Exercise 3: Mapping nlp1-10's Three Claims to Where This Course Proves Them — Possible Solution ==================================================================== CLAIM 1: SCALE ------------------------------ Per this chapter's own finding-box, scale is proven in llm1-7 (real pretraining data and parameter counts) and llm1-8 (scaling laws — why bigger follows a predictable, measured curve rather than an arbitrary guess). Together these two chapters replace nlp1-10's own general statement about scale with actual numbers and an actual mathematical relationship governing how performance changes as scale increases. CLAIM 2: SELF-SUPERVISED PRETRAINING TAKEN FURTHER ------------------------------ Per this chapter, this claim is proven in llm1-7, which extends "nlp1-5's and nlp1-9's own context-prediction training signal to an entire corpus, predicting the next token instead of a single missing word." nlp1-10 asserted that LLM pretraining pushes the same self-supervised idea word2vec and GloVe already used further than this course's own chapters did; llm1-7 is where the actual training objective (next-token prediction across a massive corpus) is presented in full, rather than referenced abstractly. CLAIM 3: ONE FLEXIBLE ARCHITECTURE VS. MANY TASK-SPECIFIC PIPELINES ------------------------------ Per this chapter, this claim is proven in llm1-6, which explains "why a single decoder-only Transformer can be steered toward tasks nlp1-6 and nlp1-7 each needed a separately built, separately trained pipeline to perform." nlp1-10 pointed at the real, concrete fact that its own sentiment classifier and NER tagger were separate pipelines; llm1-6 is where the architectural reason one LLM can instead be steered toward both kinds of task is actually explained. WHAT "PROVING IT MECHANICALLY" MEANS, AS OPPOSED TO ASSERTING IT ------------------------------ Per this chapter's own framing, nlp1-10 "asserted all three honestly, but didn't have the room to prove any of them mechanically." Asserting a claim means stating it as true, supported by reasonable justification, without walking through the actual underlying mechanism that makes it true. Proving it mechanically means showing the specific structure, numbers, or process responsible for the claim being true — real parameter/data counts and a real scaling-law equation for "scale," the real next-token training objective for "self-supervised pretraining taken further," and the real architectural reason one Transformer can serve many tasks for "one flexible architecture." The difference is between "this is true, and here's why in general terms" and "this is true, and here is the exact mechanism that makes it true." WHY THIS WORKS AS AN ANSWER ------------------------------ It maps each of nlp1-10's own three named claims to the specific chapter of this course responsible for delivering on it, using this chapter's own finding-box as the source, and explains precisely what distinguishes an honest assertion from a mechanical proof in terms this chapter itself uses.