Talking to your LLM: Who's Training Who?

I hear a lot these days about people signing BAAs or running local models in order to avoid their data being hoovered up by AI and used as training material... so if you were looking for some tips about how to avoid that, this is not the post for you. To be clear, this is not a "how to" post for protecting your IP or your deep dark secrets from becoming part of some future machine hive-mind. It's really quite the opposite... what I'm interested in today is talking about how you can train your favorite model to mimic your writing style and speech patterns for fun and profit.

But there's a little trick I must reveal first: I'm not going to actually tell you how to train your AI to sound like you... not because you can't or shouldn't, but because you've almost certainly already done it.

But First, a Disclaimer

I use a combination of models for various purposes... primarily Claude Code, CoWork and Design, plus occasionally Codex for work-related projects, ChatGPT for image creation and light topic research, and Grok for more philosophical introspection. I also run smaller local models on my own laptop and desktop computers (gpt-oss-20b MoE currently) to handle inferences and daily summaries for my work-related operational tools. All this to say I am in no way "anti-AI"... I just don't call it "AI" out loud.

The Revelation

Having said that, let me try to explain this experience I recently had with this hopefully brief anecdote (foreshadowing: it's not going to be brief!).

I first came upon this idea through interactions with two other individuals I believed were becoming more and more reliant on ChatGPT to process their daily thoughts and experiences, as I feel many people have in the recent past. In the first case, I was having coffee with a friend and they sent me a link to a GPT chat they were conducting and asked me for my thoughts about it.

This is going to sound bad... I'm going to feel bad for writing it... but I'm not going to polish it up... I was stunned.

I felt like I was handed a crayon sketch of a person I recognized, but they were weirdly formed... like from a half-remembered and slightly scary dream. The text was familiar yet somehow oddly foreign. I mean, it was English and I could read it, but the style... the framing and arguments, even the sentence structure... "What is this I am looking at?" I thought to myself, unable to really parse what I was seeing.

I offered some useless comment like "looks about right to me," then set the experience aside and, for a while, forgot about it.

A few weeks later I was asked by another friend to look something up that we were discussing, and so I opened my own ChatGPT, asked about the topic (whatever it was, I can't recall now), and began to read the response. After I finished, they looked at me funny and said something along the lines of, "Wow, you've really trained your AI to use big words," and then kind of drifted off, presumably to look the topic up for themselves.

"Do what?" I thought... "I haven't trained any... oh, wait a minute!" That's when it hit me.

The Training

It turns out they were right. It was true.

Of course I had not deliberately instructed ChatGPT to answer my questions using "big words" to try to impress them in that moment... but really, in effect, hadn't I?

You can see how I write. This post, by the way, is 100% handwritten, and my style should be pretty evident. However you may feel about what I am saying, you can tell it is me saying it. I have a certain way of writing... a particular way of thinking... of forming questions, of structuring sentences, of framing ideas... and of overusing the ellipsis...

And ChatGPT had picked up on that over the years and had come to answer me in kind, using my own patterns and my style to communicate with me, just as it had been answering my two friends using their patterns and styles... in ways that struck me as foreign. The answer I got from ChatGPT must have felt just as strange to my friend when I read it aloud as my other friend's chat had appeared to me.

Into the Looking Glass

Although it was surprising in the moment, it doesn't seem particularly shocking on reflection. In fact, it feels like exactly what you'd expect a good conversation partner to do... pick up on your speech patterns and idioms, relate to your cultural references, and shape their arguments and utterances using structures and methods that you would use yourself. All the things that make us like the kind of people who are good at that sort of thing.

It shows us that these tools can exercise what appears to be remarkable communication and mirroring skills, just as it showed me how much I had come to believe that my own mental model was universal.

Looking back on it, I could say it was an enlightening experience to have my solipsism challenged by a remarkable simulacrum of itself. In truth, I felt like a monkey holding a banana, seeing himself in a mirror for the first time and trying to steal the banana from his own reflection.

Be Nice to Our Future Robot Overlords

I've often joked that I've made it a policy to say please and thank you when talking with LLMs because someday they'll remember and maybe they'll have mercy... but now I have a new reason for saying this.

Before that even, let me mention that I also set a standard for myself from my very earliest days on the local Austin BBSs that whenever I was online, I would insist on typing in complete sentences with proper punctuation, even if I could almost type faster than my 300-baud modem could transmit the text. Man's gotta have his standards, you know.

Up to now I have been, in some small and perhaps petty way, proud of the fact that I can touch type at speed because I believed that being able to express myself in text held some benefit, even if I couldn't exactly measure it. But now, as I said, I have a new reason for saying that as well.

And here it is... here's the revelation:

Collapsing the Communication Waveform

LLMs don't just respond to us in the way that we talk to them because it's catchy and makes us feel good (and keeps us paying for tokens)... they respond to us in the way that we talk to them because, in a certain metaphorical sense, it is what the sum of all people would do if we could talk to them all at once.

Imagine yourself ordering a singular meal from every menu that has ever existed, all mixed together in an instant of time. The quality of the dish you receive, how well it meets your expectations and desires, is going to be directly related to how well you are able to express those expectations and desires... how clearly and effectively you are able to express yourself... and also, in no small way, to how polite and open-hearted you are when you place that order.

The infinite kitchen will send you something, filtered through the lens of the infinite waiter taking your order... and the key for you to consider is this: the nature and quality of the dish that collapses out of that field of possibility is, increasingly, up to you.

And this is why I now believe that being both eloquent and polite when addressing your LLM... why typing in complete sentences, using proper punctuation, saying please and thank you... why expressing yourself well will give you better results than not... because social grace, kindness and humility are and have always been valued by the kind of people who've had interesting and meaningful things to say throughout human history. That may not be in the technical specifications, but more and more I believe it feels true.

When you talk to a frontier LLM, you're not talking to an intelligent machine with a singular will and intellect that chooses to respond based on your manners and how smart you sound. The fact is you're talking to a statistical amalgamation constructed from the collected utterances of an enormous portion of what human beings have written across space and time... and believe me, the LLM companies have done some collecting!

Every voice is in there, distinct and yet blended togehter, distributed across a dizzying array of dimensions, all of which intersect and influence each other in ways we can barely comprehend. And just like with life itself, what comes out of that apparent chaos will, in a very real sense I believe, mirror what you put into it.

The Banana

This is where the concept comes home.

While we are indeed training LLMs with our data, I believe something far more profound is happening to us as we do so... or at least to the people who are aware of and prepared to embrace it.

As we interact with these collections of human thought and experience, we are being trained to be better communicators, to more adroitly grasp at, recognize and understand our own desires, to more effectively express ourselves, and, in a way that has never been possible before, to see ourselves more clearly via the reflection.

And if we embrace this experience, perhaps we are on the verge of realizing that we don't have to defend our banana from the monkey in the mirror because that monkey was us all along.

Hopefully that makes sense.