November 21, 2026
Heard in AI

Why Mark Cuban, a self-described AI fan, says AI is 'not smart'

Mark Cuban, who calls himself AI's biggest proponent, argued on The Diary of a CEO that AI is statistical brute force rather than intelligence and that humans keep the edge. Host Steven Bartlett pointed to fast, compounding progress.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on The Diary of a CEO, episode published October 26, 2026

Picture a busy four-way intersection, with cars crossing in every direction. Mark Cuban said he could stand there holding a phone with video and "the best models ever," next to a seeing-eye dog, and he would still choose the dog. "I'm letting the seeing eye dog take me across the street before I let the AI take me across the street," he said.

Cuban is a billionaire investor, a former investor on Shark Tank and the founder of the online pharmacy Cost Plus Drugs. He made the case in a long conversation with Steven Bartlett, host of The Diary of a CEO, in an episode published October 26, 2026. He is no skeptic about using the technology. "I am the biggest proponent of AI that there is," he said. All the same, he said again and again that "AI is not smart."

Statistics that predict the next word

Bartlett asked what Cuban meant when he said there would always be someone smarter than the AI. Cuban started with timing. A model from Anthropic, Google or OpenAI "gets released. It's immediately out of date," he said. "The brilliance of this conversation is not incorporated in any model anywhere."

Then he described how he thinks these systems work. AI, he said, "just knows how to use statistics, basically, to predict the next word. It's predictive. It's not deterministic." By that he meant the answers vary. Ask four different AIs the same question and "you're going to get four different answers." Change the wording slightly, in a way that should not change the meaning, and the same AI will answer differently. "That doesn't work in the real world," he said.

OpenAI's own explainer on how its models are developed partly supports him. Models generate text in tokens, which can be words or pieces of words. Several continuations of a sentence can be plausible, and randomness in generation means the same question can get different responses. The same explainer also describes more than word prediction: development includes data preparation, pre-training, further training after that, and evaluation.

Cuban's other objections were about judgment. AI "doesn't know the consequences of its actions" and "doesn't have inherent ethics," he said. It can't take in and absorb video, "at least not yet," he added, leaving room for that to change. Even with better models, he expects the systems to stay behind events: "There's always going to be a latency. And so many things can change in a blink of the eye."

His plainest example came from a toddler. A two-year-old in a high chair knows that if it pushes its sippy cup off, mom will come running, and it knows it will crack up. "AI's not going to figure that out ever," Cuban said.

Bartlett's exponential curve

Bartlett pushed back. He said it is tempting to treat the technology as if it were standing still, but with each new release he had watched progress climb steeply, and he described a graph shaped like an exponential curve. "My Tesla drives itself and apparently the crash rate is significantly lower than if a human was driving," he said. If that keeps compounding, he suggested, the results will at some point be dramatic.

Tesla publishes a safety report for its Full Self-Driving (Supervised) system that makes this kind of comparison. Tesla measures miles driven per collision and counts a crash against the system if it was active within five seconds before impact. It compares those figures with manually driven Teslas and with an estimated U.S. average. Tesla itself notes that different reporting criteria, the makeup of its fleet, missed events and the assumptions behind the national baseline can shift that comparison, and that the report counts collisions, not injuries.

Later, Bartlett said this technology "feels so different." He had not expected AI to affect driving, he said, and he now has "I don't know, 25 AI agents on my phone" going through his affairs.

Friction against an AI takeover

Cuban answered Bartlett's curve with one word: "Except." He pointed first to protests against data centers. Then he recalled Why Men Rebel, a book the two had discussed earlier and that he said he hadn't thought about since high school. If AI becomes that dominant, he said, some people will reject it outright. "There'll be communes and kibbutzes where there's no technology."

Others will attack it. Cuban described "adversarial actions," in which a self-driving system sees certain things and reacts the wrong way. Once people learn how to trigger that kind of behavior, he said, "you're going to get people who do that all the time."

He also expects knowledge to be held back from the models. When he talks to hospitals and researchers for Cost Plus Drugs, he said, he tells them they're "crazy" to publish or patent everything, because the information becomes available to every model and they lose control of their intellectual property. He said more people are keeping discoveries as trade secrets, putting them into their own model or selling them. The U.S. Patent and Trademark Office's trade-secret guidance describes the choice in those terms: a patent requires the inventor to disclose the invention publicly in exchange for a time-limited right to stop others from using it, while trade-secret protection depends on keeping the information secret.

The result, Cuban said, is a split in which Gemini, ChatGPT, Claude and Grok are not trained on the same data, and if they all give different answers, that "kind of defeats the purpose in a lot of respects." He added that he is not calling AI useless. "AI is incredible," he said. Using it is "like having four of the smartest people next to you," who will not all agree. The person still needs the skill to ask the right questions of several sources, weigh the answers and reach a conclusion.

Book smart

Bartlett asked what a person's edge becomes when everyone has the same information. Cuban said AI can write a business plan and set up the agents and bots to run it, "but I've got to be the one to say, this is right, this is wrong," and then go out and sell it. "AI is not smart. It's just grinds and grinds and grinds," he said.

He compared it with the student who is "book smart": someone who reads a book, remembers it and gets A's on tests by memorizing everything. "That's AI."

When Bartlett asked for his definition of smart, Cuban said it means knowing where to look for information, pulling it together to solve a problem, and identifying which problem needs solving in the first place. It also means working in real time. "I didn't know what questions you were going to ask," he told Bartlett. "AI is missing all that."

He gave a doctor's example. A physician with AI may have every research report at hand, but if the patient vomits all over him, the AI can't tell him what is wrong. "I have to use my senses," Cuban said, and draw on experience to ask the right question.

Security, brute force and doom

Cuban takes the same view of AI-driven hacking. He said he asks security people one specific question about stories of AI breaking into systems: did the AI invent a new attack? According to Cuban, the answer is no. Attackers take "the same list that every security person has" and, because they can spin up an unlimited number of agents, try every combination. "It's all brute force," he said.

That removes an old assumption in security, he said. Apart from a DDoS attack, which floods a target with traffic, defenders could count on attackers lacking the resources to try everything. "Those days are gone." What gets less attention, he argued, is that defenders can use the same brute force to find their own holes first and go on the offensive against attackers.

"I'm not a Doomer at all," he said. He expects some bad incidents, but "I'm more afraid of what people do because of their concerns and lack of understanding of AI than I am about what AI will do."

What would change his mind

Bartlett asked what evidence would make Cuban decide AI is smart, raising recursive self-improvement, the idea that AI systems could keep improving themselves. Cuban's test is about resources. He imagined a brilliant 12-year-old from "the worst part of the world" with no access to resources, who still sees, figures things out and finds sources of information. An AI that could ask and solve questions with that little would meet his test. But with "more GPUs, more power, better algorithms," with or without self-improvement, "I don't think we reach that point where you can do it with no resources."

Bartlett asked whether Elon Musk's forecasts, that work will become a choice and that AI-powered humanoid robots are coming, are overblown. Cuban called Musk "the best entrepreneur of our lifetimes for sure, but he's a sales guy" who has to raise a lot of money for his projects. The two trade insults on X, Cuban said, but he doesn't begrudge Musk anything.

Bartlett said that in his interviews with AI safety experts and people from the frontier labs, "I hear everything." Cuban took that as evidence for his side. "No two people give you the same answer. That tells you all you need to know," he said. If the destination were clear, everyone would see it. Instead, he said, the industry needs more power and more GPUs than in all the years before last year. "That's not intelligence," he said. "It's just brute force."

A bias he admits

Bartlett suggested early computers had started out the same way. Cuban said it was "the exact opposite." Early machines were barely capable, and it took brilliance to program them in Fortran on punch cards for mainframes that didn't have "a hundredth of the power of your phone." Then came Moore's Law and steadily faster chips. "Maybe that's why I'm prejudiced in this," he said, adding that his experience may hold him back. He still expects AI to get more powerful and more capable, "but as long as there's latency, people have the edge."

Bartlett said this was surprising from someone who built Broadcast.com during another exponential era, when people doubted the internet. Cuban remembered those doubters. When he said his company would stream, or "netcast" as they called it then, people told him, "you're an idiot. I'll just turn on the TV." Yahoo completed its acquisition of Broadcast.com on July 20, 1999, describing it as a service for live and on-demand audio and video.

"I'm not a naysayer on AI," Cuban said, but getting the benefit requires being honest about it. He once said many data centers would turn into pickleball courts. He meant it as a metaphor: for AI to reach the heights people want, he said, "you have to have less power doing more, not more power doing more," because otherwise "you run into constraints like we're running into right now."

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02
  3. 03
  4. 04

Connected ideas and articles

From the conversation

Podcast episodes

The Diary of a CEO

Mark Cuban: The Next Market Crash Is Where The Real Money Is Made!

Episode published This article draws on 0:29–0:44, 32:53–33:02, 33:40–38:14, 41:21–41:43 and 45:15–53:19 (approximate times)

Article history

Updates to this article

Tags