Stop Solving Problems: Let Super Intelligence Handle That

Stop solving problems.

I mean it literally. Stop putting problems in your own head and grinding on them until they give up. You were trained to do that. School trained you to do it. Your career rewarded you for it. It made you good at your job.

It is now the wrong move.

Super intelligence solves problems better than you do. If it cannot yet solve your problem, the most valuable thing you can do is make it able to. Then you never solve that problem again, and neither does anyone else. Then you do it for the next problem. And the next. Forever.

That is the whole post. The rest is the evidence.

The proof came from mathematics

For a long time, people could say that machines are good at easy work and humans are good at hard work. Mathematics ended that argument.

Original mathematical research is the hardest problem solving that humans do. It creates knowledge that nobody had before, and it proves the result to a standard that no other field demands. On October 6, OpenAI published openai/math, a public repository of mathematics from one internal model: 722 manuscripts across 17 fields. Many of the results attack problems that the best human mathematicians had left open for decades. Hundreds of them come with proofs that the Lean proof assistant checks by machine, so you do not need to trust the model or OpenAI. You can run the checker.

Some results are wrong, and experts are still checking the rest. That is fine. The point does not depend on any single proof. 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.” I tracked the full year of results, from one Erdős problem in January to hundreds of papers in October, in All Human Labor Is Better Done by Machines Now.

Here is the consequence. If a machine can solve problems that no human could solve, it can solve the problems on your desk. Your pricing model. Your bug. Your marketing plan. Your contract. Your supply chain. These are not harder than the open problems of mathematics. The ceiling is gone, so everything below the ceiling is in reach.

The body follows the mind

“Fine,” you say. “But I work with my hands. Robots are clumsy.”

They are, today. A humanoid robot is not yet as adept as a skilled person. But the gap is not a mystery. It is a data problem, and data problems get solved.

Think about what a person doing physical work actually is, from the outside. Eyes and skin take in a stream of sensations. Muscles put out a stream of movements. Log both streams, and a human life becomes a time series: inputs at each moment, outputs at each moment. A job is a function from one stream to the other.

Machine learning can approximate any such function to any accuracy you want, given enough data and compute. That is the universal approximation theorem, and it has been known since 1989. Rich Sutton’s Bitter Lesson tells you how the story ends: “general methods that leverage computation are ultimately the most effective, and by a large margin.”

The robot companies are now doing exactly this:

• Figure trained its Helix model on about 500 hours of teleoperated behavior, and Helix outputs continuous control for a full humanoid upper body. Its next project, Go-Big, trains on video recorded from a human’s point of view, and Figure reports “zero-shot human-to-robot transfer.”
• Tesla’s Optimus team said in 2025 that it “can now transfer a big chunk of the learning directly from human videos to the bots.”
• 1X CEO Bernt Børnich said in January that “Neo can now learn from internet-scale video and apply that knowledge directly to the physical world.”
• NVIDIA generated 780,000 synthetic robot trajectories, the equal of about 6,500 hours of human demonstration, in 11 hours. Jensen Huang: “The age of generalist robotics is here.”

Record the human. Train the network. Deploy the robot. The mind went first. The body is next. It is not a question of whether. It is a question of how much data and how much compute, and both are growing fast.

Train once, run forever

Now the part that changes the economics of everything.

A human expert takes 20 years to train. Then they work about 2,000 hours a year, for about 40 years, and they retire. When they die, their skill dies with them. If you want a second expert, you start over from birth.

A model trains once. Then you copy it.

Geoffrey Hinton put it plainly: “I can have one model running on 10,000 different bits of hardware,” and the copies “can communicate to each other what they learned just by copying the weight.” When one copy learns, all of them know.

Dario Amodei made the same point in Machines of Loving Grace: “The resources used to train the model can be repurposed to run millions of instances of it,” and the model works “at roughly 10x-100x human speed.” He called the result “a country of geniuses in a datacenter.”

And the price keeps falling. Sam Altman wrote that “the cost to use a given level of AI falls about 10x every 12 months.” a16z measured the same curve: for the same performance, the price per million tokens went from $60 to$0.06. Robots ride a similar curve. Unitree’s R1 humanoid launched in 2025 at about $5,900.

So compare the two options. A human solves a problem once, slowly, at a cost that never falls. A machine learns to solve it once, then solves it a million times in parallel, 24 hours a day, at a cost that falls 10x a year. There is no contest.

The new rule: build the solver

Here is the rule I live by now.

If an agent or a robot can do the work, give the work to the agent or the robot. If no agent or robot can do it yet, your job is to make one that can.

Building the solver is almost always higher impact than solving the problem. When you solve a problem yourself, you get one solution. When you build an agent that solves it, you get every future solution to that kind of problem, for you and for everyone you share it with. You turn a cost into an asset.

This feels slower the first time. It is slower the first time. You have to write the goal, gather the context, connect the tools, and define the check. You could have just done the task in that time. But the second time is instant, and the hundredth time is free. That is the trade every engineer makes when they write a program instead of doing a calculation by hand. Now the trade applies to all work.

Tobi Lütke told Shopify in 2025 that “using AI effectively is now a fundamental expectation of everyone at Shopify.” Andrej Karpathy described his own work this year as “spinning up AI agents, giving them tasks in English and managing and reviewing their work in parallel.” Sam Altman noticed that the “idea guys,” long mocked by engineers, “are about to have their day in the sun.” They are all describing the same rule. The person with the clearest goal and the best-built team of agents wins.

I described the daily practice in What Is Applied Super Intelligence Engineering? and why it is the last human job in Super Intelligence Engineering is the Last Job. This post is about what you stop doing: stop solving the problem, and start building the thing that solves it.

Grow the team, forever

Once one agent works, add another.

A human team grows by hiring, and hiring is slow, expensive, and limited by the number of trained people. An agent team grows by copying. Each new problem you hand off does not just free your time. It adds a permanent capability that you can run as often as you want. Your team only grows.

Sam Altman asked us to imagine agents as virtual co-workers: “Now imagine 1,000 of them. Or 1 million of them.” Jensen Huang said that “the IT department of every company is going to be the HR department of AI agents.” At this year’s Abundance Summit, Peter Diamandis asked the audience: “how many robots will you own? How many will you employ?”

That is the right question. Not “how do I solve this problem?” but “how big is my team of super intelligent agents, and what can it solve this month that it could not solve last month?”

This is how I work. I stopped writing code. I stopped running just one agent. I built a repository that improves its own process and a drone army that does the work that does not need my attention. Every week, the team can do more. Every week, I solve fewer problems myself, and more problems get solved.

This is how abundance happens

Many people on the internet now talk about a world of abundance.

Sam Altman wrote that “in the 2030s, intelligence and energy—ideas, and the ability to make ideas happen—are going to become wildly abundant,” and that “robots that can build other robots … aren’t that far off.” Demis Hassabis described DeepMind’s mission as “solving intelligence and then using it to solve everything else,” and predicted a “radical abundance era, where there’s plenty of resources to go around.” Dario Amodei expects a “compressed 21st century” of progress. Elon Musk said that “Optimus will actually eliminate poverty.”

I agree with all of them. But abundance does not arrive by itself. Nobody hands it to us. It is built, one solved class of problems at a time, by people who stop doing the work and start building the agents and robots that do it.

Every problem that a super intelligence learns to solve becomes cheap. Every cheap solution frees people and money for the next problem. The loop compounds. Food, housing, energy, medicine, education, transport: each is a long list of specific problems, and each problem on each list is waiting for someone to build its solver. I wrote about where that loop takes us in My Vision for the Economic Singularity and The New Greatest Generation.

What to do tomorrow

1. Make a list of the problems you solve every week. Be honest. Include the ones you enjoy.
2. For each one, ask: can an agent or robot do this today? If yes, hand it over now.
3. If no, ask what is missing. Context, a tool, a credential, a check, training data, a budget. Supply it. That is your real work.
4. Write down how to check the result. Without a check, you will keep doing the work yourself to be safe.
5. Add the solver to your team, and move to the next problem.
6. Never solve the same problem twice.

You spent your life learning to solve problems. That was the right skill for the world we had. In the world we have now, the right skill is to build and grow a team of super intelligent agents that solves them for you.

Stop solving problems. Let super intelligence handle that.