大家都在说什么Anthropic的价值观，用户隐私什么的，都没明白，这根本就是个生死问题。
Anthropic和OpenAI估值这么高，然后花大量的成本在训练sota模型，每次刚训练好，你给我蒸馏，然后三个月后开源。大家就都说中国模型好啊，便宜也够用，那我估值怎么办，我股价怎么办，背后那么多投资怎么办？
1
37
MCP tool catalogs can burn a lot of context.
I built/used mcp-smart-proxy to expose only 3 proxy tools with same capability, saves at least 80% tokens.
It also converts any MCP into CLI.
Repo: github.com/cybershape/mcp-sm…
Post: b.oldhu.com/post/mcp-smart-p…
33
cybershape retweeted
Today is #babashka's 5th birthday! To celebrate, share your favorite bb scripts, tips, tricks, projects with the #babashka and #clojure hashtag, star the repo if you haven't done so or RT this tweet! Thanks for using bb!
github.com/babashka/babashka
4
19
95
3,170
I was reading this blog about qdrant sizing. But the numbers just don't add up. 1 million dim-100 vector needs about 400MB, your capacity guide says we need to multiply it by 1.5 which is 600MB. But in this test, 1200MB is needed to keep all data in memory?
Vector search is thought to require lots of RAM. At least it used to, as it is possible to get as low as serving 1M+ vectors with only 135 MB of RAM 🤯
Check out the article written by @generall931 about MMAP support in Qdrant qdrant.tech/articles/memory-…
1
148
Its a shame that Garmin's all recent watches don't support Unicode smart notifications. For non-technical reasons!
cybershape retweeted
我草- - Facebook上面的一个台湾同胞说- - 他们台湾现在也在播台湾当地的春晚，重点在于- - 他们也有王力宏和小虎队 我草- - 赶场子这么高效率？ #chunwan
2
cybershape retweeted
Downloaded iPad SDK Beta 2 on my server at 17MB/s, now trying to move it to local.
1
RT @iStone1983: RT @guxiaoqiang: #GoogleCN 资深互联网人士谢文今日接受TechWeb连线时呼吁：“中国互联网企业不要幸灾乐祸，谷歌的今天就是你们的明天。” http://bit.ly/4ulnfr