I Know Kung Fu
LLMs split nature from nurture cleanly. What the model knows is its weights. What it knows about you is whatever you bothered to say.
By Geordie Everitt
The scene runs twenty-six seconds. Neo lies in a reclined chair, a thick cable jacked into the port at the base of his skull. Morpheus watches a monitor. Then Neo's eyes open.
"I know Kung Fu."
The Wachowskis shot The Matrix in 1999, but the fantasy it staged is considerably older. Instant knowledge transfer — skill acquisition without the tedium of acquisition — is one of humanity's most persistent daydreams. Moses descends from Sinai with complete moral legislation. Archimedes receives the principle of displacement in a single flash. The mystic traditions are largely literature about sudden knowing: enlightenment, revelation, the moment of grace. Even the language of genius reflects the fantasy — a great insight isn't built, it strikes.
We have always wanted learning to be faster than it is.
LLMs didn't make that happen. But they did something arguably more useful: they made the distinction between what is already known and what must be taught concrete and visible in a way it has never been before.
The Clean Split
The nature-versus-nurture debate occupied philosophers from Locke's tabula rasa onward, and for humans it remains genuinely unresolved. How much of cognition arrives pre-wired, and how much is written by experience? The empiricists said almost none; the nativists said a great deal. The modern synthesis says: it depends on the domain, and the interaction effects are complicated.
LLMs do not have this problem. The split is clean.
When a large language model is trained, it ingests a vast corpus of text and adjusts billions of internal parameters until the statistical structure of that data is, in some compressed sense, encoded in the weights. This is the model's nature: everything it absorbed before you arrived. Language patterns, reasoning structures, domain knowledge, cultural context — all of it baked into a fixed set of numbers at the end of training.
Then the context window opens. And it opens blank.
Whatever the model needs to know about your specific situation — your project, your constraints, your history with this problem, the decision made last Tuesday that cannot be undone — none of it is in the weights. It cannot be. The weights are fixed. The context window is everything that can move. Your situation lives entirely in what you put there.
This isn't a limitation. It's a feature of the architecture, and understanding it is the prerequisite for using these systems well.
What You're Actually Doing When You Prompt
Sending a prompt is not, despite the physical similarity, the same as typing into a search engine. A search engine retrieves documents that already exist somewhere. A language model generates a response calibrated to the totality of your input. The calibration can only be as good as the teaching.
The model arrives at your conversation with genuinely vast background knowledge and absolutely no situational awareness. It does not know what you do, who you're talking to, what "good" looks like for this use case, or what you tried yesterday that didn't work. It cannot infer these things reliably from a brief query. It will make plausible-sounding assumptions and proceed, because that is the natural completion of an underspecified prompt.
This is why prompts that work contain not just a question but a role, a frame, constraints, examples of the target output, and relevant context that the model couldn't possibly have. Not because the model is incapable — the training is vast — but because capability and situational usefulness are different things. The model can be extraordinarily capable and still wrong, because it's completing a different problem than the one you have.
Every underspecified prompt is a lesson that doesn't happen. You are leaving the model to fill the gaps with statistical plausibility, and plausibility is not the same as relevance.
The Blank-Ish Canvas
Here is where the blank-slate framing needs a small correction.
The model is not a blank slate. That's the point of the training. The weights carry an enormous amount — the cumulative statistical residue of an absurd volume of human writing and thought. It is a canvas that arrives pre-primed with the entire history of everything humans have written down. You are not teaching from zero.
What you are doing, every time, is teaching from your zero. The model has no record of your organization, your vocabulary, your constraints, your edge cases, your failed prior attempts. The prior work is in the weights somewhere, diffused across the parameters; your specific situation is not. The billions of neurons you light up with a prompt are responding to the totality of what you typed. The rest is dark.
This is a more precise version of the blank-slate idea. Not empty — pre-loaded. But pre-loaded with general knowledge, not with yours.
Memory Is a Loading Dock
A common objection at this point: but AI has memory now. Some systems do maintain apparent continuity across sessions.
The mechanism is less mystical than the name implies. What's called AI memory is a retrieval and injection pipeline: relevant context is retrieved from a store — conversation history, user preferences, prior decisions — and inserted into the prompt before the model responds. The model experiences no continuity, because there is nothing that experiences anything. It receives the relevant history as text, just as if you had pasted it in yourself. The result can look like memory. The process is context management.
This is worth knowing because it tells you exactly what can go wrong. If the retrieval step surfaces the wrong context, or nothing useful, the model proceeds without it — and gives no indication that something was missed. The system appears to have forgotten. What actually happened is that the loading dock didn't load.
Memory is a well-organized loading dock. The truck still has to deliver the goods to the window.
The Responsible Party
Neo said "I know Kung Fu" after thirty seconds in a chair. The fantasy was total transfer, zero effort, complete outcome. Morpheus loaded the skill; Neo had the skill. The model of learning here — if that's the right word for it — is entirely passive. The recipient receives.
The reality of working with a language model inverts this. The model arrived already loaded. You are Morpheus. Every prompt is the jack going in, and what gets transferred is determined by what you put on the wire.
The question worth sitting with is not whether the model is capable. It probably is. The question is whether you are a good enough teacher to find out.
That answer belongs entirely to you, every time the context window opens.
Geordie