Skip to content
Sixty Bytes
Newsletter
Deep Dive1 min read

How AI models actually get trained

Thumbnail for How AI models actually get trainedVideo coming soon

Training a large model takes three things. Data, compute and human feedback. Each step has its own cost and its own gatekeepers.

What happened

First, a model reads a huge pile of text. It learns to predict the next word. This step uses thousands of chips for weeks.

Next, people rate the model's answers. The model learns which answers people prefer. This step is smaller but needs skilled raters.

Last, teams test for safety and errors. Then they release it.

Read the transcript

How do AI models get trained? Three steps. Read a lot of text. Learn from human ratings. Test before release. The first step costs the most. That is the deep dive.

Sources

  1. arXiv — Language Models are Few-Shot Learners
  2. arXiv — Training language models to follow instructions with human feedback

The brief, in your inbox.

One email each morning. The stories that matter, and the sources behind them.

No spam. Unsubscribe in one click.

Narration is AI-voiced. Every story is human-checked.