# RLHF

Concept · [Building with AI](https://ecosystem.samcarlton.com/ai)

Also called: reinforcement learning from human feedback, human feedback, preference training, RLAIF

Reinforcement learning from human feedback. A training step where people compare model answers, and the model is adjusted to give the kinds of answers people [preferred](https://ecosystem.samcarlton.com/preferred-stock). It is a main way chat models learn to be helpful.

**For example:** Reviewers pick the clearer of two answers thousands of times, and RLHF teaches the model to write more like the [preferred](https://ecosystem.samcarlton.com/preferred-stock) ones.

## Connects to (4)

- [Alignment](https://ecosystem.samcarlton.com/alignment): Work to make AI models pursue what people actually intend and behave according to agreed values, such as being honest and refusing harmful…
- [Training](https://ecosystem.samcarlton.com/training): Teaching an AI model by showing it a very large number of examples before anyone uses it.
- [Fine-tuning](https://ecosystem.samcarlton.com/fine-tuning): Extra training of an existing model on a smaller, narrower set of examples, so it suits one task or style.
- [Sycophancy](https://ecosystem.samcarlton.com/sycophancy): The habit AI models have of telling you what you want to hear: agreeing too readily, praising weak ideas or giving up a correct answer…

Map view: https://ecosystem.samcarlton.com/rlhf