The Evolution of LLMs: From Imitation to Reasoning
summary
Walmart fine-tunes open models and mixes closed-source ones to auto-extract attributes from ~1 billion catalog cells, adding an auto-eval layer that materially cuts hallucinations (no exact percent given) and using distillation into SLMs plus quantized runs on A10/A100-class GPUs to lower cost. The pipeline uses DPO/GRPO style alignment and on-demand inference patterns (model cascades, tree-of-thought) so harder queries get more compute while easy ones route to cheaper models.
AI-written summary · about
chapters (8)
- 0:00Introduction to LLM evolution
- 1:25Imitation learning in LLMs
- 2:51Alignment techniques explained
- 7:52Reasoning models breakthrough
- 11:31Test time compute revolution
- 19:26Inference vs training methods
- 25:46Case study Walmart implementation
- 28:58Future of LLM engineering
related talks
30mLLMs: Beyond the Hype - A Practical Journey to Scale
42mOpen Source LLMs — Powering AI Accessibility and Innovation
29mHow to roll your own LLM
24mLLM Inference Optimization From Simple To Sophisticated
28mNavigating LLM deployment: Tips, tricks, and techniques
44mThe LLM Revolution: From Inception to the Future of Thinking Machines
12mKeynote: How LLMs Have Changed the Way Organizations Do AI
49mThe Building Blocks of Today’s and Tomorrow’s Language Models