact three core roles: the intent proposer (I, Intent), the prototype developer (P, Prototype), and the engineering implementer (L, Landing). In "Xiaosenlin," Qingqing alone took on both I and P, while the professional programmer served as L; in the "AI Scientist" project, Associate Professor He Jing single-handedly covered all three roles of I, P, and L. This also explains at a fundamental level why the OPC (one-person company) model can truly work today—it is the inevitable product jointly brought about by AI technology empowerment and new organizational structures.

Future software competition is essentially competition over evolution mechanisms. Software engineering is moving from "delivering a static product" to "establishing a living system capable of continuously generating, verifying, and evolving software capabilities."

The Crisis of "Three-Layer Folding" and the Possibility of "Three-Layer Synergy"

Beyond specific technologies and engineering, we must confront a grander proposition. The two underlying iron laws that have sustained the operation of modern society over the past few centuries—"capital cannot make money without the physical and mental labor of ordinary people" and "the price of escaping reality is extremely high"—are being completely shattered by the closed loop of "artificial intelligence + clean energy + robotics." In the face of this upheaval, human civilization stands at a crossroads that will determine its fate.

1. Two Prospects: The Abyss of "Three-Layer Folding" vs. the Opportunity of "Three-Layer Synergy"

If we drift with the current and allow technological monopoly and algorithmic inertia to run out of control, society will slide into the three-layer folding abyss of "high technology, low rationality, and strong rigidity"; but if we actively intervene and use public institutions to tame technological ambition, this is precisely also a historic opportunity for humanity to escape the heavy burden of survival and move toward a "three-layer synergistic world":

2. The Test at the Bottom Layer: Beware of "Dopamine Colonization" and Strive for True Spiritual and Social Empowerment

At the foundation of an intelligent society, the greatest hidden danger is not material scarcity but being "gently done away with." If commercial algorithms are allowed to seek only profit, short-video feeds that respond to micro-expressions at millisecond speed and AI companions that are always obedient will become a 24-hour "electronic womb" feeding instant gratification. Over time, the prefrontal cortex of the general public—responsible for logical reasoning and stress resistance—will systematically degrade. Even if "universal basic income (UBI)" is distributed in the future, that would by no means be a utopia—although the masses at the bottom would be free from hunger and cold, they would completely lose, amid addiction, the social bargaining power to defend their own rights and interests.

Technological safety nets are by no means equivalent to spiritual disarmament. We must strive for a positive prospect through institutional and cultural development:

Establish "anti-addiction guardrails" for algorithms: legislate to restrict addictive algorithmic designs that plunder attention, break information cocoons, and protect the public's nervous system from being colonized by commercial capital.

"Empower development": deeply bind basic livelihood security with community public participation, lifelong education, and offline cultural and sports activities, so that ordinary people freed from survival anxiety not only have food to eat but also have rich real community connections and space for spiritual growth.

3. The Game at the Middle Layer: Refuse to Become "Digital Tenant Farmers and Scapegoats" and Reclaim the Discourse Power of "Helmsmen"

For the professionals, super-individuals, and micro-teams caught in the middle, the AI era is by no means a natural golden age; rather, it is full of institutional traps. The most dangerous slide is to become "digital tenant farmers" and "physical scapegoats": when standardized mental labor is devoured, the middle layer can only make a living by depending on the large-model interfaces (APIs) of top giants. The giants only need to change one line of parameters, raise the interface fee once, or take a certain function back in-house in the next update to wipe out thousands of small teams overnight. More cruelly, when AI surpasses humans even in aesthetics and intuition, since AI has no physical body and cannot go to prison, the only value left to middle-layer humans may be to sign responsibility statements in medical accidents, financial blowups, and engineering collapses, serving as "scapegoats" bearing unlimited joint liability.

To break this alienation, middle-layer forces must actively strive for bargaining power in the technological ecosystem:

Promote the open-source ecosystem and antitrust: support open-source models and distributed computing power, break the life-and-death power of a few tech giants over underlying interfaces, and turn "renting land to farm" into "bringing your own tools."

Turn "the obligation to take the blame" into "the power to steer": since humans must bear legal and moral responsibility for the final results, human professionals must have the right to veto and review black-box algorithms. As long as we adhere to algorithmic transparency and strictly guard the safety bottom line, that responsibility, which only human flesh and blood can bear, will become the highest command power that no supercomputing power can deprive us of.

4. Break the Deadlock of "Top-Bottom Collusion": Step Out of Arrogant Preaching and Rebuild the Attractiveness of the Real World

Following game theory further, the most alarming social deadlock in the intelligent era is the cross-layer collusion between "top-level computing oligarchs" and "the addicted masses at the bottom." To the masses addicted to virtual gratification, the computing giants that distribute subsidies and provide free entertainment are "great benefactors," while the rational and clear-headed people in the middle who try to shut down short videos and force everyone to face reality are instead the most annoying preachers; to the top oligarchs, a docile bottom layer is easiest to manage, and the rational middle layer that demands algorithmic transparency and tries to possess independent computing power is the most dangerous.

Once top monopoly capital uses the bottom layer's majority votes to legally confiscate the middle layer's technological autonomy, humanity will fall into the thorny "cave power-unplugging paradox": when most people voluntarily stay in the warm and realistic virtual cave, how should the clear-headed people outside persuade those in the cave not to unplug their own power?

To untie this deadlock, condescending enlightenment preaching has no chance of success; the only way out is to rebuild "real connections" through action:

Let go of elite arrogance and bind interests with the masses: the clear-headed and innovators cannot just seek their own breakthrough; they must use technology to solve the real problems ordinary people care about most, such as elderly care, medical care, education, and community companionship, so that the masses personally feel that "open and rational technology" can bring longer-lasting happiness than "closed electronic opium."

Reduce the sense of frustration in the real world: people flee into virtual caves often because the real society has too low a tolerance for error. Only by establishing a more tolerant social safety net and warmer offline communities, so that ordinary people's trial and error in reality will not ruin them financially, can the real world win people's hearts back.

5. China's Answer: Cross the Pain of "Test-Taker Bankruptcy" and Turn "Super Infrastructure" into the Foundation for Common Prosperity

Projecting this model onto China's real soil, there is both unique structural pain and deep confidence for breaking the deadlock. We must clearly and directly confront three major real challenges:

The squeeze effect of super platforms: the inertia of giants to "eat the whole stack" easily leaves small and medium-sized innovation teams with no room to survive.

But at the same time, China holds a trump card that is unique in the world—a nationally coordinated "new infrastructure and full industrial chain foundation." If Silicon Valley's advantage lies in cutting-edge algorithms, China's confidence lies in "the most abundant green electricity and the largest physical manufacturing torso." With proper guidance, we can absolutely walk a highly resilient positive path:

Empower all industries with "cheap electricity and computing power": China firmly controls ultra-high-voltage power grids, green new energy, national-level computing hubs, and a supply chain covering all categories of manufacturing. As long as platform monopolies are strictly regulated, these super infrastructures can be turned into "public utilities" that tens of millions of small and medium-sized enterprises can call at low cost.

Push education to shift from "drilling answers" to "solving problems": taking advantage of the depreciation of standard answers, liberate the younger generation from mechanical answer-drilling and shift toward cultivating curiosity, hands-on ability, and the capacity to solve complex real-world problems.

Upgrade "digital work-for-relief": upgrade grassroots semi-public positions from mere "grid-based management" to warm undertakings that truly create social value, such as elderly care, ecological restoration, cultural inheritance, and community building.

6. The Final Breakout: Use the Law of "Model Autophagy" to Become a Hands-On "Modern Mohist"

Facing a vast algorithm empire, on what basis can ordinary people avoid being eliminated? The answer lies in a fatal weakness that AI systems themselves cannot overcome—"Model Autophagy."

When the whole society is consuming AI-generated content and using this secondhand synthetic data to train the next generation of AI, the algorithmic ecosystem will rapidly fall into logical degeneration and collapse like inbreeding. This means that the more powerful AI becomes, the more it desperately craves "firsthand real data" generated by human exploration and trial and error in the real physical world. The virtual world can never close the loop on itself; the real world is the source of life for algorithms.

And in China's Pearl River Delta, Yangtze River Delta, and the vast front lines of innovation, the world's cheapest photovoltaic energy storage, open-source hardware, and local open-source large models are giving ordinary people unprecedented physical confidence:

Technological self-farming (keep the tools in your own hands): ordinary geeks and small and medium-sized teams can fully use cheap green electricity, open-source chips, and local large models to build the ability to operate independently and make things by hand even when disconnected from the internet, becoming "modern Mohists" who understand cutting-edge code and can also pick up a wrench to solve practical problems.

Forge a mental moat: in an era of algorithm flooding, actively training the brain's focus and logical power is the self-investment with the highest return. The real moat in the future is no longer how much free knowledge you have memorized, but the self-discipline to withdraw from algorithmic feeding, maintain deep reading, endure boredom and frustration, and carry out long-chain logical reasoning.

[Final Conclusion] A pessimistic ending belongs only to the conformists who have completely given up thinking. As long as we insist on walking out of the warm algorithmic cave, colliding with the real physical world, and creating with our own hands and bearing the consequences, humanity will always be the true creator and helmsman of this intelligent age.