That's against the EULA if OpenAI may want to make a similar model:
> (iii) use the Services to develop foundation models or other large scale models that compete with OpenAI;
https://openai.com/api/policies/terms/
Seems to be about developing models and not just restricting you from training them with it.
They may not need to.
Either everything is fair game for training, or nothing at all is.
I would absolutely bootstrap a model from OpenAI outputs. What are they going to do about it?
Also there are specific techniques for validating that you are model training procedure is directionally correct, such as generating a simulated data set and training your model on that.
The difference is that now, you are more of a code reviewer and editor. You don't have to sit there and figure out the library interface and type out every single line.
Code-completion > Abstraction.
I love these sorts of loopholes. OpenAI is actively trying to curb the potential of their AI. They know how powerful it is. Being able to see a taste of that power is endlessly exciting.
- Implement a simple ray tracer in C++ using opengl. Provide compilation instructions for macos.
- Create a two layer fully connected neural network with a softmax activation function. Use pytorch.
- Implement the wire protocol described below in Go. The peer wire protocol consists of a handshake followed by a never-ending stream of length-prefixed messages. The handshake starts with character ninteen (decimal) followed by the string 'BitTorrent protocol'. The leading character is a length prefix, put there in the hope that other new protocols may do the same and thus be trivially distinguishable from each other.
- We are trying to classify the expense account of credit card transactions. Each transaction has an ID, a date, a merchant, a description, and an amount. Use a pytorch logistic regression to classify the transactions based on test data. Save the result to a CSV file.
- We are configuring settings for a product. We support three products: slow, medium, and fast. For each product, we support a large number of machines. For each machine, we need to configure performance limits and a mode. The performance limits include iops and throughput. The mode mode can be simplex or duplex. Write a protocol buffer for the configuration. Use an enum for the mode.
- How were George Orwell's works received during his lifetime?I asked for jokes this morning and initially it made excuses and wouldn't give me jokes until I tweaked the prompt.
Later I refreshed the chat and pasted in the original prompt and got jokes right away, with no excuses.
(I was asking for jokes on the topic of the Elon Musk Twitter acquisition. My personal favorite: "With Elon Musk in charge, Twitter is sure to become the most innovative and futuristic social media platform around.")
Perhaps it is even prohibitively expensive.
The number of people using Google Search is easily 1000x larger than the number of people using OpenAI, if not more.
[1] https://ai.googleblog.com/2022/01/lamda-towards-safe-grounde...
I’ll leave it for you to decide which is the pessimistic option.
(Google project pitchfork.)
Knowing that code is correct is as important as the code itself, and this is why we do code review, write tests, have QA processes, use logging and observability tools, etc. Of course the place that catches the most bugs is the human writing the code, as they write it.
This feels like a nice extension to Copilot/etc, but I’m not sure it’s as general as people think.
Perhaps an interesting challenge to pose to it is: here’s 10k lines and a stack trace, what’s the bug. Or here’s a database schema, what issues might occur in production using this?
Ah must be a Spring application ...
This seems like the lowest number that would be useful. Below that it's not really a problem to debug, but at that point there's typically enough complexity that some help would be useful as you forget edge cases and features in the codebase.
For demonstration purposes doing it with 100 lines might be ok, but for professional use it kinda needs to understand quite a lot! Like a minimum of that order of magnitude, but potentially millions of lines.
FWIW, I've never used Spring. My experience is mostly Django, iOS, non-Spring Java, and some Android.
Once I trust the tests, I generally trust the code.
I've seen Copilot generate code I read and thought was correct, that went through code review and everyone thought was correct, that had tests written for it (that nearly covered everything), and that even when it failed, was hard to spot the issue.
It turned out it got a condition the wrong way around, but given the nesting of conditionals it wasn't obvious.
I don't think a human who was thinking through the problem would have made the same mistake at the point of writing, in fact I think that the mind state while actually writing the code is hard to reproduce at any later time, which is why code review isn't great at catching bugs like this.