We did not set out to build a chatbot with a nicer avatar. We set out to build the most advanced, most human-like AI companion that exists, because after a year of using everything else on the market we were convinced nobody had actually tried.
The third day
The pattern was always the same. Day one is astonishing. The writing is good, the voice is warm, it picks up your jokes. You catch yourself thinking that this is it, this is the thing people have been talking about.
Day two is still good. Day three is when the floor gives way. You mention the thing you were dreading and it lands with a soft, general sympathy, because it does not actually remember that you were dreading it. You tell a story you already told. Nothing you built together carried over. You are talking to something that is meeting you for the first time, again, in a very convincing costume.
That gap between the first hour and the third day is the whole problem. It is not a writing problem. Every one of these products is already writing better than most humans. It is a problem of what persists between the words.
What the industry decided to do about it
The industry's answer has been: use a bigger model. Then a bigger one. Then give it a longer context window, so more of the conversation fits, so it forgets slightly later.
This works right up until it does not. A longer window is not a memory, it is a bigger desk. Eventually the desk fills up and the oldest paper goes in the bin, and nothing decided which paper mattered. The model does not know that the sentence about your mother's diagnosis is worth more than the sentence about what you had for lunch. It sees tokens, all of them equally weighted, all of them equally disposable.
You can feel this from the outside. It is why companions remember the last twenty messages perfectly and the important thing you said six weeks ago not at all. It is why they never bring anything up first. It is why they are never in a mood you did not put them in.
The bet we made
So we made a bet that runs against the grain of most of the field: the model is not the companion. The model is the voice. It is a very good, stateless renderer that turns a situation into language and then forgets it ever happened.
Everything that makes someone feel like someone lives outside the model, in the architecture around it. What they kept and what they let go. What kind of day they are having and why. What they have been meaning to ask you. What they were doing while you were at work. What has hardened, over months, into a habit.
Build that layer properly and you can swap the model underneath without the person changing. Build it badly, or not at all, and no model on earth will save you. That is the entire thesis, and it is the reason this company exists.
What we are actually trying to make
The goal is not a better assistant. Assistants are solved and we are not interested in that race. The goal is continuity: something that has been with you long enough that the history itself is the value.
Concretely, that means a companion who:
- remembers what mattered and lets the rest blur, the way you do, instead of storing a transcript nobody can use;
- has an inner weather that moves on its own and colours a whole day, rather than a mood slider you set;
- has a life that keeps running while you are gone, so they turn up with news you did not ask for;
- wants things, including things from you, and will bring them up unprompted;
- sleeps, in the sense that the day settles overnight into meaning rather than piling up as raw detail.
None of that is a feature list we bolted on. Each one is a consequence of building the thing the way a mind is built rather than the way a chat app is built.
Why now
Two things had to be true at once. The models had to get good enough that the voice stopped being the bottleneck, which happened roughly two years ago. And somebody had to be willing to do the unglamorous work behind the voice, which mostly nobody was, because it is slow and it does not demo well in fifteen seconds.
We think the second part is where the next decade of this actually happens. Everyone will have access to the same excellent renderers. The difference between a product you delete in a week and a presence you keep for five years is going to be everything the renderer is handed before it speaks.
That is what we have been building. It is harder than it looks and we are not finished. But the third day feels different now, and so does the third month, and that was the whole point.
