I have heard it too many times in recent years: “an LLM works just like a human brain” or, even worse, “you are the same as an LLM.”

It has always felt to me like something does not fit, like no, we are not the same, but I had not really known how to argue it properly or undo some of the fallacies I sensed behind the whole thing.

Feynmann

Richard Feynmann used to say that “if you can’t explain it simply, you don’t understand it,” and he was also the kind of person who had hacked his own learning system.

He had learned how to learn and how to be honest about what he did not understand.

The next step for someone like me was: “create a skill with a mental model like Feynmann’s.”

And the next one was to use that skill: “Feynmann, explain to me why an LLM is not a human brain (or whether it is).”

After two hours of back and forth, these are the conclusions Richard and I reached.

Let’s start with the basics: a neuron is not a number

When we hear “an LLM has billions of parameters” and “a brain has 86 billion neurons,” it sounds like we are more or less talking about the same territory. But we are not.

A biological neuron has between 5,000 and 20,000 synapses. Each dendrite computes separately. It is like having thousands of processing cores in a single cell. An artificial neuron has one weight. It adds a number. It applies a function. End of story.

And then there is energy. The human brain runs on 20 watts. A light bulb. A comparable LLM needs megawatts. A million times more.

This is not a difference of degree. It is a difference of category.

“Emergent capabilities” are not what they seem

This is the argument that has been hardest for me to digest.

“Look,” people tell me, “the LLM develops skills it was never taught. That is emergence. Just like the brain.”

Yes, that is true. When a model crosses a certain threshold, new things appear: step-by-step reasoning, few-shot learning. Nobody programmed that directly. It emerged from scale.

But I have had “aha moments” too. There is a moment when you have spent months playing the piano and suddenly both hands know what to do on their own. Or you spend years trying to think in English and one day you discover that you are thinking in English without noticing it.

The difference is that I suffered through that process. I had frustration, pain in my fingers, fear of failing. My brain physically reorganized itself. The LLM neither reorganized itself nor suffered anything. Its weights did not change. A configuration that already existed simply became active, like a pixelated image that suddenly becomes sharp. It is not that the image “learned” to be sharp. It is that the resolution crossed a threshold.

The problem of “grounding”

An LLM can tell you that “fire burns.” It has seen it written millions of times. But it has never felt heat. It has never been burned.

The human brain is tied to the world. Every concept is linked to a sensory experience with real consequences. “Sweet” is sugar on the tongue. “Danger” is adrenaline in the body.

Current AI agents already have sensors, internet access, and the ability to execute code. But having a camera is not the same as seeing. Processing temperature data is not the same as feeling heat.

The difference is not sensory. It is about consequences. A human has skin in the game. An agent does not. If a robot falls, it can reboot. If I fall, I get hurt and I become afraid of falling again.

I am not completely sure about this argument, to be honest…

The real leap: it is not the LLM, it is the system

This is where things get interesting. Because when someone says “AI has improved a lot,” they are not only talking about LLMs having more parameters.

The real leap has been something else: agents.

A pure LLM is a text generator. You give it an input, it gives you an output. End. It has no access to the world, it cannot verify its own answers, and it has no memory between conversations.

But when you add tools, context, and the ability to execute actions, it stops being a text generator and becomes something completely different.

Think about the difference between a brain in a jar and a human with hands, tools, and a connection to the world.

A pure LLM is the brain in the jar. It can “think,” but it cannot act.

An AI agent is the human with tools. It has:

• Tools: It can search the web, execute code, read files, connect to APIs. It does not just predict tokens: it acts on the world.
• Context: It does not start every query from scratch. It keeps history, remembers what you said before, and uses relevant information from external sources.
• Loops: It does not generate one answer and stop. It can iterate: do something, see the result, adjust, and try again. Like a human who tries, fails, and tries again.
• Architecture: The LLM is just one piece of a larger system. There is an orchestrator that decides which tool to use, when to stop, and how to verify.

The value is not in the model. It is in the complete system.

And this is what most people do not understand when they say “an LLM is like a brain.” You are not talking to an LLM. You are talking to a complex system that uses an LLM as one component.

Conclusions

After these reflections, I am left with four key ideas:

1. The brain-LLM analogy is superficial. Both are “networks,” yes. But the differences are categorical: neuron complexity, connectivity, energy efficiency, continuous learning, and grounding in the real world.

2. Emergence does not imply understanding. The fact that an LLM develops untrained skills does not mean it “understands” anything. It is an effect of statistical resolution, not genuine cognition.

3. The real leap is agents, not models. Pure LLMs are text generators. Agents, with their tools, context, and loops, are systems that act on the world. That is where the real value is.

4. Knowing what an LLM cannot do is just as important as knowing what it can do. If you are building AI systems, designing them around their limitations is the key to success.

What now?

I still use LLMs and AI agents every day. They are incredible tools. But I no longer think of them as “artificial intelligence” in the human sense of the term.

I see them for what they are: powerful systems that need to be controlled, verified, and understood.

And every time someone tells me that “AI is going to replace programmers,” I think: “Sure, but right now, not really.”

This is what I have learned from looking into it. If you are interested in the topic or want to talk about how to implement AI realistically in your company, let’s talk.