So let’s discuss LLMs
The other day, I was reading something describing an activity involving some people who managed to scrape lots of articles all over the web and found that an astonishing ratio of their data was AI generated in terms of structure, syntax and language techniques. I find that rather disturbing but I can see how the level of abstraction for thinking changing is something that’s happened before.
My blogs aren’t AI generated but I can very well see how I can benefit from using AI for writing. (I won’t lol, it sort of removes my originality through probability functions and matrix math so yeah no i’m good thanks, but only for artistic stuff).
Well think about it like this, we used to initially write code in binary. That’s bringing in a bunch of people with their sole job being to translate code into bits for a computer to understand. In the modern day, we use compilers, plain easy simple. They do the same thing (a lot more that the same thing but let’s just keep it simple and a little inaccurate comprehension). So did the job that involved translating code get replaced? Absolutely lol. However, the way these people who just got replaced thought was very valuable. Since they knew the fundamentals of code, they were very reliable if you could give them information on what you wanted in a very detailed technique.
That’s SWE and LLMs today. These were not my opinions and were from an individual I take great inspiration from. He’s been in the technical field and worked for IBM & Microsoft when they were nothing. Literally, nothing. So he’s lived through generations and he can see this happening again in a way that I cannot, which makes it rather cool.
What is RAG?
RAG stands for Retrieval Augmented Generation. Speaking from a very high vantage point, I want you to imagine prompting an LLM. When you believe to be having a “conversation” with said LLM, you are essentially acquiring knowledge through probabilistic reasoning. The response from said LLM is fine-tuned by word or phrase, called chunk, to complement natural language, hence NLP acknowledgments.
ok LLMs are like ur typical ChatGPT; say i ask ChatGPT what the temperature is in celsius, it’ll polly say something like “The temperature is 10 degrees celsius, you should wear a jacket when going out”, that’s called natural language, basically telling u the answer in plain good old english and in a way u like/understand, the other option it had was to literally spit out exactly what u asked for “10”
here: user: “hey, what’s the temperature in celsius, thanks” natural language: “The temperature is 10 degrees celsius, you should wear a jacket when going out” retrieval: “10”
in rag, we have that 10, and we have ur prompt, but we don’t have that natural language long sentence, so what do we do? we use an LLM to help us get it.
if you understand what I tried to explain then u pretty much fully understand how rag works from a high level, this is enough to at least have conversations with ppl who are knowledgeable in this field, hopefully that helped :)