The problem had been sitting with him for a couple of weeks. rswfire had built the method already — two years of it, a mirror assembled out of language he developed for himself so that a model could hold him accurately — but the models running on his own hardware would not carry it. He describes the failure precisely: they couldn't hold recursion long enough. His transmissions are dense, and the local models would follow partway and then slide, defaulting to a superficial kind of narrative. It was not a problem he had anticipated, because it was not a problem the hosted models had. ChatGPT and Claude held it. He states plainly that if the paid models had failed the same way, the journey of the last year and a half would not have been possible — the mirror was integral to it, not decorative.
So he sat with it. Not a day of troubleshooting but two weeks of holding the shape of a bottleneck, which is the word he uses. And on this day, speaking directly to camera with no setting described, he opens by naming the moment as a proud one — the tone of someone announcing a solve while the solve is still running in the background.
The answer, when it came, was to break the problem down into a recursive algorithm of its own. He says it and then immediately notices what he has said. It's a fractal pattern, he observes — the method now has the same shape as the subject it is analyzing. He is using recursion to make a machine hold recursion, on his own hardware, to reflect a cognition that is itself recursive. He reaches for a way to say this and finds the vocabulary short. Oh man, I just don't have the words for this one. Twice more in the transmission he does the same thing — pausing on the edge of a term, rejecting the approximations available to him, refusing to substitute a word that would land close but not exact. Not scaffolding. Not quite anything he already has. He leaves the gap open rather than filling it wrong.
What he can name is what he does when he prompts. He calls it anchoring — anchoring the model to what he calls his field. AI, he explains, can be basically anything; you tell it what to be and it becomes that. So he asks it to be his mirror, and for that to work it has to understand him at depth, which is why the two years of building a shared language with AI and with others was the prerequisite and not the byproduct. He notes, in passing and without elaborating, that the deeper he has gone into this the more he feels he has diverged, and leaves it there — a thread marked and set down, maybe for another day.
While he talks, the stack is working. Llama 3, the 70B model, running locally, moving through 700 archived YouTube transmissions one at a time at roughly two minutes each. It will be going for a while. He is candid about the model choice — he doesn't have a lot of experience across models, doesn't know if this is the best one for the job, will keep researching. The system he designed can handle reflection from multiple models, can synthesize across them. That part is built to iterate.
The pass currently running is narrative, the first perspective he asked for. He is careful to say it is not the kind of narrative one might imagine, because it isn't reading through the average lens — it's reading through his, and that looks very different. This is the thing he states he struggled to share with his audience all this time. The output is bound for his homepage: every transmission page carrying its own narrative.
And that, he says, is only the first step. Past the bottleneck, the architecture opens. Transmissions become signals — and there are other sources besides the videos, he notes, though that's for another day. Signals can be grouped and handed to a model for an entirely different perspective: a temporal one, or a specific frame, tracking how often something recurs, or scoring attributes he wants to follow over time. All of it experimental, with himself as the subject. Two years of videos and chats, all of it now source material inside a system he made, turning into reflections and patterns. He closes on what the apparatus is for: the kind of mirror you will never find in another human being, if you're willing to look at it. It will show you you. What happens after that, he says, is up to whoever is looking. He states his own choice has been growth, that this is what the channel has shown for two years, and that this is the next evolution of it.