The useful training data is when you clarify your intent, when you tell the model a different approach would be better, when you consistently refactor towards Y and away from X, and so on. The training data isn’t the code, it’s the session transcript. (Anthropic would call this a “distillation attack” against their model, but in this case the model is you!)
replyI would imagine your interactions with it are more important than the output.
replyFunny. I use it through Opencode Go which gives more use than I can use, but didn't realize it was actually free on Zen. Will switch to that I guess
replyEvery lab trains their models with AI generated code at this point.
replyHopefully, 'validated' AI code
replyWhat do you think you're doing when you accept an edit, press thumbs up, or don't ask for modifications after an edit.
replyThats not exactly 'validated'. Feels very noisy, it is not a good bar for either
- does this code do what the user actually asked
- is this code actually 'good'
replyThere would be so many examples of coding projects that these models began or attempted to work in, that were abandoned because the models were floundering.
I would imagine the labs have some decent ways to produce novel requirements and then actually validate they are met, without the noisiness of implicit human feedback.
That said, the more I think about it, you are right, there's probably also very good ways to extract signal for all these sessions.
This is exactly what RLVR is, and the reason that models have improved so much at verifiable domains like coding and math while not so much on unverifiable ones like writing and UI design.
replyTraining on ai generated content is how the models got a big jump in capability
reply