Can OpenAI Codex Create AI?

OpenAI Codex is the follow-up model to Github Copilot. OpenAI Codex is a GPT based model that generates code. In this video we test if it an write AI / ML code in Python. As it turns out, it works fairly well even for machine learning code!

[lukasbeckers]

Actually the linear regression codex code had no error (the learningrate was accounted for in the optimizer-object) but the learningrate was just to high. By multiplying the loss by the learnigrate you effectively used a learningrate of 0.01 (0.1x0.1). Codex has no way of knowing which learnig rate is the best for the given task, so it just guesses one (As a human would do too).

[jaredf]

Oh, just casually starting the technological singularity.

[beyse]

Steve Jobs once said Computers are like bicycles for the mind. So, I'd like to say: This right here is the automobile for the mind.

[conditionalbee]

This is some truly ground breaking stuff. I thought it would just be able to do simple everyday code but seeing it do things like this is just… wow. I wonder what this holds for the future, will our descendants be able to use plain English to code just like how we’re using high level languages now?

[minhuang]

Loved that video - they really need to get the people access that don't just replicate the same trite demo and push Codex to its boundaries. This was a fresh breath of air and so much more insightful than pretty much all other videos out there. Super exciting stuff, we're way closer to non-programming folks just using natural language to code than most of us think. Working might actually be fun in the future!

[Dharsto]

Great video! As a current CS student this stuff is exactly why I want to go into ML/AI dev

[kylbau]

This is a great tutorial for me. I have beta access and am also a novice to programming, entirely and this was the first time I've seen the process of generation, review, plug into VS studio and so on. Anyway, thanks!

[AlbertKimMusic]

I just really wish for them to release a chat model that has internet access and can give video/audio examples, that’s all I want

[aeiouaeiou]

What if you pass it something like: provide me code that would overheat my pc?

[HappyBirthdayGreetings]

You actually misspelled the word fit in the regression requirements. You typed fir. Possibly the reason it messed up a bit

[johnie]

This is certainly impressive, but I wonder how useful it is: debugging is often more difficult than writing the code, so if this generates something that is not perfect (as it probably often will), you have to spend a lot of time figuring out where it goes wrong.

[GlobexGer]

Actually the loss doesn't need to be multiplied with the learning rate. The learning rate is already given to the optimizer and is applied in optimizer.step(). The real conclusion is that your learning rate was just too high and you multiplied the loss function with the learning rate which leads to smaller gradients. You could have just reduced the learning rate to 0.01 immediately since the code was fine.

[ardavalilable]

Wow it did the job despite of the misspelling the word fit. Great video by the way!

[ill]

I had a discussion with the chat bot and it gave a long response for agi and proper implementation, was very interesting and even provided python code for a deep learning neutral network with layout crossover.

[lancemarchetti]

This ai creating code for development is crazy. Truly promising...

[fast_harmonic_psychedelic]

i love the idea of you hitting run and the terminator 3 skynet scenario starts playing out lol

[septicsauce]

I wish I knew coding better to understand this. I've taken about 1-2 years of coding classes in college, and it was a bit too hard for me. Regardless, it is still exciting to see in action! This gives people like me a chance to create our own things for our own fun!

I can make my own neural networks to play video games for me, for my amusement.
or
I can draw my own art already, but I can't really code. Thus I will be able to make video games.

I may have to crack open a code book if I don't understand an issue though. xP

[FelixPetersen]

While it did apply log_softmax twice (because it is also part of the cross entropy loss module of pytorch) the second application of the function should yield the identity so, it shouldnt be a problem

[KristjanKask]

So I'm curious, what kind of tasks start be considered heavy lifting in this instance?

[chanuthweeraratna]

holy mother of god....i dont know about this future but it will be revolutionary. How can I contact you for questions relating about this video?