1:08 GPT Background 2:35 Bigger Data! 3:30 Bigger Models! 4:15 Zero-Shot LM Transfer 5:24 Zero-Shot NLP Tasks 5:57 Question Answering, Allen Institiute Demo 7:22 Transformer Decoder 7:54 GPT-2 Translation 8:20 Larger Models, 17B Turing-NLG 9:20 Misuse: Reddit GPT2 + BERT Bot

Great video - concise, yet covers a lot of material I had questions about!

A much needed video!!! Thanks

@connorshorten6311

4 жыл бұрын

Thank you!!

In just a few years we went from GPT-2 to ChatGPT

Perfect!

@connorshorten6311

4 жыл бұрын

Thank you!!

Hi Henry! I have a question, why is it called "unsupervised"? I mean, the training is performed on prediction of the next token, so why would you consider it unsupervised? The model learns P(x(t+1) | x(t), x(t-1)... x(0)), right?

I had no idea that GPT existed 3 years ago.

I asked for a short story involving harame. What I got however was this. As harambe lay in a pool of his own blood, he spotted the glint of a gun. The ape cried. "How come I, man?" but it would not bear him there. "Well-why, why not? Who are you?' A woman, at once, asked. "I am a slave of Master Cunha. He is my man. He holds nothing to me but my faith. I do not have any slaves, but Cunha keeps them all away from me." He glanced at her, then at Cunha. There seemed to be a sharp tug of his hand. "There is something else!" said he. "No, not the one-" "I suppose that it will be better. We shall get him a chance to give him a lesson now. But if he wants to do it I shall let you know, then I will tell her you are one." "I am sure he shall." The man took his hand. "I am afraid I am not going to come but by your invitation. I have already told him you are of his own, and that you must not speak to him. Give him the opportunity to see you. He has to see you," he ordered, after a pause. "But I don't know," said he. "I'm sure he understands what you are doing and knows what

Jesus. Is there not a GUI for this where I can just drag and drop like 2000 text files into a box and click "train"

heh, Shuttlecock

I believe a man used this to make fake content messages that he claimed were from, it has ruined my life. Is there a way to detect the fake words? Please help

Huh

What type of dog is Henry?

@connorshorten6311

2 жыл бұрын

Maltese haha

@Aieieo

2 жыл бұрын

@@connorshorten6311 I have a pug named Henry

Quick suggestion to slow down and explain more. Seems as though you are quickly reading to us at times. Good topics though.

@jonathancarpenter399

3 жыл бұрын

Just set video to .75 or .50 playback

2:40 "the way this dataset is constructed is a huge contribution to this paper" .... proceeds to show that "women who use vibrators both alone and with a partner experience greater" I think +3 karma was the cutoff? So all it takes is for three adolescents who really need to get outside more, for it to become groundbreaking paper-worthy knowledge. lol

@grownupgaming

Жыл бұрын

All I am saying is lets not have the our future masters and overlords hold reddit to be their bible=)

@nullbeyondo

Жыл бұрын

Training a model on everything would basically allow your model to have an understanding of.. well, everything. Example: reddit is full of typos so the AI could learn if someone is being rude even if the user used "fck" to "fk" to and so on because all these words would be assiocated in its neural network. Even better! You could train your model on only sexual topics and every time a user sends a message (let's say in your Discord server), you could let the model "rate" how sexual it is. And take its probability, if it is very high, you ban them. In a nutshell, to defend against something, you've to understand it first. And no, the AI wouldn't make mistakes. If someone basically said "sex", it wouldn't ban them unless they start talking in a specific way. That's how powerful language models can be at understanding human language if fine-tuned correctly. What's shown here is just a "general" dataset which means it should've everything in it.. and I mean everything. Uncensored data! You could find weirder stuff in its dataset if you searched enough.