GPT-4: What, Why, How?

preview_player
Показать описание
OpenAI's GPT-4 makes significant improvements over GPT-3 and 3.5, but does it live up to the hype? It will soon support images, and is already multi-lingual, supporting context lengths up to ~32,000 tokens. It works great but we know next to nothing about it. I also talk about how to get access to GPT-4, along with a demo using ChatGPT.

Outline
0:00 - Intro
0:50 - New Features
1:42 - Evaluations
5:32 - Multilingual
7:07 - Images
8:19 - Hallucination
9:19 - Training & RL
11:40 - Malicious Use
14:40 - How it Works
18:20 - Getting Access
19:56 - Demo
24:06 - Script Generation
26:51 - Programming Demo
32:58 - Takeaways
34:29 - GTC & Raffle
35:57 - Wrapping it Up

Social Media

Sources:

GTC sessions I'm attending:
Рекомендации по теме
Комментарии
Автор

Among the big pile of clickbait GPT-4 posts, yours was truely informative like a fresh breath of air. Thanks.
Keep up the quality work 👍

monajmilan
Автор

OpenAi should change their name to ClosedAI

KeinNiemand
Автор

Hi there, I am a Native german speaker and used Chat GPT all the time in German. So it is absolutely capable of writing in German.

Even more complex topics like Law was no problem. I also tested it, with some one I know in its knowledge about tax law in Germany. It made some little mistakes but got the overall stuff pretty right. Alls this was of course in german.

sir_no_name
Автор

bro I literally left a comment on the previous video asking for a GTP-4 video, 5 minutes before you uploaded this one lol

Trevorsouloriginal
Автор

I'm surprised to see GPT-4 out already.
I kind of expected it to also be able to generate images, but just being able to read them as input is already a huge step to make it usable for more use cases.

sebastianjost
Автор

Man they really got me with the image input. When I heard about that I immediately purchased GPT+, only to then hear (burried deep in their developer livestream) that image input won't be released any time soon. I want my $24 back 😭

xntumrfoivrnwf
Автор

Great video, thank you!
As a side note: of course, GPT-3 was also multilingual (I've done testing a long time ago). Even though the training data was mostly English, it wasn't limited to English.

ilianos
Автор

Max responses are down to 50 now every 4 hours

metaphysicalfuck
Автор

The "if __name__ == '__main__'" thing in Python is so that you can import the file from another, without running the logic at the bottom of it.

TankorSmash
Автор

referencing the title
what: bigger model, more data
why: more money
how: we don’t know

Graverman
Автор

I remember a few years ago being impressed by openAI's desire to be open as should every research be.

Honestly now I hope they crash and burn and people that actually care about openness take their place

pulsarhappy
Автор

It's a good thing that OpenAI is transparent with their limitations even with the release of GPT-4. The AI is improving but we still can't completely use AIs as a reliable sources until it has an almost 0% error margin. I think for now we can stick to other AIs that we can utilize for automation. Like Bluewillow for images or robotic process automations and such.

WilfEsme
Автор

I think at 3:50 you misinterpreted the graph, I'm pretty sure the dark green represents the improvement from the no vision to w/ vision

HatsuneSquidward
Автор

I already expected multi-modality considering palm-e but gpt 4 is not proto-agi or anywhere near agi it cannot self-improve or do anything alarming. gpt 5 will be crazy though it would probably be what decides how far llms really will go until the technology inevitably plateaus (only improves when computation improves no crazy architectural jumps).

ethanwmonster
Автор

its funny that "openai" went full closed once they started milking the gpt cash cow. but fortunately for people open alternatives are beeing worked on, such as this llama model or whatever. i wont mention the other garbages that try to mimic chatgpt and fail to align 5 words into a coherent sentence.

genkidama
Автор

You can somewhat guess the size of the models by the API pricing.

LetondAtreides
Автор

Coding with GPT4 will be such a great experience ! (hopefully...)

billykotsos
Автор

Personally, I think they did great with GPT-4. Its practical utility has gone up considerably... And Microsoft is going to be adding it pretty much everywhere...

Nvidia stock should skyrocket from the massve increase in the need for their GPUs - hope the supply can keep up with the demand.

LetondAtreides
Автор

I wonder how it would do at writing the docstrings for your code (28:15).
GPT-3's pretty good at writing boilerplate, but not very good at understanding how to use obscure libraries (I was playing with godot-rust). It did pretty well considering that it couldn't compile anything.

nathanbanks
Автор

ChatGPT (3.5) is quite ok with obscure languages. Example from real life :
My daughters highschool had spring ball week ago, invitation and information (of course..) on facebook, written by someone in roughly same age group (and i suspect, non-native speaker).
I had to read it multiple times to understand most important parts. Then asked my daughter to confirm that i got important parts ok. As i was playing around with ChatGPT that time, i did copy-paste and asked it to produce more formal version of it, with better structure.
It did it OK .. in English. Then i asked for translation back to original language. First try was exact copy of original text (cheater!). When i pointed it out, it did what i wanted, good structure, grammar etc. (my wife is teacher, she said it was a-ok:)
Not bad for language with ~1 million native speakers ..

matikaevur