Karina Nguyen: commitment

Transcript context

…Let's actually get into that. That's really interesting. I want to talk about evals, but let's follow that thread. So talk about how this helped you create Canvas. So when I first came to OpenAI, I really had this idea of, "Okay, it would be really cool for ChatGPT to actually change the visual interface but also change the way it is with people." So going from being a chatbot to more of a collaborative agent, and the collaborator is a step towards more genetic systems that become innovators ultimately. And so the entire team of applied engineers, designers, products, research got formed in the air almost out of nothing. It's just like a collection of people who just got together and we rapidly started iterating with each other. Actually Canvas is one of the... I would say the first project at OpenAI, where researchers and applying engineers started working together from the very beginning of the product development cycle. And I think there's a lot of things that we have learned on the way, but I definitely came with the mindset of, "We need to do a really rapid model situation such that it would be much easier for engineers to work with the latest model possible, but also learn from user feedback or early internal dog food. How do we improve the model very rapidly?" And it's really hard to kind of like figure out how people... when you deploy a product, how people would be able to use it. And so the way you synthetically train the model is physically figuring out what are the most core behaviors that you wanted the product feature to do. And for Canvas, for example, it came down to three main behaviors. It was how do you trigger Canvas for prompts like, "Write me a long essay," when the user intention is mostly iterating over long documents? Or, "Write me a piece of code," or when to not trigger Canvas for prompts like, "Can you tell me more about President..." I don't know, some of the general questions. So you don't want to trigger Canvas because the user intention is mostly getting answer, not necessarily iterate over the long document. The second behavior is how do we teach the model to update the document when the user asks? So one of the behaviors that we taught the model is actually have some agency and autonomy to literally go to the document and select specific sections and either delete it or edit, so highlight it and rewrite certain sections. Sometimes the user would just say, "Change the second paragraph to be something friendlier," and we would have to teach the model to literally find the second paragraph in the document and change it to a friendly tone. So basically you teach both how to trigger edit itself, but also how do you teach the model to get higher quality edit for the document? In case of coding, for example, there's also the question of how good the model is of completely rewriting the document, versus having a very specific target edits. So that's another layer of decision boundary within edit itself is, "Let's select the entire document and rewrite completely, or do you want to have a very targeted custom behavior." fic target edits. So that's another layer of decision boundary within edit itself is, "Let's select the entire document and rewrite completely, or do you want to have a very targeted custom behavior." And when we first launched the model, we would bias the model towards more rewrites because we saw the quality of the rewrites were much higher. But over time you are shifting based on user feedback and what you're learning from iterative deployment. Lastly, the third behavior that we taught synthetically the model is how to make comments on any document. So the way we used that is we would use o1 model to seem a way of user conversation, let's say like, "Write me a document about XYZ." But then we used o1 to produce the document and then we injected user prompt to be like, "Oh, make some comments, critique my piece of writing or critique this piece of writing that you just made." And then we taught the model to make comments on the document on very specific [inaudible 00:17:45] So it's also what kind of comments you want the model to make. Do they make sense or not? How do you teach the quality of that? And it all came down to measuring progress via very robust evals. But, yeah, this is how you used o1 and a synthetic data generation for the training.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.