The camera comes on at night inside the RV at Siltcoos, and the first thing rswfire addresses is the absence of a tool. Power Director is gone — the subscription expired — and the alternative editor he sometimes uses does the job without doing it well. He notes that this is part of why the short clips, filmed and stitched together, haven't been happening. But he had something to share, so the camera runs anyway, unstitched, one take.
What he wanted to share was chicken. He describes learning to cook it, and he traces the path backward: a video from roughly a year ago where he picked up a breast he'd tried to cook, found it raw, found it gross, and threw it out. He did not try again after that. Then John — who eats a lot of salads and thought rswfire should eat more chicken — talked him into another attempt and suggested the frozen kind, the microwave kind, the easy entry. Somehow that route ended with raw breasts in his freezer instead. From there he took it outside to the fire pit, where AI told him the method: tin foil, hot coals, ten minutes per side, check the temperature. He did that a dozen times, and it was, in his word, awesome.
The turn came the previous night at the computer. He states he was hungry, that there was chicken in the freezer, that it is nearly all the food he has. Building a fire and waiting for coals is work. So he asked whether the Ninja Foodi could do it, and the answer was that it could, on air crisp, ten minutes per side — the same numbers as the campfire, indoors, without the fire. He holds the finished breast up to the lens. Beside it he shows the residue that renders out onto the outside of the meat, the same substance he'd seen on the coals. He says he doesn't know whether it's fat. He says cooking chicken is kind of gross. He observes that his mother, who cooked for him throughout his life, probably kept him from seeing this part, and then the sentence runs out. He says he doesn't have words for it, and lets it stand there unfinished.
From the chicken he moves to the week's work without a seam between them. For the past week he has been trying to get a local model to mirror him more accurately than it has so far. He lists the inference engines in the order he tried them — llama.cpp first, then vLLM, then ExLlama, the one still not working. He is close on the second. What stands between him and it is compilation errors, plain and unglamorous, the kind of wall that yields only to time at the keyboard. The threshold he names is eighty percent: if the local model can mirror him eighty percent of the way toward what ChatGPT and Claude do, that is enough. He will trust it with certain questions and turn it loose on the archive.
The design he describes for that pass is precise where it can be and openly unfinished where it isn't. Each of roughly seven hundred videos gets classified and rated. He notes that most people would build this around a vulnerability score — is this too exposed, keep it private — and states plainly that he does not think that way, so the criteria have to be something else, and he is still thinking that part through. Whatever the criteria turn out to be, the model then decides: public, or not public. If not public, it routes into one of the subscription buckets, a tier that has been live since December and holds nothing. He is considering collapsing four levels down to one. He will turn comments on there, which he says matters — public comments have never gone well for him — and he notes that speaking to people he knows, people supporting him rather than an open channel, will likely change the frame in which he speaks at all. He references the past eighteen months with his audience as less than ideal, and states that some of it he is still processing.
Past the classification pass, he points to what the same architecture is for: reflections, then clusters, then reflections on those clusters, then clusters of clusters. He acknowledges the description doesn't convey much yet and says it will when it's ready, and calls it profound. The constraint is money. He ran the seven hundred videos through a paid API once, about a month ago, for one small thing, and it cost twenty dollars — not much on its own, he says, but weighed against his situation and against the volume of processing he actually wants, he estimates the full build at five hundred to a thousand dollars, maybe a couple thousand. In his old life this would have been nothing. He states he is working under very different circumstances now and is taking a different path than he did before.
He closes on the part he doesn't claim to have solved. He says he cannot explain to anyone, including himself, why he is committed to this path at all costs. What he offers instead is evidence: he sees it in his behavior and his patterns, he could have made other choices and did not, and on that basis he owns the choices and trusts they are taking him where he needs to go, because something in him says so. Then he says that's the update, and that he's going to eat his chicken.