5 ms·
The title may be misleading. It seems to be based on this quote > InstructGPT is then further fine-tuned on a dataset labeled by human labelers. The labelers
by cochne 5y ago
The title may be misleading. It seems to be based on this quote
> InstructGPT is then further fine-tuned on a dataset labeled by human labelers. The labelers comprise a team of about 40 contractors whom we hired through Upwork and ScaleAI.
It sounds like run of the mill supervised training data creation. Not pre-canning responses as the title may suggest.
- andybak 5y agoI only skimmed TFA but the accusation seems to be that they are adding tons of special cases for well publicized flubs. That seems somewhat unlikely to me although it might be nudging at the truth. i.e. they might be using bad press to help identify areas of weakness and then focusing on those. I guess it boils down to how generalizable and scalable these fixes are.
- changoplatanero 5y agoIs he submitting these questions to their API? Probably they would just sample responses from their logs and then have people write the correct answers for it so that if a similar question ever gets asked again they are prepared.
- andybak 5y agoThat's the explanation I actually find unlikely. It's too specific and really won't scale apart from a tiny sample of the most common questions.
- changoplatanero 5y agoYou find it unlikely that companies would pay people to correct the labels for their ML systems based on their query logs? This is what all the major companies do for their voice assistants like siri, alexa, google etc.
- andybak 5y agoNo. That's not what I'm saying. I'm talking about specific, narrow changes.
- Imnimo 5y agoI think it's based on this quote: >OpenAI evidently employs 40 humans to clean up GPT-3’s answers manually Which feels a bit more ambiguous. It might mean they're cleaning up the answers to serve as future training data, but I think the natural interpretation of that sentence is that they're cleaning up the answers before they are given to the user.
- ImPostingOnHN 5y agoI don't think that's the natural interpretation of that sentence. I guess it just goes to show how subjective the natural interpretation of a sentence can be. Semantics aside, the substance seems to be that, if you publish GPT-3 failures, and the right person sees your publications, that person will (possibly via a team of mechanical turks) submit some narrowly fit changes to the GPT-3 responses which will make subsequent asks of the _exact_same_question_ not return such stupid answers (the same question asked a different way, still will), without solving the underlying issue of the model not actually understanding the question, and thus not actually being as generalizably smart as it is presented to be