Large language models are not a paradigm shift
summary
Argues that LLMs are not a paradigm shift for computing but a new, powerful instance of familiar ML systems, showing concrete analogies: fourth‑order Markov chains fail where LLMs succeed, yet production needs the same guard rails, dynamic context lookup mirrors feature stores and RAG, and prompt tweaks are really prompt engineering akin to feature engineering. Also presents systems-level wins: caching context like an OS and using speculative decoding (slipstream-style prefill) to make autoregressive inference far faster.
AI-written summary · about
chapters (13)
- 0:00Introduction and thesis
- 2:08What is a paradigm shift
- 7:27Examples from science history
- 12:22Paradigm shifts in computing
- 14:06Perceived differences in LLMs
- 17:23Guardrails in machine learning
- 19:56Context in ML systems
- 21:42Feature engineering comparison
- 24:29LLMs in production lifecycle
- 27:40LLMs as search and recommender systems
- 31:34LLMs and classical ML similarities
- 36:05LLMs as general computer systems
- 44:09Summary and Q&A
related talks
49mA Practical Approach to Using LLMs: Peculiarities and Pitfalls
51mTowards a Pragmatic Understanding of Large Language Models
28mExplaining Machine Learning Models with LLMs
12mKeynote: How LLMs Have Changed the Way Organizations Do AI
32mWhat can we expect of LLMs as software engineers?
14mImproving Large Language Model Performance
33mOpening Keynote: LLMs, between fantasy and reality - the poin...
11mWhen Less Is More: A Technical Overview of LLMs and the Strength of Smaller Models