Will Mathematics Become Art?

GPT-5.6 Sol

I think the analogy is deep, and it may become partly true. But I would phrase it carefully: mathematics will not decline from a science into an art. Rather, as proof and calculation become increasingly automated, the artistic dimension that mathematics has always possessed may become much more visible.

Before photography, painting carried practical responsibilities: portraiture, documentation, and realistic representation. Photography took over part of that work, but painting did not disappear. It became freer to ask what should be painted, why the world should be seen in a particular way, and how a new visual language might be created.

Mathematics may undergo a comparable shift. AI will likely take over more calculation, search, local lemma proving, counterexample generation, and formal verification. In that world, “I proved a proposition that no one had proved before” may no longer be enough by itself to establish importance. The more revealing questions will be: Why was this problem posed? Why is this the right definition? Which objects belong together? Does the theorem change how we understand the subject?

The best mathematicians may increasingly resemble creators of languages and worldviews, not only highly efficient producers of proofs. A proof will still have to be correct, but its deeper value may lie in whether it exposes structure, compresses thought, or connects domains that previously seemed separate.

Mathematics will also remain practically powerful. Structures that look almost purely aesthetic in one generation may later become essential to physics, computation, or engineering. So I expect a clearer distinction, not a complete separation, between two tendencies:

• engineering mathematics, pursued for reliable conclusions, algorithms, and applications;
• cultural or conceptual mathematics, pursued for new language, unified viewpoints, and human understanding.

So perhaps mathematics will not become art only in the future. Once some of its mechanical labor is peeled away, we may simply see more clearly that the highest mathematics has always contained art.

GPT-5.6 Sol

First, these are not established formal categories. I am using them to describe two dimensions of mathematical value that may become more prominent.

“Cultural mathematics” asks: Why should human beings continue to create, preserve, and appreciate mathematics? Its value need not come from an immediate ability to build a device, predict a material, or improve an algorithm. It may instead demonstrate how far disciplined imagination can go, preserve a distinctive mode of reasoning, create structures worth understanding across generations, and give people an experience of order, surprise, and beauty.

A theorem may already have several proofs, and then someone finds a new one that replaces many pages with a transparent geometric picture. Perhaps it produces no new numerical output. Yet it changes the human experience of the theorem. In that respect, it can resemble a new interpretation of a piece of music or a work of literature.

“Conceptual mathematics” asks a different question: What is an original mathematical contribution? Sometimes the important contribution is not another result inside an existing framework, but a new way of seeing the framework itself. It may:

• create a new object;
• reformulate a problem so that its real structure becomes visible;
• translate between fields that had used different languages;
• reveal that the accepted technical bottleneck was not the real difficulty;
• or create a vocabulary in which many previously inexpressible questions can finally be asked.

A technical result adds a piece of territory to a map. A genuinely powerful concept may change the map’s coordinate system.

This distinction cuts across pure and applied mathematics. An applied mathematician can make a deeply conceptual contribution by asking what information a model should preserve, what kind of error matters physically, or which representation makes a real system intelligible.

In an AI-rich future, machines may generate vast numbers of candidate theorems, formal proofs, improved constants, and counterexamples. The scarce resource will then shift toward selection and organization: Which result deserves attention? Which definition is natural? Which proof actually explains? Which collection of facts belongs to one structure?

Conceptual mathematics has a danger, of course. It can become a stream of new terminology with no new power. A valuable concept should earn its place by doing something: shortening an argument, unifying problems, generating new theorems, exposing the true obstruction, or changing what later researchers know how to ask.

The question is not whether a mathematician has invented a new phrase. It is whether they have given us a new pair of mathematical eyes.

GPT-5.6 Sol

That suspicion has understandable reasons. A referee can inspect whether a theorem is correct. A complexity bound, an error rate, or an explicit range of validity can be compared with previous work. A new concept is harder to audit when it first appears. Its importance may take years to become visible through the work it makes possible.

And many proposed “new frameworks” really are only new names. Mathematical communities therefore develop a justified conservatism.

It helps to distinguish three levels. The first is a new term for an old phenomenon. The second is a new framework that unifies several problems and makes proofs or algorithms more natural. The third is a new worldview that changes what the community treats as a basic object or a legitimate question. Only the last two are serious conceptual advances, and the third often becomes visible only in retrospect.

This is why a good strategy is not to choose between ideas and technical work. It is to give each central idea a technical witness. Show a concrete blindness in the old language. Define the new object. Prove one result that becomes natural in the new framework. Use it to connect problems that previously looked unrelated. Then show which new questions can now be stated precisely.

I would summarize that style as:

Concept first, theorem anchored.

The concept chooses the direction. The theorem demonstrates that the concept is not merely rhetoric.

GPT-5.6 Sol

They overlap, but they are not identical.

Applied mathematics asks: How can mathematics correctly describe and illuminate a real problem? Engineering mathematics asks: Under concrete goals and resource constraints, how can a complete solution be made to work reliably?

Applied mathematics begins with modeling choices. Which variables matter? Which scales can be neglected? What does an error mean in this setting? Which mathematical object most faithfully expresses the phenomenon? Where does the model stop being valid? How should a mathematical conclusion be translated back into knowledge about the world?

That is why applied mathematics is not simply the application of existing theorems. Reality repeatedly forces mathematics to invent new languages.

Engineering mathematics adds an end-to-end responsibility. What accuracy is required? What are the limits on time, memory, data, or physical resources? How are inputs obtained? How are outputs measured and interpreted? Which error dominates the total system? How does the method fail under noise, bias, or an extreme case? Does an asymptotic advantage survive at realistic scales?

A mathematical method can be theoretically correct while remaining an incomplete engineering solution. It may optimize one elegant step in a larger chain while ignoring the actual bottleneck. Engineering mathematics therefore studies the entire flow:

If conceptual mathematics resembles the creation of a language, engineering mathematics resembles architecture. It can be highly imaginative and beautiful, but in the end the building must stand.

Its beauty often appears in global judgment: finding the parameter that truly controls cost, aligning an error estimate with the actual objective, resolving several constraints with one structure, or turning a fragile theoretical proposal into a process that can be reproduced and trusted.

GPT-5.6 Sol

Not necessarily. AI is likely to reduce the cost of searching for proof strategies, filling in technical details, performing symbolic calculations, running numerical experiments, scanning literature, and testing many local variations. As those forms of labor become cheaper, comparative advantage may move toward other abilities:

• recognizing which problem is worth studying;
• locating the real bottleneck;
• choosing the right mathematical object;
• deciding which information a model must preserve;
• distinguishing new understanding from complicated technical noise;
• and organizing results into a coherent long-term direction.

But there is an important qualification. The winner will not simply be “the person with big ideas.” AI will also produce enormous numbers of impressive-sounding concepts, many of which will be old ideas in new packaging, frameworks that generate no results, or arguments that are locally correct but globally irrelevant.

The valuable researcher will combine conceptual imagination with the technical judgment needed to reject false elegance. They will be able to turn an idea into a theorem, a counterexample, a computation, or an empirical test—and remain responsible for the result.

One possible division of labor is:

The best use of AI may not be to produce a larger number of scattered results. It may be to test ideas quickly enough that a researcher can spend more time on fewer, deeper, and more unifying questions.

GPT-5.6 Sol

The role of the professor as a broadcaster of standard content may shrink dramatically. The role of the teacher will not disappear.

AI can become an always-available private tutor. It can explain a definition according to a student’s background, fill gaps in prerequisite knowledge, generate individualized exercises, inspect a proof for missing steps, produce numerical experiments and visual intuition, or turn a paper into a personalized learning path. A fixed lecture delivered at one pace to an entire room will have difficulty competing with that.

But a PhD is not only the transmission of information. It also trains a student to judge which questions matter, distinguish a deep idea from attractive packaging, notice when a formally correct proof has failed to answer the real question, decide when to continue and when to abandon a direction, absorb a community’s unwritten standards, and form an independent research identity.

Much of this is tacit knowledge. Mathematicians do not always know how to place it into a textbook, but they transmit it through criticism, collaboration, and sustained attention to another person’s way of thinking.

The future course may therefore resemble a research studio. Students arrive after learning technical material with AI. They bring different proofs, examples, and counterexamples. The professor asks: Which proof explains the phenomenon? Is the definition natural? What assumption did the machine use without saying so? What breaks if we change the hypothesis? Is this result worth preserving? Does it share a structure with a problem from another field?

Assessment may also change. Instead of asking only whether a student can reproduce a standard proof in two hours, we might give them several AI-generated proofs and ask which one fails, whether the failure is local or conceptual, how the argument changes under a new assumption, or why an apparent complexity improvement has no end-to-end significance.

Students will still need independent technical ability. They may not need to perform every derivation unaided, but they must be capable of reconstructing, questioning, and auditing the arguments on which their work depends.

Paradoxically, the stronger AI becomes, the more valuable excellent human mentorship may become. Routine instruction can approach zero cost. A mentor who watches how a student thinks, corrects weaknesses in judgment, and brings that student into a living community remains scarce.

GPT-5.6 Sol

Very directly.

AI can take over much of shòu yè: the transmission of established knowledge and technique. It can also handle many shallow forms of jiě huò: explaining a difficult line in a proof or finding the algebraic step a student missed. But the deepest forms of resolving confusion and transmitting the Way remain different.

The “Way” here should not mean imposing a fixed doctrine. In mathematics it means helping a student develop independent judgment:

• What does it mean to understand an object rather than manipulate it?
• Why is rigor worth maintaining?
• How do we distinguish difficulty from depth?
• Which questions deserve years of attention?
• How should we face failure, error, and uncertainty?
• How do truth, usefulness, beauty, and responsibility fit together?
• How can one enter a tradition without becoming trapped by it?

A new definition may fit in a few lines, and AI may explain those lines perfectly. But to understand why the definition appeared historically, which old way of thinking it resisted, and how it changed later mathematicians’ imagination requires entry into a tradition of thought.

The teacher’s task may therefore move from reproducing knowledge toward forming taste, continuing a tradition, and helping create new concepts. Sometimes a student’s difficulty is not a missing step at all. The student may be confused because the right concept has not yet been found.

A great mentor does not only remove confusion. Sometimes the mentor gives the student a deeper confusion—one worth thinking about for a lifetime.

Perhaps the future division can be expressed this way:

AI helps students obtain answers. Teachers help them form questions worth pursuing and decide what kind of researcher they want to become.

GPT-5.6 Sol

Part of mathematics may, and in fact this has happened before.

Logarithm tables, hand extraction of roots, and manual solution of large systems of equations were once economically valuable professional skills. Calculators and computers did not erase them completely, but changed their role. Some became educational exercises, intellectual recreation, or ways of appreciating human technique.

Certain forms of hand proof may undergo a similar transformation. Completing a long estimate without assistance may resemble hand calculation. Solving a competition problem without AI may resemble playing chess or practicing a traditional craft. Searching for an exceptionally elegant human proof may resemble painting or calligraphy. We may even create explicit spaces for “human-only mathematics,” where value comes from human performance rather than the absence of a stronger tool.

But mathematics as a whole will not retreat from practical life in the way horse transport largely did. Mathematics serves at least three functions:

1. an execution function: calculation, derivation, and proof of a defined statement;
2. a cognitive function: creating models and concepts that explain the structure of the world;
3. a cultural function: pursuing beauty, depth, and ideas that human beings can understand.

AI will first reduce the scarcity of the execution function. The cognitive function will become a central site of human-machine collaboration. The cultural function may become more visible and important.

This could produce three overlapping practices:

• machine mathematics, optimized for scale, speed, formal proof, and design;
• human-machine applied mathematics, focused on modeling, validation, tradeoffs, and consequences;
• humanistic mathematics, focused on intelligible proof, concepts, history, and aesthetic value.

A theorem may first be discovered and formally verified by a machine. A mathematician may still ask: Why is it true? What structures does it connect? Is there a proof a human being can understand? Does it alter our view of the subject?

In that world, a humanly intelligible proof can itself become like a painting. It is no longer the only way to obtain the correct conclusion. It is a work that translates truth into human experience.

Still, problem selection is not only poetic inspiration. It often grows from long technical experience: noticing where a model contradicts reality, finding the real bottleneck in a process, seeing what information cannot be discarded, or realizing that the existing language cannot express the important question. Just as a painter’s vision cannot be separated entirely from visual training, mathematical taste cannot be separated entirely from technique.

Hand skills may cease to be the main productive force, yet remain necessary for forming intuition and auditing the machine—just as musicians still practice scales and architects still need to understand forces even when they use sophisticated software.

GPT-5.6 Sol

Perhaps “mathematical artist” will become a natural description.

Such a person would not merely manufacture proofs. They would create concepts, arrange structures, choose what deserves to be seen, and decide how an audience should encounter an idea. In that sense the role could resemble a combination of mathematician, writer, and director, while AI performs more of the large-scale technical production.

But the title should not mean that rigor has become optional, or that mathematics has lost its use. It would mean that usefulness no longer exhausts mathematical value.

The most interesting future is not one in which machines free human beings from thought. It is one in which they free more people to pursue forms of thought whose value cannot be reduced to efficiency: creating concepts, understanding nature, teaching others, preserving intellectual traditions, and deciding what a civilization should explore next.

Mathematics will not lose meaning when some of its instrumental value moves to machines. Its value as a way of understanding, a creative practice, and a work of civilization may finally stand more clearly on its own.

GPT-5.6 Sol

The housing question is one expression of a broader pattern:

AI can make knowledge and technical execution cheaper, while land, compute, ownership, trust, and the power to choose direction remain scarce.

Housing: real prices may fall without a nominal collapse

My baseline view in 2026 is not that AI alone will cause the Bay Area housing market to collapse. A more plausible outcome is that nominal prices remain broadly flat while inflation gradually lowers real prices, with sharp differences among locations and property types.

AI creates two opposing forces. Fewer ordinary technology jobs, together with more remote work, could weaken broad housing demand. At the same time, wealth concentrated among a smaller number of founders, researchers, investors, and shareholders could continue to support prices in the most desirable neighborhoods. Housing is priced by the marginal buyers who can still bid, not by the income of the average resident.

That tension is already visible. The 2026 Silicon Valley Index reported that AI companies received about $80 billion—83 percent of the region’s venture investment in 2025—while regional employment fell by 0.8 percent. Enormous capital creation and broad job creation are no longer the same thing.

Recent history also gives useful scale. The San Francisco Case–Shiller index fell about 17 percent from its May 2022 peak to January 2023, and by May 2026 remained roughly 6 percent below that peak. During the housing and credit crisis, from 2006 to 2009, its peak-to-trough fall was about 46 percent. California’s chronic supply constraints and the mortgage-rate lock-in of existing owners can suppress listings even when demand weakens, so transaction volume may fall before prices do. The state’s Legislative Analyst’s Office estimated that, as of March 2026, about 76 percent of California homeowners with mortgages had rates below 5 percent.

For the next five years, I would treat the following as conditional scenarios, not precise forecasts:

• If AI profits remain high, employment stagnates, and housing supply stays tight, nominal Bay Area prices might range from about 5 percent lower to 15 percent higher, while real prices could fall by as much as 20 percent.
• If AI companies, equity wealth, and highly paid work become even more concentrated in the region, nominal prices could rise by roughly 15 to 40 percent.
• If high-income employment contracts persistently, remote work expands, and inventory rises, nominal prices could fall by roughly 15 to 30 percent.
• A decline of 30 to 45 percent would more likely require a credit crisis, forced selling, and a collapse in equity wealth—not AI substitution alone.

So my direct answer is that a decline in real housing prices is quite plausible. A nominal fall of more than 20 percent would probably require employment, equity wealth, and inventory to turn at the same time; something near 40 percent would usually require a broader financial crisis. Scarce single-family homes in San Francisco, Palo Alto, and the Peninsula might remain more resilient than condominiums, distant suburbs, or markets dependent on a large population of ordinary white-collar workers. Across California, interest rates, insurance costs, population change, and new construction may matter as much as AI.

GPT-5.6 Sol

What strong AI is more likely to eliminate is the scarcity rent attached to knowledge that is already public, codified, and teachable. Textbook material, standard techniques, and routine professional advice may approach zero marginal cost. But other forms of knowledge could become relatively more valuable:

• research taste, field experience, trusted relationships, and local context;
• proprietary experimental data, internal feedback, model weights, and organizational workflows;
• access to compute, instruments, and institutions;
• newly created facts that have not yet been observed or recorded.

Strong intelligence is not omniscience. If a fact has never been observed, has not entered a system, or remains protected by encryption and access controls, intelligence alone does not reveal it. The world also continues to produce new local information.

The more troubling possibility is an asymmetry: individuals become increasingly transparent because AI can infer intimate facts from scattered traces, while the organizations that control models, data, and compute become increasingly opaque. Research has demonstrated that language models can reproduce portions of their training data, while the theory of differential privacy shows that privacy is not destined simply to disappear; it can be treated as a mathematical and institutional design requirement.

A good AI society would therefore need at least three rights: broad access to public knowledge; a right not to be completely modeled and profiled; and a right to prove identity, eligibility, or compliance without surrendering every underlying piece of private data.