WEBVTT

00:00:00.000 --> 00:00:05.170
In this chapter you meet the large language model, a program trained on a huge amount

00:00:05.170 --> 00:00:10.400
of text whose one core skill is guessing the next small piece of text, called a token.

00:00:10.800 --> 00:00:13.309
A token is not a word and not a letter.

00:00:13.309 --> 00:00:16.847
A short word may be one token, a rare word several, and

00:00:16.847 --> 00:00:20.450
the same text can count differently on different models.

00:00:20.850 --> 00:00:26.198
It writes an answer one token at a time, feeding each new token back in, and because the

00:00:26.198 --> 00:00:31.850
choice among likely tokens involves some chance, the same question can get different answers.

00:00:32.250 --> 00:00:33.980
Training came in two stages.

00:00:33.980 --> 00:00:39.605
In the first, the model practised guessing the next token over a very large amount of text.

00:00:39.605 --> 00:00:43.065
In the second, people wrote examples of good answers and

00:00:43.065 --> 00:00:46.650
ranked model answers, and the model was adjusted to match.

00:00:47.050 --> 00:00:50.677
It only knows what was in its training up to a cutoff date.

00:00:50.677 --> 00:00:55.225
Anything later is not in its knowledge unless the application supplies it.

00:00:55.625 --> 00:00:58.330
It can state wrong things with confidence.

00:00:58.330 --> 00:01:01.808
Researchers argue this happens partly because training

00:01:01.808 --> 00:01:05.350
and testing reward guessing over admitting uncertainty.

00:01:05.750 --> 00:01:09.310
You will learn how a model builds an answer, why answers

00:01:09.310 --> 00:01:12.425
vary, and why a fluent answer can still be wrong.

00:01:12.825 --> 00:01:17.698
To use this chapter, watch the video, read the lesson, and try the lab where you

00:01:17.698 --> 00:01:22.450
press a button and watch a made up table of chances produce different answers.