>the entirety of human knowledge in a couple of gigabytes of data which can be queried in natural languageLLMs are the new magnets. How the fuck does this even work.
Have you tried asking an LLM "How do large language models work?"
>>108893693It is lossy compression you can query in natural language.
>>108893704Yeah, it give me a load of wank about tokenization, weights and embeddings.
>>108893720LLMs being lossy is the exact reason they don't work properly. Everything they shit out is an approximation.
>>108893693>the entiretyobjectively, factually, incorrect. Are you 5?
unlike magnets, you can read the white papers that explain how they workor just ask an LLM.
>>108893693>How the fuck does this even work.Lossy compression and users dumb enough to accept erroneous results after querying the database in a natural language.
>>108893693> whats lossy compressionthink of it as a giant jpeg and the queries giving you a slice view of said image. The final image is going to look rather poorly the higher the compression rate. That was one of the first jokes when gpt3 launched and everyone was hyped ot learn about it and discovered that gzip could essentially do the same albeit with severe limits.
