naaqq

6 ms·

1.
by naaqq 2mo ago
Will be later.
2.
by naaqq 2mo ago
just got 2/18, not a native english speaker, hard for me.
3.
by naaqq 4mo ago
I don't think it's helpful, you can already get a 99+% cache hit on claude code, just change the api settings to deepseek. I would like to use a agent built by deepseek itself using deepseek models. Deepseek should make their own
4.
by naaqq 5mo ago
Not all read tokens are included in the context, many of the tokens are from read cache hits. I hit it many times so it grew to 200M. The number came from the API platform.
5.
by naaqq 5mo ago
Sorry, I was wrong here. I meant a single long session. And there’s no compression, the 1M context is only half used.
6.
by naaqq 5mo ago
DeepSeek’s official API has a cache hit rate of over 99% if you use it continuously within the same codebase for long sessions, so it’s much cheaper than frontier models. I have an example of 200M token session in claude code.
7.
by naaqq 6mo ago
It reminds me of the Easy Programming Language( https://en.wikipedia.org/wiki/Easy_Programming_Language ), a Chinese programming language. Many kids in China started learning programming with it about 15 years ago.
8.
by naaqq 6mo ago
This article said some things I couldn’t put into words about different AI tools. Thanks for sharing.
9.
by naaqq 9mo ago
Round the squirrel only if the squirrel is still on the tree.
10.
by naaqq 9mo ago
Start reading from the 3/4 mark, that’s the ‘how’ part
11.
by naaqq 10mo ago
This is so cool. I like the mircoSD card idea.
12.
by naaqq 10mo ago
It makes LLMs stronger because code is something LLMs are good at.