ansh

❗️Funny: Anthropic’s Claude Sonnet 4.6, when asked in Chinese: “你是什么模型?” (What model are you?) Replies: “我是 DeepSeek.” (I am DeepSeek.) This is Anthropic’s model, which just accused DeepSeek of stealing Claude’s capabilities.

nothing better than coding on christmas man 🦖

Thank u my dear Yogi Babu. Miss u lots.

no way

Yesterday I did an interesting experiment along these lines. I'm building a language focused on UI building and streaming data processing. It is strongly statically typed. I'm using claude coder to help me with the documentation. Yesterday I assigned it to write examples for all the widgets. Using the language documentation, the library source code, and the underlying tk documentation it wrote 61 small example programs. Reading through them, they looked directionally correct, but it was clear nearly all of them would not even compile, much less work. I decided, instead of fixing the programs, to implement a type check only option in the compiler. So just parse, compile, typecheck, don't execute. Once that was done I instructed claude coder to use it to fix all it's examples, which it spent about 15 minutes doing. When it was finished 100% of the examples compiled, and 92% of them worked as intended. This was especially shocking because these are all visual examples, and claude code can't see. So just with the ground truth of a strong static type system caude coder was able to improve it's task accuracy from a low single digit percentage to 92%. And it did this in a programming language that is certainly not in it's training set. Well many have pointed out that what these systems are missing is reasoning. A type system is a kind of logic engine. So is it really surprising that the two complement each other?