However, before setting out on said Sellwood Loop (taking Clinton Street to the train crossing at SE 12th, and Orange Line station, which I'd pass through later on the Max, after following Springwater Corridor to the SE Linn and my favorite Sellwood station, the one with the portals), I made the above video.
Polymaths tend to be good at self-learning, or at least pride themselves on that ability. However I'll be the first to admit a good teacher oft supplies the activation energy, the boost, it takes to get over the hump vs-a-vs some skill or topic area. These teachers come out of the woodwork sometimes, as if sent by central casting (I look over my shoulder for the camera crew that must be getting this).
Along those lines, I'm pretty avid in my studies of techie topics, such as a Jupyter Notebook version that'd give me ways to access AI company APIs or choose some local model.
I learned about Herdr, for getting several agents going at once in a terminal session. I took in a few vids on how query key and value transforms help the contextual semantic balls (the systems) to form based on incoming prompts. The LLM learns to splash around in whatever neighborhood within the embeddings, spinning out strands, or trains, of thought.
However in a more mundane vein, they queue in an orderly fashion as postponed vids, books, articles, movies, magazines... the throttling factor, as always, is my own ability to take in.
One needn't apologize for intake limitations, any more than a car maker need apologize for upper limits on speed vs-a-vs some specific make n model. Of course any model has its limits. I'm glad when I'm somewhat able to keep churning through it all, but my appetite for this and that varies. I seek a balanced diet.
Karpathy is a good tour guide. I watched his presentation at Stanford shortly after his return to OpenAI, wherein he gets an LLM to generate pseudo-Shakespeare, having trained on his corpus, a monkey far ahead of the others in approaching bardic benchmarks. LLMs also generate credible Haikus. Hello Hal.
That last topic, LLM word2vec type embeddings (those crazy diamonds the AI companies seek to press), got me into a NeuralNine presentation on SQLite extensions that turn it into a vector database, suitable for querying about cosine similarity between simple phrases, based on an imported, in-memory model.
