Poems, the Personality Model, and the System Prompt

Your AI agent has a system prompt. It's a few paragraphs of text you usually configure in the settings. Usually it describes how the AI should behave specifically with you. What to do, what not to do, what matters to you, what to keep in mind while doing its work — and so on.

This prompt is taken into account on every request. Even when you type "hey how's it going?", the model receives the entire system prompt along with that request. Every. Single. Time.

But this is not a post about how to configure your agent. This is a post about you. And about me.

Your personal - "non_artificial" - intelligence has a system prompt too. As well as a trained model. And yes - it also has billions of parameters.

My whole life is reinforcement learning. Watched a film, read a book — the model's weights shifted a little. Talked to someone — the model got fine-tuned. Any feedback from the universe nudges the weights a little. Recalculates. Not all of the weights. And usually not really much.

Just as with large language models, I can't read my entire model and fully grasp how it works. A property of emergent systems. Deal with it.

When we are kids - the model is small and has default weights. Learning happens at lightning speed. A single input can flip half of all the weights. Childhood traumas? Welcome! This this way please. Feel free to settle down.

With age the model gets bigger. Havier. The cost of fine-tuning goes up. To really change something, you need either a long time or a very strong impact. A strong impact over a short stretch of time is what we usually call a "crisis".

Who is training your/mine model? Everything.

Any signal dissolves into the weights of your latent space. The whole cultural background around us. Every series we watch. Every YouTube clip. All the music. All the poems.

Speaking of poems.

For a language model, poetry is a strange and very important modality. Poetry works not only on the semantic field but on the very structure of the language model itself. Brodsky (Nobel Prize in Literature 1987) put it well:

There are times when, by means of a single word, a single rhyme, the writer of a poem manages to find himself where no one has ever been before him, further, perhaps, than he himself would have wished for.

Verse writing is an extraordinary accelerator of consciousness, of thinking, of comprehending the universe.

On top of that, poetry is the oldest memory tech: meter and rhyme are redundancy + constraint. Through them, an error-correcting code is built into a text. It's no accident that pre-literate cultures stored their laws, genealogies, and cosmology in verse — so that the text could preserve/repair itself in transmission.

And it it's not only Brodsky, Dante, or Kipling who use this. Advertising rhymes and rhythms have lodged themselves in my head forever. "I'm lovin' it…", "The holidays are coming…" ...coming to settle into the weights of our model for good. Perfect examples of a hack on the fine-tuning system. Loving it or not... you don't have much of a choice.

And no, you can't erase it. Nothing can be erased from the model — you can only lower the weights of individual elements. But they're still there. For good. For good. And you know it.

And yes — the weight of some individual parameters can be elevated. That's exactly why, to recalibrate the weights, I rewatch films or reread books that matter. I want them to keep carrying enough weight in my model.

So what about the system prompt?

It is not the part of the model. It's a set of clear instructions that are always in force. And there can't be many of them. Due to a technical limitations.

Any poem you read dissolves into the weights. But the few you've learned by heart are promoted to the rank of system prompt. Personal system prompt - the only text that's always loaded, instead of being fetched from unreliable memory upon request.

That's why I memorize poems. Only a few. (Remember? - Technical limitations). But I make sure the most important things stay in my personal system prompt intact. Because it shapes everything else — even when some weights in the model start to drift.

An AI can't rewrite its own system prompt. It can read it, it can discuss it with you. But it can't fix it on its own. At least not without you.

AI can't rewrite its system prompt. But you can. That's the difference. And that is your power.

What's in your system prompt?