5 ms·
The problem with GPT-3 is it’s too agreeable. It will agree to anything and promise anything - which is not a good idea in this context, to put it mildly.
by thorum 4y ago
The problem with GPT-3 is it’s too agreeable. It will agree to anything and promise anything - which is not a good idea in this context, to put it mildly.
- sixQuarks 4y agoPoster said 5 years, not now
- cookingrobot 4y agoIt’s possible to prime it to take different personas. If you start with a setup like “a conservation with a disagreeable chatbot” and give an example of some unfriendly back and forth it can pick it up and be unfriendly. The recent case of the “sentient bot” at Google was apparently primed to be agreeable.