Tinker Mixture of experts

by Samy (Kondasamy Jayaraman) · drawn with Colorfigures by April Zhu

Loading Larkspur’s toy experts model…

Pick a token

“Questions go to the billing team through”

Chapter 3 of 7

What the experts learned

Training kept sending similar tokens to the same rooms. Larkspur’s team asks what its toy’s rooms ended up good at.

This chapter: Do experts split by topic, or by kind of token?

About this chapter

You will learn

  • What the toy’s rooms took most often
  • Why experts split by kind of token
  • What a large model’s authors found

Words you’ll meet

Token
A piece of text a model reads: a word, part of a word, or a mark.
Specialisation
What an expert ends up taking most: here, measured over the toy’s training text.

Step 1 of 4

In training, each kept going to the rooms that helped. Over time each room got its .

Predict first.

What did the toy’s rooms end up taking most?