I've always hated interfaces that try to complete my sentences for me. It started with suggested replies in email and IM apps. I sure noticed it when they started showing up in the llm chat interfaces and it really bugs me.
This is the one case I don't mind it. Suggested responses feel like they cheapen human interaction, but here I'm talking to a robot that really does tend to know what I want next, and also is highly unlikely to be offended by a less-than-heartfelt response.
How does showing the suggested answers to the user make the conversation better for model training?
They could take any conversation without suggested answers, truncate it to just before a user message, have the model predict suggested answers and then train it on the difference between predicted and actual answers, right?
The idea is that a thread can have many reasonable follow-ups that the user would've accepted, so it is wrong to punish the model for predicting a follow up that is different from the user message, as that prediction could've been accepted by the user if it was given.
RL training, the second phase of LLM training, is based on "I did X, was that good/bad?" and that 1 bit of information is the training data.
So you give the user a suggestion, and the user accepts -> good
You give the user a suggestion, and the user refuses and types something else -> bad (plus some supervisory training data)
The main performance enhancer in LLMs is getting high quality training data. So, first, any extra training data will help. Second this is training data that's directly relevant to their product, and thus higher quality than many other sources.
I'd believe any model provider is mining the shit out of every last customer interaction they can get, not just this.
I started getting prompts about "how is claude doing?" as a separate thing in Claude Code, that I noticed yesterday. So they're (also?) soliciting direct feedback about satisfaction with the session.
And if you do provide feedback, they also collect the session. So it's a way for them to collect prompts, answers and overall grade for how good the answers are.
This isn't really convincing, since you can do this even without showing the prediction at all. Simply ask the model to predict what the user will send, then show the actual next prompt, and done. The only reason to show this would be to influence the user's next prompt, which the article doesn't touch on.
One annoyance I have is the suggested prompt is not a bad idea, but not what I want to do next. But it interrupts me and sometimes I go with it. So I don't think it's a accurate prediction, more like a self-fulfilling prophecy.
They could take any conversation without suggested answers, truncate it to just before a user message, have the model predict suggested answers and then train it on the difference between predicted and actual answers, right?
So you give the user a suggestion, and the user accepts -> good
You give the user a suggestion, and the user refuses and types something else -> bad (plus some supervisory training data)
The main performance enhancer in LLMs is getting high quality training data. So, first, any extra training data will help. Second this is training data that's directly relevant to their product, and thus higher quality than many other sources.
I'd believe any model provider is mining the shit out of every last customer interaction they can get, not just this.
Interesting thought at least.