Beat GPT 4 & Build the Best AI For You
summary
Shows concrete gains: GPT‑4o scores 81% on FashionMNIST vs a tree model 90% and a ResNet 95%, and a LLaMA 8B fine‑tuned to generate code diffs is ~40% more likely to fix errors than GPT‑4o (with >2x token latency). Claims that a few hundred labeled examples can change model behavior, recommends few hundred labeled examples and use evals as P0, and describes building QTorch to make distributed PyTorch iteration fast on Kubernetes so teams can fine‑tune LLaMA 8B for code diffs and iterate quickly.
AI-written summary · about
chapters (11)
- 0:00Fine-tuning importance
- 1:06GPT-4o limitations
- 2:32Business value of fine-tuning
- 4:17ML velocity advantage
- 6:44Domain knowledge vs general models
- 8:05Fast experimentation
- 11:15Fine-tuning basics
- 12:56Success metrics
- 14:20Infrastructure speed
- 16:47ML agility platforms
- 18:34Key takeaways
related talks
21mBeat GPT 4 and Build the Best AI For You: Mobilizing First-Party Data with Fine-Tuning
21mBeat GPT 4 and Build the Best AI For You: Mobilizing First-Party Data with Fine-Tuning
33mBuild Better AI with First-Party Data and Fine-Tuning
38mIntroducing the New Colleague - ChatGPT
43mState of GPT
01h 00mInside GPT – Large Language Models Demystified
28m🧼 From GPU-poor to data-rich: data quality practices for LLM fine-tuning
01h 05mCan you trust your (large language) model?