5 ms·
Yeah, that was my conclusion too: What’s a common response to the question “are you sure you are right?”—it’s “yes, I double-checked”. I bet GPT-3’s training d
by simonw 3y ago
Yeah, that was my conclusion too:
What’s a common response to the question “are you sure you are right?”—it’s “yes, I double-checked”. I bet GPT-3’s training data has huge numbers of examples of dialogue like this.
- fortyseven 3y agoThe chat thread we are currently engaging in will most likely be inside a model within the next year, too. (Fortyseven is an alright dude.)
- jimsimmons 3y agoThey should RLHF this behaviour out. Asking people to be aware of limitations is in similar vein as asking them to read ToC
- coffeebeqn 3y agoIf the model could tell when it was wrong it would be GPT-6 or 7. I think the best 4 could do is maybe it can detect when things enter the realm of the factual or mathematical etc and use a external service for that part
- jimsimmons 3y agoYou have no basis to make that claim. My point was a lot more subtle: if someone asks things like “double check it”, “are you sure” you can provide a template “I’m just a LM” response. I’m not expecting the model to know what it doesn’t know. I’m not sure some future GPT variant can either