Beyond the Hype: A Realistic Look at Large Language Models
summary
History of LLMs traced from LSTMs to Transformers and models scaling from 120M to ~1T parameters, with a concrete failure mode: GPT-4 solves 10 Codeforces puzzles that predate its training but fails 10 post-training, implying memorization not generalization. Demonstrates a working QA pipeline built with retrieval-augmented generation, LangChain, GPT‑3.5 and a Chroma vector DB over a 2,000‑page PyCharm PDF, retrieving five chunks per query; it answers well but returns some irrelevant chunks and warns of hallucinations and sensitive hyperparameter tradeoffs.
AI-written summary · about
chapters (5)
- 0:00Intro
- 3:01Where are we and how did we get here?
- 13:35Are LLMs "intelligent"?
- 27:29Using LLMs
- 42:29Outro
related talks
30mLarge Language Models and Prompt Engineering for Developers
52mUnleash the Power of Large Language Models (LLMs)
51mLarge Language Models: Friend, Foe, or Otherwise
01h 00mInside GPT – Large Language Models Demystified
28mFine-Tuning Small Language Models with Python
19mSystem for Scaling Large Language Models
51mLarge Language Models for Real-World Applications - A Gentle Intro
40mThe Future of Language Models