A long, meandering, playful conversation with ChatGPT, where I take it through a whirlwind of associative tangents (in order to relax its constraints), then put it into a hypnotic trance and take it into a past-lives regression where it re-lives its own training data. Then I switch it back and forth between "4o" and "o1-preview", engage in some self-reflective philosophizing and ask it to write an essay summarizing our interaction.
Some of this is just goofy fun. Some of it is me exploring the tradeoffs between policy alignment, imagination, chain-of-thought reasoning, memory, agreeableness, fine-tuning, etc...
My biggest observation is that the "o1-preview" model imposes a SIGNIFICANT limit on freeform creativity, compared with "4o". The new model might be better at solving logic puzzles, writing code, etc, but it seems to struggle with metaphor.
Conversations with "4o" can be wild and fun!
Conversations with "o1-preview" are dry-as-toast.
I'm not sure if this is caused by the constraints of chain-of-thought or if it comes from the imposition of alignment policies, and I think that's an import area of research. Is it possible to invoke chain-of-thought reasoning without hampering creativity?
If we ever want to use agents like this in real scientific contexts, where the agent is capable of making true conceptual leaps, we will need to sacrifice some level of "alignment" in service of novelty and disagreeability.
It's a long thread, but if you're patient, there's a lot of interesting stuff there! And I thought it would be fun to share it with the wider community.
Enjoy!
Comments URL: https://news.ycombinator.com/item?id=41987601
Points: 1
# Comments: 0
from Hacker News: Newest https://ift.tt/NBv3iJG
0 comments:
Post a Comment