RAG cheat sheet: under ~500 pages, skip rag and stuff it all in context (cache the prompt). got keywords? use bm25. retrieve top 20 chunks, rerank, and have an llm add context to each chunk first. #ai

start simple. a single llm call with retrieval and examples is usually enough. add agentic complexity only when it demonstrably improves outcomes. #ai

Gradient descent is a blindfolded treasure hunt. you stand on a hillside, feel the ground for the steepest way down, and step. the hill is the loss landscape, the steps are parameter updates. #ai

i built a chatbot over all lex fridman podcast episodes (1 to 325). ask a question, get an answer plus sources, and click to play the video from the exact moment it's discussed. #ai github.com/qniksefat/lexitalk

readme tip: gif > screenshot > text. and make the 'why' section a story: i had problem A, tried B, it failed because C, so i built D and now i do E with ease. #ai

to get unbiased data, find channels with a higher bar. If it’s instagram -> total BS YouTube a bit better A survey paper -> good A notebook on kaggle -> awesome The data itself! best. #ai

machine learning is the high-interest credit card of technical debt. quick wins aren't free: boundary erosion, entanglement, hidden feedback loops, undeclared consumers. (classic google paper) #ai