Thread by @kunlgrg

alright so learned the most used and imp terms in ai space (from @gkcs_)
some of the most heard and common ones are:
- llm: predicts next token from input (yes divides our input into tokens)
- quantization: playing with neural network weights basically

- transformers: also indicates next output token from input but consider it as a core part of llm, it has attention block linked with ffnn and has multiple blocks of these (ex: consider it as a engine for a car)
- fine tuning: teaching our llm specific to our use cases

- vector db: basically grouping the words who has similar meaning in a n-dimensional space (ex: group of all fruit names, group of company names, etc)
- rag: retrieval augmented generation, providing the specific docs/context to llm specific to the user query

mcp: model context protocol, external servers which llm can call to perform actions and get the work done which llm directly can't (ex: booking an emirates ticket, emirates mcp server can book it for the user and can send the details to the core llm)

reinforcement learning: remember sometimes gpt gives us two response and gives us option to select which is more preferable well when we select one we were reinforcing the model which response is good and which isn't and in future you should answer acc to our preference

distillation: see llm's are expensive right, so we can have a small lang model, fine tune it to give response as close as the big llm or just like big llm and help us decrease cost well this process is called distillation

do lmk if i wrote something wrong, i just put down my raw, exact thoughts of what i understood :))
thanks, gonna keep sharing my learnings in public

@threadreaderapp unroll