I Asked ChatGPT To Write an Actor Critic Agent ...

Is the hype around ChatGPT justified? How useful is it as a tool for writing code? It turns out the answer is somewhat complex, and we're going to see why in this video.

Learn how to turn deep reinforcement learning papers into code:

Get instant access to all my courses, including the new Prioritized Experience Replay course, with my subscription service. $29 a month gives you instant access to 42 hours of instructional content plus access to future updates, added monthly.

Or, pickup my Udemy courses here:

Deep Q Learning:

Actor Critic Methods:

Curiosity Driven Deep Reinforcement Learning

Natural Language Processing from First Principles:

Just getting started in deep reinforcement learning? Check out my intro level course through Manning Publications.

Here are some books / courses I recommend (affiliate links):

Come hang out on Discord here:

[GregMurray-uf]

That's a relief! My DS team asked it to solve some of our harder Sr. DS level problems and it failed gloriously. About 2 years ago, I thought we were 10 years away from what ChatGPT can do now though (I'm not in hard AI), so that is a bit disconcerting from a job security standpoint.

[GroxEmpire]

The output of ChatGPT is character limited. You can ask it to continue the unfinished code it has written by saying "Continue"

[JousefM]

Thinking about doing it for a simple neural network - let's see how it does!

[dilithjayakody]

I think another thing to keep in mind is that there's an impact from previous prompts of the chat on the current prompt. This is why it kept trying to include DQN and replay buffers in its responses when asking about Agent57 and NGU. Try using a new chat or tell it to ignore the text from previous prompts. That seems to give better results, at least for me.

[yandychang]

you can ask it to continue from where it left off. it will continue to write code.

[hao]

I did the same earlier. Later some guy told me I can say ‘continue’ 😅

[deanvanaswegen]

You should think of chatgpt like an intern as opposed to an AI code generator. Give it a task check the result and then give it feedback on how to improve. You can reply to it and point out the errors and it will fix them. Often you can run code and paste any runtime errors into the chat and ask for a fix. It definitely isn't as good at more niche technical topics but that would just come down to the dataset being very general. Actor critic agents are probably a very very small subset of the original dataset.

[billykotsos]

14:15 Prompting is really key it seems to getting it to work well !

[budokan]

I am using it the same way. While some experts are inclined to downplay its success, no one has ever made a robot that I can formally ask questions and get reasonable answers in NLP. And I believe OpenAI will take it out of the static training it's trapped in. Soon it will be in a position to actively self-educate on the web and replace today's search engines.

[jonhoye]

When it craps out on the code mid program, you can ask it to continue.

[matrix_multiply]

The other day, it performed value iteration pretty well for me.

[billmurphy]

I routinely get suggestions which use non existent syntax or functionality in a given language. When I ask it about it it just says 'Yes this is wrong, sorry for the confusion'. The level of certainty with which it offers nonsese is comparable to a ponzi schemer.

[yunpeng_ma]

Just wonder how ChatGPT is trained, seems there’s no detailed information available