Hacker[1hr Talk] Intro to Large Language Models
Andrej KarpathyYouTube ↗
Further reading 3
When does input become an instruction? →A typo can still be valid code. A data format can turn out to express computation. Follow an input inward and identify who decides how to interpret it.This video helps answer · Where should untrusted input stop?How far is next-token prediction from answering a person? →Separate three questions: how an input becomes a representation, how training changes a network and how a capable text predictor becomes an assistant. An explanation at one level does not settle the others.This video helps answer · When does a text predictor become an assistant?How much does a successful demonstration establish? →A program gives the expected output. Between “it did” and “it will” lie the test distribution, formal assumptions and deployment context. Those boundaries are easy to miss under pressure to ship.This video helps answer · Does running a model explain its decision?
Caption notes
Using Llama 2, Karpathy distinguishes pretraining, assistant tuning and inference, then examines tool use and vulnerabilities such as prompt injection.
Original passages · 6
we don't know how these parameters collaborate to actually perform thatwe swap out the data set now and we train on these Q&A documentsusing tools and existing Computing infrastructure and tying everything togetherprompt injection is about hijacking the large language model giving it what looks like new instructionsit's just a next word prediction neural networkprompt injection is about hijacking the large language model