Kyle Hill on How ChatGPT works internally

[jedishrfu]

Kyle Hill gets into the internals of how ChatGPT works:

 

[Feynstein100]

Not complaining but I thought we weren't allowed to discuss pop sci in the forum

 

[berkeman]

Not complaining but I thought we weren't allowed to discuss pop sci in the forum

There are some PF forums where it may be appropriate, depending on the subject. I haven't watched the video above, but it sounds interesting. Posting videos as sources in the technical Physics and Math forums is almost never a good idea.

 

[jedishrfu]

This guy is a more serious if not silly science reporter and has some really good video content. He gets into some of the conceptual and technical challenges faced by the GPT team to make it work well. He cites the GPT-3.5 paper and provides insight into how things are done.

 

[Borg]

It's a decent video that gently describes the basics of Large Language Models (LLMs) and how they function. It's not in the PopSci vein of "ChatGPT will end life as we know it!!!"

 

[jedishrfu]

As a programmer, this was along the lines of something I wanted to see. I posted an earlier one in another thread where the presenter went through the GPT paper looking for things. But I haven't seen a followup to it.

Kyle is very emphatic when he says GPT is not sentient and that it doesn't know what its generating and then goes to show the extensive training and statistical analysis being used to select the best choice of words which was quite enlightening for me.

 

[jedishrfu]

Here's another interesting video on using ChatGPT:



The presenter talks about using the temperature setting to change GPT responses and shows some related tools you can use to explore GPT models.