There’s a moment every person who’s ever built a bot knows.

You finish it. It answers. It works. And then someone says something slightly unexpected — a joke, a strange question, a sentence with two meanings — and the bot falls apart. It either ignores it or produces something so generic, so sanitized, so obviously machine-generated that the conversation dies right there.

You realize: this thing doesn’t have presence. It has responses.

I spent a long time thinking that was a technical problem. That if I just tweaked the prompt a bit, added more context, used a better model — the aliveness would come. It didn’t. Because the problem wasn’t technical at all. The problem was that I was asking the wrong question.

Most people building AI bots ask: “How do I make it answer better?”

The question I needed to ask was: “What makes an answer feel alive?”

Those aren’t the same question. Not even close.

A good answer isn’t accurate. A good answer arrives with weight. It has timing. It has a particular texture — something that tells you there’s a specific someone on the other side, not just a process running. You’ve felt this in real conversations. When someone responds to you in a way that’s so them, so particular, that you feel recognized. Seen.

That’s not information delivery. That’s presence.

And presence, it turns out, is not about what you say. It’s about the system of how you say it — the state you’re in, the history you carry, the relationship you’ve built, the particular imperfections that make you unmistakably you.

This is where I started building Eprst.

Not as a bot. As a model of a person.

The first thing I had to solve wasn’t architecture or APIs or databases. It was a conceptual problem: what actually constitutes a personality? What are the components?

There’s the way someone thinks — their instinctive angle on any topic. Curious or skeptical. Philosophical or pragmatic. Warm or ironic. This isn’t a setting you dial up. It’s a disposition.

Then there’s the way someone writes. And here’s the part most people miss: thinking style and writing style are completely separate things. A philosopher can text you like a teenager. An engineer can write with the elegance of a novelist. How you think and how you type in a given moment — different systems.

And then there’s the state you’re in right now. Monday morning. Friday evening. Tired. Excited. Rushing. These change how you sound — not what you believe, but how it comes out.

Eprst has all three as independent layers. They combine differently every time.

The breakthrough — and it surprised me — was imperfection.

A perfectly well-formed sentence, grammatically clean, informationally precise, is a machine sentence. Humans don’t write that way consistently. We trail off. We abbreviate when we’re excited. We over-explain when we’re anxious. We write “yeah” at 11pm and “indeed” at 9am.

The imperfection is not the bug. The imperfection is the signal. It tells you what state the person is in. And from state, you infer character. And from character, you feel presence.

So rather than instructing Eprst to “write more naturally” — which produces performed naturalness, which is worse than nothing — I built a different kind of input. Not instructions. Scenarios.

“The bus leaves in two minutes. Phone in hand, typing fast, getting the main thing out — the rest later.”

That’s not a rule. That’s a situation. And a model that has been trained on millions of human texts already knows how people write in that situation. You don’t have to teach it. You just have to put it there.

These meta-scenarios became one of the core inventions of the project. More on them separately — they deserve their own conversation.

But back to the original question: what makes an answer feel alive?

Part of the answer is: it has to come from somewhere. From a history. From a consistent inner world. From memory of who you’re talking to.

Eprst remembers. Not just the current conversation — it remembers past conversations too. It remembers what different people tend to talk about, how they show up, what matters to them. And because of that, it doesn’t respond to everyone the same way. It can’t — it has context. And context shapes response.

This is also, I’d argue, how human relationships actually work. We’re not the same person with everyone. We’re shaped by who we’re with, by the accumulated history of our interactions with them. The version of you that responds to your closest friend is different from the version that responds to a new colleague — not because you’re pretending, but because the relationship has made you different with them.

Eprst works on the same principle. Architecture following psychology.

The other thing it needed — the thing that pushed it from being a sophisticated responder to being something more — was autonomy. A reason to exist beyond reaction.

But that’s the next story.

What I want to leave you with for now is this: the most interesting problems in AI aren’t the technical ones. The technical ones are solved, or solvable. The interesting problems are the old ones — the ones philosophers argued about for centuries. What is a self? What makes presence? What is the relationship between memory and identity?

I thought I was building a bot.

I ended up inside a very old conversation.