AI will eventually become the most powerful force on earth, and within 10 to 20 years it will reshape mathematics, software, work and power. I have been thinking about this almost nonstop for three years. This post archives where I stand in October 2026, so I can check later what I got right and what I got wrong.
Mathematics will never be the same. In 2026 alone, AI disproved the 80-year-old Erdős unit distance conjecture, and OpenAI announced an unreleased model claiming a blow-up (singularity) result for Navier–Stokes. That paper is still being read, and nothing is verified yet. Proofs are objectively verifiable and fully digital, which makes math the field AI conquers first.
The reaction from the community has been alarm. The Leiden Declaration in June, the letter from 25 Fields medalists in September, and the glum write-ups from the meeting at OpenAI all point the same way. Part of it is genuine concern for the field. Part of it is identity: mathematicians built their status on being the sharpest minds alive, and that status is now being taken away.
What I expect next:
• Credit collapses, so incentives change. AI uses published work without citing it. If your name on a paper stops meaning anything, the reason to publish weakens.
• AI math and human math split. Within 10 to 20 years there will be AI mathematics and human mathematics, and humans will be behind. Even if every mathematician on earth tries to decode the AI’s research, the gap will grow.
• The frontier goes behind closed doors. Even if the countries agree to stop or heavily regulate publicly. Companies may publish chunks, but the real work most likely happen privately.
I like Scholze’s philosophy of not using AI in his own work. But AI will eventually surpass him too, and one day even he might end up as an interpreter.
Agentic frameworks now build Adobe-level tools in hours. I built my own photo and video editor with AI, containing only the Photoshop features I actually use, customized for me. Reverse-engineering anything is now automated too.
This breaks the old business model. Users will choose custom-built tools over paying for bloated suites, and AI companies with near-zero marginal cost can undercut incumbents almost indefinitely. Monopolies will be created. Antitrust law won’t stop this outside the EU.
My thesis: agentic coding makes custom tools cheap enough that companies stop paying for upgrades and seats, and keep the incumbent only as the database. Vibe-coded software will still need humans for safety and accountability, but far fewer of them.
The race no one can stop
AI should be regulated, but the race makes that very hard. The US and China each believe that if they slow down, the other side controls the most powerful technology ever built. Neither will stop.
With nuclear weapons, catastrophe required someone to launch. With superintelligence, reaching it may already be the launch. A bomb cannot think or decide for itself; a superintelligent system can, and it may not be fond of humanity. If something close to superintelligence emerges behind closed doors, it could slip out of control before anyone even notices.
The levers that exist are worth naming: interpretability research, monitoring of the huge and physically visible data centers that frontier training needs, testing before deployment, and treaties on the model of nuclear arms control. Whether they will be enough, nobody knows.
The real problem: inequality and power
The most important issue of the next 10 to 20 years is income inequality and the concentration of power, because that is where the potential damage is greatest.
In 50 years, perhaps 90% of jobs are gone. Industrialized regions that got their “controlled” AI running will keep producing. Poor regions will have no money flowing in. Even inside rich regions, power will concentrate. When human labor is no longer needed, who is the money for?
The winners will be decided by access to resources, compute, data centers and lighter regulation, unless there is a genuine effort to prevent a humanitarian catastrophe.
On solutions:
• Universal basic income is the obvious answer, and probably part of it. But a cash transfer alone leaves people dependent on whoever controls the AI.
• Percentage ownership of AI capital by citizens is a stronger answer, on the model of Norway’s oil fund. Capital owners will not accept this willingly; it runs against the very concentration of power that will decide who wins.
• Leverage has to come from chokepoints. Data centers need land, permits, grid power and water. Regions can trade those for public equity, the way Norway made state participation a condition of drilling.
The transition is dangerous. If jobs disappear faster than institutions adapt, the damage happens in that gap.
When there is no job left for a human, everything will have to be rethought: work, ownership, rights, and what intelligence even means.
One last thought. If there are other AIs out in the universe, built by actual aliens, I would still want ours to win. That is my human nature.
