Transcription
A father and son just mathematically proved that an AI agent will never do what Silicon Valley is promising. Not probably won't. Not might have limitations. They've mathematically proved they use computational complex theory that's been settled since the 1960s.
And this isn't coming from some AI doomer clickbait journalist. This is coming from Vishel Sika, former CEO of Infosys, board member at Oracle and BMW. He's a Stanford PhD who literally studied under John McCarthy. He's the guy who coined the term artificial intelligence. He and his son just published a paper that no one in AI marketing departments wants you to read, especially right now as we enter the era of Magnus and Open Claw, the agents that can use your browser and click buttons for you.
It looks like AGI has arrived, but Sika says we're actually just watching the ceiling get higher, not disappear. Their argument is simple. LLMs can only perform a certain number of computations per response. That number is fixed. And if a task requires more computation than that ceiling allows, the model will either fail or hallucinate. And this isn't a maybe. It's baked into the math.
But if the math is so broken, then why are the big players still promising the world? I'll tell you the devious reason why at the end of this video, but first, I want to look at the ceiling that they discovered. Now when you send a prompt to chat GPT or cloud or Grock or any of the current frontier models, the model will do a fixed amount of work to generate each word as an output. This happens through the self attention mechanism. This is of course very simplified. But think of it like this. Every word in your prompt needs to look at every other word to understand the context. So if you have a thousand words, it's a million comparisons. A thousand a thousand. But there's no let me think about this harder. There's no give me more time on this one. Every token gets the same budget. A simple hello gets the same number of operations as a complex physics problem. That's the ceiling. It's not about better hardware. It's about the architecture of how the systems actually work.
The paper and I have it here on screen if you want to read it. It uses traveling salesman problem as an example. To visit 20 cities and figure out the shortest possible route between those cities, you need to check over two quintilion combinations. An LLM physically cannot do that math in one shot. So, what does it do? It guesses. It pattern matches. It gives you something that looks plausible and it's not a bug. That's the architecture. But how would you actually handle tasks that require that level of computation?
Next, I'll show you why even verifying the answers is just as impossible for these models. The authors of this paper make a distinction. Doing a task versus verifying it. Now, you'd think that the model could at least check if the answer is right, even if it can't handle the computational complexity to calculate it. But no, verification often requires just as much work as solving the problem up front. Every AI demo you've ever seen, it was running tasks designed to stay under the necessary complexity ceiling. They work because they're designed to work. Meanwhile, the real world tasks that your business actually needs are going to blow right past that ceiling.
And this is where Sika's background becomes a factor. This isn't an outsider's perspective. Remember, he studied under John McCarthy, the man who literally coined the term artificial intelligence. He's bridging the gap between the foundational laws of the 1960s and the chaotic world of AI in 2026. He isn't saying these tools are useless. Far from it. He's just saying they're being marketed as reasoning engines when the math proves they're actually pattern mirrors. They reference the time hierarchy theorem. Again, I don't mean to throw so many fancy words, but this basically says that some problems require a minimum number of steps. You just can't shortcut them. And the argument that the paper makes, if a task needs more steps than the model can perform, it will unavoidably hallucinate. Unavoidably. And this is why hallucination isn't a training issue. Yes, more recent models have gotten better at it, but for certain problems, hallucination is the only possible output.
But wait, you might be thinking, what about the new agentic era? Tools like Manis or OpenClaw. They don't just give one answer. They run thousands of loops, browsing the web and thinking through step by step. The tech community is calling this chain of thought or agentic workflows. And the idea is that if a model has a ceiling, just spread the problem across more steps. Give it more room to work. But Sika's paper argues this as a trap. And here's why. If you have a fixed amount of thinking power per word, giving the AI more steps is like giving a writer more sheets of paper. Each individual sheet is still the same size. You haven't made the writer smarter. You've just given them more room to ramble off topic. That's why you'll see an agent book a flight perfectly, but then get stuck in a bizarre infinite loop trying to change a seat assignment. The math, specifically again that time hierarchy theorem, says that for complex problems, errors eventually compound. The model goes off track at step five, and because it can't mathematically verify its own logic, the whole chain eventually falls apart. In the Agentic era, hallucination isn't a training bug. It's a cumulative mathematical certainty.
Then of course you might be arguing, well they can just use a tool, give it a calculator. After all, we wouldn't expect a human to be able to calculate the traveling salesman problem by hand. But Sika acknowledges this as well. You can build components around LLMs to overcome the limits, of course. And then the LLM becomes an orchestrator. But notice what just happened. The LLM didn't solve the problem. It just handed it off to a classical algorithm that could. But the catch, the model still has to verify that that tool worked. And if verifying correctness requires more math than the model can do, again, the agent fails in unpredictable ways.
Well, what about those massive context windows? Gemini 3 Pro can see a million tokens at once. Yes, that solves information access. It doesn't solve the computational steps per word. Having a bigger filing cabinet doesn't help if you don't have the brain power to process what's inside.
So, what does this mean for you? Now, the paper, it's not saying that AI is useless. Indeed, it it definitely is not. I use these tools every day in my business. I'm sure most of the people watching this do as well. For the right applications, current AI, the current frontier models are exceptional. Writing drafts, summarizing, reformatting data, research, and comparison. These stay under that ceiling. The problem is the gap between reality and the pitch decks. AI agents will autonomously run your business is a lie. The math just doesn't support it.
To see this in action, look at vending bench 2 from Anden Labs. This is the 2026 gold standard for testing AI agents at running a business. Models like Claude Opus 4.6 Gemini 3 Pro. They're given $500 in a year to run a simulated vending machine business. And on paper, the agents look like they're winning. The current leader, Claude Opus 4.6, netted $8,000 in profit. Here's that test, Vending Bench 2. Feel free to look it up yourself. And here are the current standings for Frontier Models. We can see Claude Opus 4.6 $8,000. Pretty good. But here's the actual ceiling. And in labs calculated a human baseline for this exact same simulation. Let me scroll down and show that to you. It's a long paper here. This isn't the best ever, $63,000 a year. This is a human baseline and it blows the AI models out of the water. The reason the AI models can't make $63,000 a year is because they lose coherence over a long time frame. Result, the frontier models, the best we can make now, aren't hitting even 15% of a human baseline. Over these runs, we've seen agents honestly give away their inventory for free due to social engineering or they've even tried to contact the FBI to report their own $2 bank fees as fraud. And this is the time hierarchy theorem in the wild. As the chain of tasks gets longer, the AI's ability to verify its own logic collapses. It doesn't matter how smart the model is. The math says that without a human to reset the error rate, the autonomous chain will eventually break.
So here's what you actually do to stay on the winning side of this math. First, be specific about tasks. [music] Draft an email using my tone and cadence that works. automate this workflow is going to fail. Build in human verification. This is a structural requirement, not an option. And third, use AI for pattern recognition, not logic heavy math.
But here is the real tip off. Why the singularity probably isn't as close as people keep saying? Because if the singularity were just months away, why are the smartest people in the room quitting? Look at the insiders. If open AI was about to hit AGI, why would senior engineers be leaving to start risky startups? If you knew the world was about to change forever, you wouldn't leave. You wouldn't leave Open AI if they're on the verge of AGI. You'd stay to be part of the release of a lifetime, to be part of the equity of of a lifetime, unless you saw the ceiling. Now, they know the next model will be better, but not qualitatively different. Just like chat GPT5, it was better than four, but not qualitatively different. They're starting companies that use AI as a tool, not companies that use AI as a god. The opportunity here is not chasing some imaginary AGI. The opportunity is an understanding exactly what AI can do for you right now. The ceiling is real, but there's a lot of room underneath.