Weights vs. Words in Fine-Tuning & Enhancing Agentic AI Workflows
summary
Compares LoRA fine-tuning against smarter prompts and software scaffolding by building a restaurant search engine using a react loop harness with three tools, then running $20 of GPU credits to fine-tune Quen and compare outputs to Claude Sonnet vs. Quen. Uses an A100, R=16 low‑rank adapters to cut trainable params from millions to ~57k, gets mixed wins where telemetry-guided tool calls improve success but does not match Claude Sonnet consistently, so fine-tuning helps only after scaffolding is exhausted.
AI-written summary · about
chapters (13)
- 0:00Weights vs. words
- 1:06AI hype cycles
- 3:51Harness engineering
- 5:55React loop architecture
- 8:21Model performance in search
- 10:28Domain-specific fine-tuning
- 11:52Low-rank adaptation mechanics
- 16:15Fine-tuning hardware/data
- 17:37Datasets for knowledge distillation
- 19:23Fine-tuning with LoRA demo
- 22:30Fine-tuning vs. scaffolding
- 23:14Regulatory considerations
- 24:17Conclusion
related talks
36mPanel Discussion
39mThe Future is Fine-Tuned: Deploying Task-specific LLMs
19mIntro to Fine-tuning
21mFrom Data to Precision: Fine-Tuning LLMs for Proprietary Query Languages
21mBeat GPT 4 and Build the Best AI For You: Mobilizing First-Party Data with Fine-Tuning
33mFine-Tuning Large Language Models with Declarative ML Orchestration
21mBeat GPT 4 and Build the Best AI For You: Mobilizing First-Party Data with Fine-Tuning
48mPracticalities of Fine-Tuning Llama 2 with AI Studio