Who are you in this scene?

The scene everyone remembers wrong

There is a moment in The Matrix that most people remember, and most people remember it wrong.

Neo sits in a chair. They plug him into a machine. Ten hours of combat training load directly into his brain. He opens his eyes and says the line everyone quotes: “I know kung fu.” In most people's memory, he becomes a weapon at that point.

That is not what happens. What happens next is that he steps into a sparring program against Morpheus and loses. Then loses again. Morpheus puts him on the floor while the entire crew watches from a screen upstairs.

It is a strange scene, because Neo's knowledge is perfect. He did not learn kung fu, he loaded it. There is no move he does not have and no technique he is missing. And he still loses to a man who learned the same thing the slow way.

Morpheus does not tell him he is missing information. He says something else:

“There is a difference between knowing the path and walking the path.”

I went back to that scene recently, and not out of nostalgia. I went back because I keep watching it happen inside real companies.

What Neo is actually missing

Neo is not short on knowledge. He is short on a different kind of knowing altogether. He knows what to do. What he lacks is knowing when. He has no sense that something is wrong half a second before it goes wrong, and he cannot tell that the move he just selected is correct in the book and wrong in this fight, against this opponent, at this moment.

Michael Polanyi called this tacit knowledge, and put it in one line: “we can know more than we can tell.”

Every commercial organisation runs on it. The salesperson who hears a shift in tone on a call and knows the deal is slipping. The category manager who looks at a forecast and says the number is wrong before she can explain why. None of them can write that down as a procedure. Which is precisely the point: if it could be written down, it could be loaded.

That knowledge gets built one way only, through a loop: attempt, failure, correction, attempt again. What the loop produces is not more knowledge. It produces discrimination, the ability to tell good from nearly good, quickly, without reasoning it out. We call that judgment, and it is what separates someone who knows from someone who can.

The trap in the word discipline

The obvious response here is that it takes discipline. Put in the hours. No shortcuts. Alan Watts argued that this response is the actual problem. Most people who teach a discipline, he said, teach it with a kind of violence, as something unpleasant you simply have to put up with. He preferred another word: skill.

The distinction looks like semantics. It is not. Discipline is something you force on yourself against your own will. It is defined in advance as unpleasant, which is why it breaks. People do not quit because they are weak. They quit because the whole thing was framed as a fight against themselves, and those are hard fights to win over a long period.

A skill is something you acquire. Nobody disciplines themselves into learning an instrument. They come back to it because they want to know what happens next, and the repetition stops being the price you pay and becomes the thing itself.

This changes the question a manager should be asking. Not “how do I get my people to stick with this”, but “how do I make the work interesting enough that they want to come back to it”. Every company that mandated AI adoption from the top did exactly what Watts describes, and got the predictable result: performative compliance, shallow usage and total abandonment the moment the pressure came off.

What AI actually did

Most talk about AI is pitched too big to be useful. AI did not “replace expertise”. It did something very specific: it brought the cost of knowledge down to zero.

Anyone in your company can produce a marketing strategy, a competitor analysis, a spec, working code or a work plan in two minutes. The output will look professional, use the right language and be well structured. What AI did not bring down to zero is the number of rounds.

Now we stop, because the casting is wrong

Up to here I have assumed something, and you have probably assumed it with me: that we are Neo. AI is plugged into the back of our heads, loading us with knowledge, and we walk into meetings carrying expertise we did not earn. But look at the scene again and ask who actually fits the description.

Who arrived in the world holding all existing knowledge without acquiring any of it? Who knows every move, from every school, and has never taken a hit? Who has read every book on negotiation ever written and has never lost a deal?

That is not you. That is the model.

A language model is Neo after the download, exactly. It was loaded, not trained. It has read everything and done nothing. It holds perfect knowledge of consequences and zero experience of one. That is not a metaphor, it is a technical description of what a language model is: a system that learned from text, with no body, no stakes, and not one instance where something cost it anything.

Once you see it that way, the behaviour stops being mysterious. The absolute confidence in a wrong answer. The recommendation that is right in the book and wrong in the account. The reasoning that sounds airtight and falls apart on first contact with a customer. That is not a bug. That is what knowledge without hits looks like.

So who are you? Four roles

• Neo: the part most people try to takeIt happens every time we compete with the model on its own ground: knowing more, remembering more, producing faster. That competition is lost before it starts, not because we are bad at it, but because we picked the one axis where the machine is genuinely better. There is only one Neo in the scene, and the part is taken.
• Morpheus: the part the room needsMorpheus does not know more than Neo. He knows less. What he has is hits. He is the only person in that room who can look at a technically perfect move and say it will not work, and be right. That is your position relative to the model: not to supply knowledge, but resistance.
• The director: the role that is also a trapThe director decides what the scene is supposed to achieve and what counts as a win. In AI terms, that is defining what the system is for and what good looks like. It is a real role, and it tempts you to skip the dojo. You can only direct a scene you could have acted in.
• The crew upstairs: the marketThey cannot see what was loaded into Neo's head. They see one thing, which is who is on the floor. That is your market, and it is not impressed by the download.

The agent you deployed is the purest Neo of all

If you have put an AI agent in front of customers, you have staged this scene literally. You took a model, loaded it with a knowledge base, prompts, playbooks, tone of voice and objection handling, and declared that it knows kung fu. Then you sent it into its first real conversation.

It lost. They always lose the first time. Not for lack of knowledge, but because the real customer was not in the catalogue.

What you did next, reading the transcripts, finding the exact turn where the conversation went sideways, adjusting and running it again, is not optimisation. That is the dojo. You are running the agent through the same loop that built your own judgment, except this time you are standing on the other side of the mat.

Which is also why you cannot build a good agent in a domain you do not understand. The model supplies the knowledge. What you would lack is any way of telling that it is losing.

Three failures, by their right names

• Playing Neo instead of MorpheusHanding over the output instead of applying judgment to it. Mediocre work from a person used to look mediocre. Mediocre output from a model looks excellent.
• Directing a scene you could not act inYour ability to evaluate AI output is limited by your ability to spot when it is wrong. Outside your expertise, AI gives you confidence without expertise, which is more dangerous than ignorance.
• Skipping the loop that builds judgmentAI is most valuable to people who have already done the reps, and most tempting to people who have not. It does not close the gap between experienced people and beginners. It widens it.

What to do about it

• Separate output you can verify from output that needs judgmentCode that passes tests, a translation you can cross-check, a summary of a document you can read: there AI is immediate leverage at low risk. Strategy, pricing and messaging: there the output looks good exactly when it is wrong. The two need different policies.
• Protect your juniors' repsThe grinding work of the first two years is where your senior people come from. Automate it away and the saving is immediate, while the cost arrives in five years.
• Measure rounds, not adoption“How many people use AI” is a meaningless metric. The question is how many rounds went into an output before it was handed over, and how often the first answer was rejected.
• Make it a skill, not a disciplineDon't mandate adoption. Build professional curiosity. People come back to the work when they want to know what happens next, not when they are told to.

Back to the dojo

One detail in that scene is easy to miss. The download was not useless. Without it Neo would have had nothing to fight with. It gave him the vocabulary. What it did not give him was the ability to choose between the moves under pressure, against a real opponent. The dojo gave him that.

The question “how do we avoid becoming Neo” was the wrong question from the beginning. We are not Neo. Infinite knowledge with no experience was never our condition; it belongs to the tool we built. The right question is whether we have earned the second role.

AI is not the enemy, and it is not the dojo. It is Neo, and your job is to be good enough to train it.

Related

• Organisational enablementMost AI projects don't fail because of the technology. They fail because nobody owns the process, the data isn't ready and nobody was trained to carry on. Organisational enablement is the work that makes it hold after we leave.
• Autonomous agentsAn AI agent doesn't just answer questions. It carries out a task end to end: it reads, decides by rules we define together, acts in your systems and reports back. In a project Shahar Peer led before founding Pele, at a medical equipment company, one agent built into the existing lead process lifted lead conversion by 56%.