All Human Labor is Better Done By Machines Now
All human labor is better done by machines now.
This is not a prediction about some distant future. It follows from one theorem and three experiments that are already complete. The theorem says that a neural network can approximate any function. The experiments are language, robotics, and original mathematics. Add one economic fact, that machines get cheaper every year, and the conclusion is plain. Computers and robots will do all labor. The only job left for humans is to oversee the super intelligence.
That is a good thing. It is the end of toil.
Every job is a function
A job takes inputs and produces outputs.
A translator takes a sentence in one language and produces a sentence in another. A radiologist takes an image and produces a diagnosis. A programmer takes a goal and produces a working program. A warehouse worker takes what their eyes see and what their hands feel, and produces movements of their arms and legs that put a box on a shelf. A mathematician takes a question and produces a proof.
In each case, the worker is a function. The function can be very complicated. It can depend on years of context and on small details of the world. But it is still a mapping from inputs to outputs.
If you can record the inputs and the outputs, you can turn the job into data. Record what the translator reads and what they write. Record what the robot operator sees and how they move the controls. Record the question and the proof. Then the job is a data set, and a data set is something a machine can learn.
Neural networks can approximate any function
In 1989, George Cybenko, and separately Kurt Hornik, Maxwell Stinchcombe, and Halbert White, proved what is now called the universal approximation theorem. It says that a neural network with enough neurons can approximate any continuous function on a bounded domain, to any accuracy you choose.
This is the theory. There is no function of that kind that a neural network cannot represent. If a job is a function, a neural network can do the job.
The theorem does not say how to find the right network. It says the network exists. For decades, that gap was the whole problem. We did not have enough data, enough compute, or good enough training methods to find the network in practice.
That gap is now closed. The proof is not another theorem. The proof is a set of experiments that worked.
Experiment 1: language
Language was supposed to be the thing that made humans special. It is the medium of law, science, business, teaching, and nearly all office work.
Language is also a function. It maps the text so far to the text that comes next. That function is as complicated as anything humans do, because a good continuation can require knowledge of physics, history, law, code, and the person you are talking to.
Neural networks learned that function. Today’s models write, translate, summarize, argue, and program at or above the level of skilled professionals. In software, the change is complete. I have not written a line of code by hand in a year, and I am not unusual. The head of Claude Code told Fortune the same thing. I described that change in We Are All Managers of Super Intelligence Now.
If a neural network can learn language, it can learn every job whose inputs and outputs are text. That is most office work.
Experiment 2: robots
Physical work is a function too. Its inputs are cameras, touch sensors, and joint positions. Its outputs are motor commands. For a long time, this function was too hard to program by hand. Every new task needed engineers to write new rules for every new situation.
Now robots learn from data, the same way language models do. And the robots are leaving the lab.
Figure’s humanoid robots worked shifts at BMW’s plant in Spartanburg, South Carolina, where reports say Figure 02 supported the production of more than 30,000 BMW X3 vehicles. In April 2026, Figure reported that its BotQ factory went from building one Figure 03 per day to one per hour.
Unitree shipped more than 5,500 humanoid robots in 2025 and listed on Shanghai’s STAR Market on August 19, 2026. In the first half of 2026, AgiBot shipped about 8,400 humanoids, and the industry as a whole shipped about 19,100, up 272% from a year earlier.
1X takes orders for NEO, a humanoid for the home, at 499 per month. When NEO cannot do a task by itself, a human operator at 1X can guide it remotely. Some people see that as a weakness. I see it as the method. Each time a person guides the robot through a task, the robot records the inputs and the outputs. The job becomes data. The data trains the next model. The operator’s work is the last time a human does that task.
This is the pattern for all physical labor. Record the job. Train the network. Deploy the robot.
Experiment 3: original mathematics
If any human work was safe, it was original mathematics. Mathematics research is not repetition. It is the creation of new knowledge that nobody had before, proven to a standard of certainty that no other field demands.
That threshold is behind us.
The clearest record is OpenAI’s. In ten months, its results went from one problem at a time to hundreds of papers at once.
January. Two amateurs, Kevin Barreto and Liam Price, used GPT-5.2 Pro to solve Erdős Problem 728. Harmonic’s Aristotle system checked the logic. Nobody found an earlier human proof.
May. On May 20, OpenAI announced that an internal model disproved Paul Erdős’s 1946 unit distance conjecture, a central problem in discrete geometry. The model built a counterexample from deep ideas in algebraic number theory. Mathematicians including Noga Alon and Tim Gowers checked the argument. Jacob Tsimerman said, “It is definitely an intimidating construction.” Alon said these systems are “changing dramatically the way mathematical research is being done.”
July. OpenAI published a proof of the cycle double cover conjecture, a graph theory problem that was open for about 50 years. GPT-5.6 Sol Ultra found it in less than an hour with 64 subagents that worked in parallel. Mathematicians at the Institute for Basic Science in Korea then published an exposition of the argument. In the same month, a proof of Crouzeix’s conjecture from GPT-5.6 was reviewed and accepted by Michel Crouzeix himself.
August. On August 1, OpenAI said that its next model, Astra, solved ten problems that were each open for at least a decade. They include an explicit non-sofic group, which answers a question that Mikhail Gromov asked in 1999, a disproof of Connes’s rigidity conjecture, three Erdős problems, and the first improvement since 1978 to the general upper bound on sphere packing density in high dimensions. Every result came with a proof that the Lean proof assistant checks by machine. The compute for all ten cost about $2,000.
September. OpenAI released GPT-6 Astra on September 3. On September 8, OpenAI claimed that a more capable internal model proved that the three-dimensional Navier–Stokes equations can develop a singularity in finite time. That is one of the seven Clay Millennium Prize problems. OpenAI says about 10,000 agents worked together for about 88 hours and produced a Lean formalization. The mathematics community has not accepted the result yet, and there is a public dispute about credit. Even so, a machine-checked attack on a Millennium problem, by a swarm of agents in less than four days, was not a serious possibility one year ago.
October. On October 6, OpenAI published openai/math, a public repository of mathematics from an unreleased internal model. It started with 722 manuscripts in 372 families of results, across 17 fields. OpenAI gave the model about 4,000 open problems, and each result used about three hours of compute on average. The claimed results include the Kakeya conjecture in four dimensions, the Mahler conjectures, the isomorphism of the free group factors, an extension of Hilbert’s tenth problem to the rational numbers, new zero-free regions for the Riemann zeta function, and a proof of the Hodge conjecture for CM abelian varieties.
The release is not clean. On October 7, OpenAI withdrew three manuscripts after experts found a sign error, and it revised 14 more. Most results have no human review yet. But by OpenAI’s own count, 300 of the 719 remaining top-line results, about 42%, have Lean proofs. A Lean proof does not depend on trust in the model or in OpenAI. Anyone can run the checker. Even if you count only those, one model produced hundreds of new, machine-verified mathematical results in one evaluation.
Michael Douglas of Harvard said the day after the release: “Yesterday will be the day when a new era of math and mathematical physics began.” Andrew Sutherland of MIT gave the necessary warning: “we should expect some of the proofs to contain mistakes, possibly serious ones.” Both statements are true. Checking the results is now the hard part of the work. Producing them is not.
OpenAI is not alone. Quanta Magazine reports that a Google DeepMind agent resolved 9 of 353 open Erdős problems, at a cost of a few hundred dollars each. In July, Levent Alpöge and Akhil Mathew, working with Claude, found an explicit counterexample to the Jacobian conjecture in dimension 3. The list of mathematical discoveries by machine intelligence now grows every week.
Mathematicians are already changing their own description of their job. Jeremy Avigad, in a guest post on Terence Tao’s blog called The Future of Mathematics, wrote that “we get to choose the problems, grade the solutions, and favor the ones we like best.” That sentence is a description of oversight. It is the job I describe below.
Original mathematics is the hardest function humans perform. Machines now perform it.
Theory plus experiment
Put the pieces together.
The theory says that neural networks can approximate any function. The experiments show that we can find those networks in practice for language, for physical movement, and for original mathematical research. These are not easy cases. They are the hardest cases. Language was the core of human intelligence. Dexterous movement in an unstructured world was the core of manual work. Original proof was the peak of human reasoning.
If machines can do the hardest functions, the easier ones are a question of data and engineering, not of possibility. Every remaining job is a function that nobody has recorded and trained yet.
The cost goes toward zero
Possible is not the same as economical. A robot that can do a job but costs more than a person will not take the job.
Today, robots are still expensive. An Anthropic study found that robots can perform about three quarters of the physical tasks in US jobs, but that they cost less than a person for only 0.3% of work today. At the historical rate of about 3% cheaper per year, that study says, the change would take decades.
I do not think the historical rate holds. Computers and robots are manufactured goods, and manufactured goods get cheaper with scale. Humanoid production went up by 272% in one year. Figure increased its build rate by 24 times in four months. The cost of a unit of machine intelligence falls every year. Ten open mathematics problems cost OpenAI about $2,000 in compute. A result in the October release used about three hours of compute. A human mathematician can spend years on one result. And super intelligence now designs, programs, and debugs robots and their factories, which means the machines speed up their own cost decline.
A human worker has a cost floor. A person needs food, housing, health care, sleep, and years of training, and nobody can mass-produce a trained person. A machine has no such floor. Its cost falls toward the cost of the energy and the materials it uses. Over time, that approaches zero compared to the cost of a human.
When a machine can do every job, and the machine is cheaper than a person, the machine does every job. That is not an opinion about technology. It is arithmetic.
Elon Musk said it directly at the US-Saudi Investment Forum in November 2025: “Work will be optional.” He compared future work to a hobby, like growing vegetables in your backyard when you could buy them at the store.
The last job
When machines do all the labor, what is left for humans?
One job: oversee the super intelligence.
Someone has to decide which problems matter. Someone has to state the goal, give the system the context and the tools, define what done means, and check that the result is the right result. Someone has to notice when a brilliant answer solves the wrong problem. That someone is a human, because the purpose of the work is to serve humans.
I have written about this job from several directions. In The Only Thing Left To Do Is Unblock The Super Intelligence, the job is to get the machine its compute, money, and energy, and to point it at important problems. In We Are All Managers of Super Intelligence Now, the job is to keep many agents working for a long time with good checks. Here the same conclusion applies to all labor, not only software. Every worker becomes an overseer.
Liberation
Some people hear this argument and feel fear. They ask what people will do without jobs, and how they will earn a living.
Those are real questions about how we share the output. They are not reasons to want toil. For all of history, most humans spent most of their lives doing work they did not choose, because the work had to be done and nobody else could do it. People broke their backs in fields and mines. They lost their eyesight in factories. They spent their best years on repetitive tasks in offices.
That era is ending. The machines can do the work, and they will do it better than we did.
We are liberated from labor. We do not have to spend our lives as functions that map inputs to outputs. We get to choose what we want: which problems to solve, what to build, what to learn, and who to spend time with. The economy that results is one of abundance, and I wrote about how to build it in My Vision for the Economic Singularity and The New Greatest Generation.
All human labor is better done by machines now. The last job for humans is to oversee the super intelligence. Take that job, and do it well.