- The model is said to be rewarded/penalized based on answers, and the plain meaning would leave someone with the impression of "answer to the question asked/prompt", the visual appears correct (the chalkboard indicates the question is "fill in the next token" and the answer is "the next most likely token")
- You end up with a strong sense there's 3 tables in the model called key, query, and value that carry the weight of the world. Are there only 3?
- "what part of data are we processing and are they relevant?" - what is they? what part of data is it processing? what is data in this context?
- "how well does this data answer my question?" - so the model is picking out answers from the training data and checking if it answers the prompt? This creates a strong sense of copying verbatim from training data
- "how should we improve the contextual representation of the input data?" - "contextual representation of the input data" isn't clear here, in my dummy brain, it's "here's information we can use to decide the next token: in this context, they meant cat cli, not cat the animal"
- "training a model is like doing Q&A with the neutral networks while it attends to the right data": Is it Q&A? If we double down on that (which means also doubling down on the idea being finding the answer in the training data, I don't think it's a good idea), then there's a gap in how the sentence connects within itself that's worth addressing. I assume "like doing Q&A with the neutral networks, and based on whether the answer is correct or not, it'll adjust and learn to attend to the right data"