I've done a lot of post training and data collection for post-training
I think if you're not OpenAI/Anthropic sized (in which case you can do better) you're not going to get much value out of it
It's hard to usefully post-train on wildly varied inputs, and post-training is all most people can afford.
There's too much noise to improve things unless you do a bunch of cleaning and filtering that's also somewhat expensive.
If you constrain the task (for example, use past generations from your own product) you get much further along though.
I've thought about building a Chrome plugin to do something useful for ChatGPT web users doing a task relevant to what my product does, then letting them opt into sharing their logs.
That's probably a bit more tenable for most users since they're getting value, and if your extension can do something like produce prompts for ChatGPT, you'll get data that actually overlaps with what you're doing.
I think if you're not OpenAI/Anthropic sized (in which case you can do better) you're not going to get much value out of it
It's hard to usefully post-train on wildly varied inputs, and post-training is all most people can afford.
There's too much noise to improve things unless you do a bunch of cleaning and filtering that's also somewhat expensive.
If you constrain the task (for example, use past generations from your own product) you get much further along though.
I've thought about building a Chrome plugin to do something useful for ChatGPT web users doing a task relevant to what my product does, then letting them opt into sharing their logs.
That's probably a bit more tenable for most users since they're getting value, and if your extension can do something like produce prompts for ChatGPT, you'll get data that actually overlaps with what you're doing.