• training data: to train the network
• validation data: to select the best model/fitting strategy
• test data: show how much error you have on it to not bias the model any choices (to give information on how the model actually performs)

Finding training data for AI is hard. So instead:

Intentional training data

• curated for training data
• Spent time thinking about bias, control, etc.

Training set of convenience

• Dataset that just comes about
• Problematic:

Accidentally introduce bias into the data: Googling images of CEOs, which is convenient, results in all white males for a bit.

“What we have been building since ChatGPT at H4.

• No pretraining in any way

Basic Three Steps

Goal: “helpful, harmless, honest, and huggy” bots.

1. Retraining step: large-scale next token prediction
2. Incontext learning: few shot learning without updating parameters
3. “Helpful” steps
   1. Taking supervised data to perform supervised fine tuning
4. “Harmless” steps
   1. Training a classifier for result ranking
   2. RLHF

Benchmarking

Before we started to train, we have a problem. Most benchmarks are on generic reasoning, which evaluates 1), 2). Therefore, we need new metrics for steps 4) and 5).

the transformational generative syntax is a linguistical precept proposed by Noam Chomsky which has the interesting conclusion that meaning is supported by structure, rather than the other way around as generative semantics suggests.

This means that you can first come up with generic, independent structure to a sentence, then fill in the sentence with meaning.

For instance, “colorless green ideas sleep furiously” is a sentence Noam Chomsky proposes to have perfect structure but failes to be filled with meaning, supporting the transformational generative syntax theory.