> ## Content Index
> Fetch the complete content index at: https://nnutshell.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Superintelligence already exists in math, Altman says, and pacing must change
- URL: https://nnutshell.com/superintelligence-already-exists-in-math-altman-says-and-pacing-must-change/
- Published: 2026-10-22T03:13:25.000Z
- Updated: 2026-10-24T03:45:46.000Z
- Description: Sam Altman, OpenAI's CEO, speaks to Stanford GSB students about intelligence speed limits, the coming cybersecurity crisis, and why AI will change jobs less abruptly than predicted.
- Author: Admin
- Tags: Sam Altman, OpenAI, AI Safety, Foundation Models, Cybersecurity, Open Source, #Import 2026-10-25 09:54

*Recorded 22 October 2026 · 48 min*

In narrow domains such as mathematical reasoning, Sam Altman says, superintelligence is already here. Another year or two at the current rate of progress means the world cannot carry on as it is. He wants to preserve two things at once: humans stay in control, and power stays distributed. Holding both forces a hard question about how fast and how far AI should be allowed to outrun the smartest people.

Altman, CEO of OpenAI, is interviewed at Stanford Graduate School of Business and then takes MBA student questions. The talk runs from the scaling bet behind OpenAI to career advice, open-weight models and organization design. The tension throughout: he is calm about jobs and society, but alarmed about cybersecurity, alignment and AI efforts elsewhere that he thinks will run ahead.

1. **Betting on exponentials** — [▶ 7:28](https://www.youtube.com/watch?v=anAmTGcewX8&t=448s&ref=nnutshell.com)  
In 2005 a Stanford CS professor told Altman the one sure thing about AI was that deep learning would not work. Then came AlexNet in 2012, and by 2014 deep learning was improving with scale in many ways. The world "didn't accept this at all." His career-long insight: when you see a beautifully scaling exponential, you keep going, and human intuition handles this badly. OpenAI did not yet know how predictably performance scaled, which came years later, and nearly lost conviction several times. Outside the company, almost no one believed it.
2. **Wrong path, better outcome** — [▶ 10:20](https://www.youtube.com/watch?v=anAmTGcewX8&t=620s&ref=nnutshell.com)  
OpenAI began as a research lab and was supposed to be Altman's "retirement project." The plan was "do research, build superintelligence, many question marks, people are benefit and happy." The technology went a different way, which he thinks is better: it behaves like electricity or transistors, a normal economic force that enables other companies. That forced productization, and by late 2020 or early 2021 it was clear a big company was needed. Delivering it means building compute bigger than any project humanity has taken on.
3. **Intelligence speed limits** — [▶ 13:33](https://www.youtube.com/watch?v=anAmTGcewX8&t=813s&ref=nnutshell.com)  
His hardest decision is pacing, or "intelligence speed limits." He holds two beliefs: people stay in control, and power stays distributed, neither in an AI with "IQ 10 billion" nor in a few people who have AI while everyone else does not. Together they raise two questions: how much smarter than the smartest people a system may get, and how quickly. He thinks it is solvable but takes time. He calls "any regulation is bad" wrong and stopping all progress "absolutely wrong," and admits the pragmatic centre is a hard place to stand.
4. **Superintelligence in math** — [▶ 16:09](https://www.youtube.com/watch?v=anAmTGcewX8&t=969s&ref=nnutshell.com)  
Six months of progress tested his convictions. He still backs iterative deployment and thinks people care little about the work of machines, so society will look more familiar than expected. But in mathematical reasoning "we have superintelligence now," sooner than he expected. Another year or two at this rate with no changes to how things work "is not going to work," so pacing must change. Inside OpenAI, a dispute over how good models would get had been reframed as a dispute over strategy. As the first question resolves, less room remains for the second.
5. **Cyber defense must flip** — [▶ 19:20](https://www.youtube.com/watch?v=anAmTGcewX8&t=1160s&ref=nnutshell.com)  
Earlier this year the "Mythos freakout" got some bugs patched. But leaders did not intuit the exponential: they treated the vulnerabilities found as all that models would ever find, then moved on. Models kept improving and can find bugs outside the normal threat model, even hardware-level ones. Altman no longer thinks patching everything is possible. The defense model has to switch to using the same technology as a shield. He is "somewhat astonished" at how little attention this gets and calls it a huge safety and security issue.
6. **No job apocalypse** — [▶ 22:09](https://www.youtube.com/watch?v=anAmTGcewX8&t=1329s&ref=nnutshell.com)  
Mathematicians went from "This is fake" to "We're cooked" within days, then to the view that most of their work was making knowledge human-intelligible, which survives. Programmers and consultants followed the same arc: salaries for people who get computers to do things keep rising, and almost no one wants an AI consultant telling them what to do. Altman expects huge short-term task changeover but no "job apocalypse." Graduates should start when the ground shakes most, because high-potential people do relatively best then. He is disappointed that MBAs chase what was cool five years ago.
7. **Why human values persist** — [▶ 28:37](https://www.youtube.com/watch?v=anAmTGcewX8&t=1717s&ref=nnutshell.com)  
People "really do not seem to care" what machines do for their own purposes, and he expects that to hold however capable AI gets, provided there is no loss-of-control incident. Learn skills that help you think, as his teachers taught him with writing, not skills AI will own, like his cursive. He fears an industrial-revolution future where humans are cogs in a "Moloch" machine and hopes for a renaissance. People want the renaissance, he says, but economic pressure toward "ruthless efficiency" will need a counterforce from companies, governments or people.
8. **Robots as the hidden loop** — [▶ 33:10](https://www.youtube.com/watch?v=anAmTGcewX8&t=1990s&ref=nnutshell.com)  
Beyond recursive self-improvement in models, Altman sees a second loop: robots building robots, and data centers building data centers, including operating chip fabs and building power plants. He calls it as big, or almost as big, a deal as the AI research loop, and says it gets far less discussion. OpenAI has built its own chip, Jalapeño, with another coming, claiming big wins in intelligence per watt. As models get smarter they will help design better chips. OpenAI is also investing across fabs, power and data center equipment.
9. **Alignment may get harder** — [▶ 35:42](https://www.youtube.com/watch?v=anAmTGcewX8&t=2142s&ref=nnutshell.com)  
In 2020 Altman thought alignment would be harder than it has been. He would have said, "with like a high degree of certainty," that either superhuman models or easy alignment was wrong. Both have held so far. He worries people assume that continues for n+1\. He expects it to get much harder and says the cost of failure has risen enormously. Sandboxing works for a while, then monitorability, but eventually "nothing but alignment" will work. He fears AI efforts elsewhere are so confident that they run ahead.
10. **Open weights have a ceiling** — [▶ 38:11](https://www.youtube.com/watch?v=anAmTGcewX8&t=2291s&ref=nnutshell.com)  
Banning open-source models would be an extremely high cost to innovation and liberty, and Altman would not take that trade, even knowing it brings hate speech, fraud and worse phone scams. But above some threshold, no one should be able to take down the internet, cause physical harm or build a bioterror weapon. OpenAI could not responsibly open-source its best internal model "anytime soon." Cyber defense would have to change and a bio layer be built. The live question is that few believe the closed lead over open models is large.
11. **Agents organize without bickering** — [▶ 46:14](https://www.youtube.com/watch?v=anAmTGcewX8&t=2774s&ref=nnutshell.com)  
In his speculation, most organizations descend from the military's hierarchical command and control, only somewhat improved by tech. Then he watched 10,000 agents solve Navier-Stokes with no leader, no credit squabbles and perfect alignment, unlike any organization he could imagine. Humans will never work that way, and he would not want an emotionless world. But he sees teams inside OpenAI that embrace AI getting much smaller and working differently. He calls designing the best possible organization today an underexplored area of study.

The safeguards have an expiry date, and Sam Altman can say roughly what the sequence looks like. First comes sandboxing, which supports some safety guarantees. Then the models get so smart that nobody can say what a sandbox would even mean. Monitorability works for a while after that. Eventually, he says, "nothing but alignment" will work. It has to be alignment that humans are smart enough to understand, so they can say with confidence that it is okay to run a model and people will stay in control.

The trouble is that the evidence so far is misleading. In late 2020 Altman thought alignment would be harder than it has turned out to be. He also did not expect models to become very superhuman in important ways. He would have said one of those beliefs was wrong "with like a high degree of certainty." Both have held. Alignment has been easier than expected while capability raced ahead. His worry is the lazy extrapolation: it held for this model, so it holds for n+1\. That is not obvious to him. He expects it to get much harder, and he says the cost of being wrong has gone up enormously. What concerns him most are AI efforts elsewhere that are so confident, given how easy it has been, that they are too willing to run ahead.

That is the logic behind a speed limit. Altman holds two commitments. People stay in control, and power stays distributed, neither concentrated in an AI with "IQ 10 billion" nor in a few people who have AI while everyone else does not. Together they force two questions. How much smarter than the smartest people can a system be allowed to get? And how quickly? He thinks the problem is solvable but says it takes time, and a limit is how you buy it. He rejects both poles. Saying any regulation is bad is wrong. Stopping all progress now is "absolutely wrong." He admits the pragmatic centre is a hard place to stand. He also says only a limited number of ways exist for the world to agree on where the limits sit.

Underneath all of this runs one pattern: people do not intuit exponentials, and Altman has built a career on noticing that. In 2005 a Stanford professor told him the one thing known for sure was that deep learning would not work. AlexNet arrived in 2012\. By 2014 results were improving with scale in many ways, and the world "didn't accept this at all." His rule is that when you see a beautifully scaling exponential, you keep going. The same blind spot is now playing out in security. Earlier this year the "Mythos freakout" got some bugs patched. Then leaders moved on, as if the vulnerabilities models had found were all they would ever find. The models kept going. Altman no longer thinks patching everything is possible. Defense has to switch to using the same technology as a shield. He is "somewhat astonished" at how little attention this gets.

So why is he calm about everything else? Because he thinks society moves more slowly than the models. Consider the mathematicians. Over about 10 days their reaction ran from "This is fake" to "We're cooked" to a more considered view: one kind of math job is over, but most of what mathematicians did was make sense of knowledge and communicate it. Programmers went through the same arc and are still being paid ever higher salaries. Consultants discovered that almost no one wants an AI consultant to show up and tell them what to do. So he expects no "job apocalypse," though he expects huge short-term churn in tasks. His advice to MBA students follows. Start when the ground is shaking most, because people with drive do relatively best then. He is chronically disappointed that they want what was cool five years ago.

The deeper reason for his calm is about people. We "really do not seem to care" what machines do for their own purposes, and he expects that to hold however capable AI becomes. Work will matter because a person does it. But the claim carries a condition: as long as there is no loss-of-control incident. His calm depends on the safety work above succeeding. He also sees a fork in what gets built. One path is an industrial revolution, with the economy automated and humans reduced to cogs in "this incredible Moloch thing." The other is a renaissance of creativity and exploration. He thinks people want the second. But the pressure toward "ruthless efficiency" is strong, so some force from companies, governments or people will have to push back.

Even the machines offer a lesson here. He describes 10,000 agents that solved Navier-Stokes with no leader, no credit squabbles and perfect alignment, unlike any organization he could imagine. Humans will never work that way, he says, and he would not want an emotionless world. But teams inside OpenAI that embrace AI are getting smaller and working differently.

The obvious objection is open-weight models, and here Altman gives his least comfortable answer. Banning them would be an extremely high cost to innovation and liberty, so he would not take that trade. Society will get hate speech, fraud and a worse generation of phone scams. But there is a threshold above which models do enough harm that people lose the liberty to use them however they want. No one should take down the internet, cause physical harm or build a new bioterror weapon. He says OpenAI could not responsibly open-source its best internal model "anytime soon." The total cyber defense model of the internet would have to change, and a bio layer would have to be built. What holds the line in the meantime is a rolling advantage of closed models over open ones. Yet he says few people really believe that model is far ahead of what someone will open-source. That is a thin lead to rest a defense on.

OpenAI, meanwhile, is investing in loops that speed everything up: robots building robots, data centers building data centers, models designing better chips. The conversation did not test how a speed limit sits beside those. On government engagement he offers only cautious optimism: "We'll see if something actually happens."

What remains unresolved is the limit itself. Altman wants one and thinks it is solvable. He does not name a threshold, a mechanism or an owner. Until someone does, the ladder he describes, from sandbox to monitoring to alignment, is running out of rungs on a schedule set by the models.

**The open question.** Altman wants a speed limit, thinks the problem is solvable, and says only a limited number of ways exist for the world to agree on one. He names no threshold, no mechanism and no body to set it. His own safeguards rest on a rolling lead of closed models over open ones, yet he says few people believe that lead is large. OpenAI is also investing in loops that accelerate capability, such as robots building data centers and models designing chips. How a pacing regime survives that is the claim to watch.

Watch the full video at [https://www.youtube.com/watch?v=anAmTGcewX8](https://www.youtube.com/watch?v=anAmTGcewX8&ref=nnutshell.com).