Claude 3.5 Sonnet is impressive

Initial impressions

I am trying out Claude 3.5 sonnet (both the model and the projects they introduced recently) and am loving it so far. A very anecdotal experience:

I asked about very specific kinds of learning resources for some subjects - ChatGPT (4o) and Claude (3.5 sonnet), and Claude won by a huge margin.

  • ChatGPT answers were predictable, Claude responded with completely hidden gems

  • ChatGPT links were mostly missing or broken, most of Claude links were accurate (a couple weren't)

  • While both gave items that were not aligned perfectly with my criteria, ChatGPT did nothing to explain the limitations of its response wrt my ask, instead claiming to align with all the criteria, while Claude pointed at the exact limitations of some of the items on the list - this is a huge factor in how much we can trust model responses.

I am sure there are various factors at play here, but I suspect that a large part of it is in the alignment fine tuning bit - ChatGPT seems eager to please and claim alignment, while Claude seems more bold and willing to admit misalignment.

Also, the project based management of chats in Claude is super useful, even for private organization (haven't tried it for collaboration, but I suspect that will be even more useful).

It was only a matter of time that someone gave ChatGPT a run for its money. I believe it's a good thing - will spark more innovation all around. Definitely one of the cases where choice benefits users.