2 ms·
OpenAI makes ChatGPT 'more direct, less verbose'
- Jimmc414 3y agoInteresting that MMLU and HumanEval scores dropped slightly.
- eastoeast 3y agoI feel they've leaned too far in one direction, and will (hopefully) adjust behavior again soon. E.g. "Create a list of 20 potential names for a startup about Ovens" Chatgpt returned only a single response to this. When prompted for 10 more options, it only return two new options. Seems it is forcing itself not to count properly to satisfy brevity requirements.
- tus666 3y agoThe more I use ChatGPT, the useless I realize it is. It looks impressive at first, until you realize it is like a trick pony. Don't get it to haul your luggage. It gives apparently good summaries of sometimes complex topics. But then you ask it to explain things in more detail...welcome to a special hell of circular reasoning, inherent contradictions, and surprising about-faces that occur even when you ask a complementary question (complementary as in congruent not admiring). ChatGPT: <long verbose answer...> Me: Oh so your saying X implies Y but not Z? (which previous answer did) ChatGPT: Apologies for causing confusion, I misspoke.... (go on to give a worse response now) Me: But your first response was OK, I just wanted to clarify XYZ... ChatGPT: <now gives a response with directly conflicting statements> How many times have you seen this?