3 ms·
The article suggests all or nearly all of GPT-3's bad answers get fixed a few days later... This then suggests those 40 people are reviewing every input and ou
by londons_explore 5y ago
The article suggests all or nearly all of GPT-3's bad answers get fixed a few days later...
This then suggests those 40 people are reviewing every input and output for bad responses.
Seems like a lot of work...
- baobabKoodaa 5y agoGPT-3 is likely tagging a small portion of its responses as requiring human review, not all of them.
- ImPostingOnHN 5y agothe ones that get published and get enough attention likely get prioritized