2 ms·
What I would really like to see is an analysis of the weight that individual chunks of context contribute to the final probability for an output. That might al
by benlivengood 6y ago
What I would really like to see is an analysis of the weight that individual chunks of context contribute to the final probability for an output. That might allow for better prompting; when GPT-3 gets the wrong answer it would be fairly obvious what was lacking in the prompting given the context it thinks has the most influence. Also, good prompts would presumably have a higher weight.
- jadbox 6y agoOnce discovered, you could have a kind of outline builder where you provide all the right linking outline context to have it generate appropriating paragraphs around it.