Chapter 1
Gathering the raw digital ingredients
Chapter 2
Washing and prepping the data
Chapter 3
Slow-cooking in the neural oven
Chapter 4
Taste-testing and seasoning
YOUR GOAL
Master of The digital cookbook for hungry algorithms
Ch 4 · Taste-testing and seasoning
007skaty AI Recipe Hallucinations
No fake food, no fake food, Ground the recipe to keep it good!
007skaty Michelin star taste tests show how human feedback improves artificial intelligence
007skaty Like chefs relying on critics, AI developers use Reinforcement Learning from Human Feedback to turn raw text into helpful assistants.
007skaty Google fine-tuned its Med-PaLM AI model to pass professional medical licensing exams
BETA
007skaty Drag nutritious data ingredients to master the recipes that keep digital AI families healthy.
Fun Fact
InstructGPT triumph: A tiny model tuned with human feedback beat a model 100 times larger.
Quiz
What does the acronym RLHF stand for in the context of training and aligning modern artificial intelligence models?
007skaty Actively balancing datasets and testing models worldwide ensures inclusive digital outputs for everyone
Quiz
What percentage of the time did human judges prefer the outputs of the smaller, RLHF-aligned InstructGPT over the massive GPT-3?
Quiz
What is the process of taking a pre-trained general AI model and training it further on a specialized dataset called?
Quiz
Which specialized medical AI model, developed by Google, was fine-tuned to answer professional-grade clinical questions?
Quiz
What is the term for the process where users trick an AI into bypassing its safety guardrails to generate harmful content?
Quiz
Which AI safety technique, introduced by Anthropic in 2022, trains models using a written set of ethical principles?
Quiz
What is the term for when a large language model confidently generates false or fabricated information?
Quiz
What does the acronym RAG stand for, representing a technique that lets AI look up external facts in real time?
Quiz
What is the term for systematic, unfair favoritism or prejudice inherited by an AI model from its training data?