What differences do you notice in performance or features? Which one do you think is better?
3 Likes
002 is better, more human, that what I found out
3 Likes
It is more censored, often interrupts its answers on the first word or in the middle of a sentence. Of course, there is a slight hint of NSFW in my queries, but then what is the point of content filters then?
4 Likes
Even in Korean, the 002 model exhibits more human-like qualities in terms of interaction compared to existing models. While Claude 3.5 Sonnet was previously considered the most capable and human-like AI for Korean, the 002 model now demonstrates more natural processing in some Korean language tasks.
However, I believe the pro exp 0827 model excels in long-term and extended reasoning. Because the 002 model generally prefers short responses, the 0827 model remains more suitable for tasks requiring extensive reasoning.
Translated by Gemini.
I agree. The exp-0827 model will try, and try, and try again to solve a problem. For example, this is a snippet from generated output when asked to solve a geometric problem:
QUOTE
We have two equations:
1. z² + y² = x² (from the similar triangles)
2. x² + y² = z² (from triangle ABC)
Notice that equation (1) is incorrect. We made an error in setting up the proportions from the similar triangles. Let’s correct that:
END QUOTE
The model then proceeds to set up the proportions from similar triangles afresh. If that doesn’t work, it will try a different approach altogether. That is remarkably similar to what a human problem solver would do. It produces way more generated output, because the solution attempts that proved fruitless are in the output as well, but it gets the job done (=finds the correct answer) much more frequently than the -002 model does (which favors shorter responses).
1 Like