An idea sketched two years ago assumed expertise-as-licensable-artifact would be a premium-tier product. The economic floor just dropped through that assumption. Worth re-examining what the idea was actually about now that the substrate has changed.
The open-weights frontier is genuinely usable from a home rig now. The bottleneck has moved, it isn't capability anymore, it's memory bandwidth and patience.
A retrospective on a thread of posts about personal AI ownership, what's clarified, what's still murky, and what I'd revise about the earlier framing now that more of the substrate exists.
Meta released Llama 2 last week with a permissive commercial license. Five months after the leak made local inference possible, the legal layer just caught up to the technical one.
If the goal is a working representation of how a specific person reasons in their domain, what's the actual stack? Walking through corpus, model, retrieval, eval, and where each piece breaks today.
The model providers scraped the open internet and called it fair game. The next phase needs an actual market for the data, and the structural pieces of that market don't exist yet.
If you could capture not just what someone knows but how they reason, and run that on local hardware, what would the artifact actually look like?
Capturing how someone writes is not the same thing as capturing what they know. Both are interesting capabilities. The conflation between them is going to cause problems.
Meta released LLaMA to researchers on February 24th. Three days later the weights were on 4chan. The rest of 2023 will be shaped by what happens next.
Four ways the engineering community is talking about AI right now. Two of them are wrong. Here's the framing this site is going to use, and what to expect from the threads that follow.
