As for deepseek r1, I’ve only tested the chat with rather complex abstract questions so far with VERY good results, but I still need to integrate it into my workflow with Cursor ai via API, although they’ve already added Deepseek 3 which works decently as well from what I see, although it doesn’t have this “reasoning” capability.
In my case and regarding the current problem with o1 Pro, now it’s not giving me the “Finished Thinking” errors anymore, but it’s practically unusable since it forgets the context very, very quickly, and even if it’s in Pro mode with “reasoning” activated, it only takes a couple of seconds generally to analyze things that used to take minutes (although it never got results because it failed before).
I’ve asked some basic questions regarding his context and limit and his answer really caught my attention, since I remember that with Plus and o1 or o1-mini I could do things like that without problems, with much longer files and and he kept the context in a very decent way, this was what he answered me:
From what I’ve seen in reviews, videos, etc., this shouldn’t be the case. I mean, how can we expect it to solve complex things if we can even give it a pass between messages with so little context?
I understand that the answer says that I shouldn’t send them all together, it’s understandable since it would be 4000 lines of code, but what I’m doing is simply consecutively sending 4 or 5 files with 800 lines of code each, explicitly indicating that we should analyze them together to determine the logical or conceptual problem, and immediately after I send a piece of code it does a basic analysis in a few seconds, I repeat it again with the next one and the same thing, and by the third message it’s not even able to maintain the flow of what we were talking about.
This is something that I don’t think happened even with the most basic models, and are things you can do with Claude Sonnet, Deepsek, Gemini.
There is clearly a problem here.