It's the same thing as with set theory. Once you let yourself talk about sets of sets without any restrictions other than the language itself puts on what you are talking about, you'll end up with sort of contradictory recursion that ends up in a paradox. The message is not "set theory is incomplete, or invalid" but rather "we were a bit too cavalier with words and not everything we can say about sets makes sense just because it's grammatically correct so we need to be more careful about defining what a set is and what it can and cannot contain".

For me "provability" is a direct equivalent of unrestricted sets of sets. If you don't restrict what provability means and how it can be used in context of the rest of math you end up with contradictory recursion that makes you believe there are true but unprovable things. It's easier to spot how bonkers it is on sets because there you end up with a set that is and isn't its own element at the same time (because sets are so generic that can produce paradoxical recursion on themselves without any other concept involved), but things being unprovable and true at the same time is the same category of absurdity.

Deepseek Harness supposedly has nice plugin system which others do mostly lack.

But current system is close to perfection for me.

Gemini doesn't have adjustable reasoning effort (at least on the graph) so each of its curves is just one point.

Damn, Luna on max is as good on DeepSWE as Kimi k3, I think I dismissed this model unjustly.

That actually sounds amazing.

AI can write custom plugins for you. So this means the tool is infinitely flexible for you, even without any community.

Compare this to Zed where I can't make a hexviewer for binary files or player for audio files for myself without recompiling Zed's source code.