@jleni

Making models replaceable before they return the favor. @_kunobi_ @_zondax_

Zug, Switzerland
Joined August 2008
Models are becoming interchangeable. But AI providers are quietly bundling the model, your memory, your relationships and your identity into one subscription. The intelligence should be replaceable. You shouldn’t be.
1
6
14
847
Juan Leni retweeted
Combe 有一个很小但很有用的小优化。 `Command+Shift+T` 可以直接把一个 Panel 提升为独立 Tab。这种场景在你经常 split 时, 然后发现太多了panel 是否就非常适用。 建议 ghostty 和 其他做 Terminal 都学习下这个小优化。
2
5
836
Juan Leni retweeted
It seems Kache by @_kunobi_ cached around 7.5 GBs worth of Cargo crates, but thanks to this one of my projects is down from 13.5 GBs of space to 4 GBs, while the other one is down from 3 GBs to a mere 740 MB. Not the mention the faster build times.
1
1
13
281
Juan Leni retweeted
Didn't get the chance to try it out til now, wasn't building anything, but I'm back, just cleaned up 13 gigs of a project, let's see how well this will fare, if everything goes well I'll make a post and tag u to it 🫡
1
2
51
Juan Leni retweeted
Replying to @3povC @ImLunaHey
Fair: sccache can share a cache directory, and our CI benchmark does that. We should update docs. The difference is physical storage. In our Firefox 151 benchmark on APFS, a second worktree added about 3 GB with Kache versus 16.7 GB with sccache. Glad to get feedback on the benchmarks too! We aim to be fair! We just built this internally because rust target/ dirs were unbearable!
1
1
2
98
Juan Leni retweeted
Small thing I built in my spare time that has really helped agentic rust development: github.com/RhizoNymph/rust-b… It uses Kache backed by minio to set up a remote build server and shim all cargo commands (configurable what commands stay local) to run through it. If it can't reach it, it uses a local Kache instance instead and settles up the build artifacts when it can see it again. It also has a bootstrap command to set up the server and any number of clients as long as you have ssh access to them. Before I had this I always told my agents not to let subagents run rust builds because they were killing my laptop doing it in parallel but this fixed that entirely by introducing the caches and a build queue.
4
2
20
903
Juan Leni retweeted
Replying to @alswl @x_czyt
Use kache. We actually started with this concept and have been using this in production for while. At the same time, we have contributions from Mozilla, Apache, SpaceX and others! Lastly, we have a few surprises and innovation coming soon!
3
3
38
Juan Leni retweeted
Replying to @realfxw
Use kache. We started the concept internally, have been innovating since. There are a few other options that just followed our design.. But we have a few other internal innovations that we are open sourcing soon. If you have any feedback please ping us! Always open to issues, feedback and contributions
3
4
48
Juan Leni retweeted
Did you try it? We would love to hear some feedback! We built this internally just because of the same issue and we soon got people from mozilla, apache, SpaceX and others contributing!
2
3
62
When jev-fly ?
1
1
105
Juan Leni retweeted
Replying to @shuizhuyu @samzong
Any feedback is appreciated!
1
2
29
Juan Leni retweeted
Replying to @huacnlee
Kache mostly solved this for me (plus a 4tb ssd)
1
2
3
793
Juan Leni retweeted
Replying to @huacnlee
推荐 kache ,其次 gmc 也内置了 kache ,可以体验下
1
2
11
4,030
Juan Leni retweeted
Obviously yes I agree with most of this statement. BUT I don't want to train a classifier I just want to reach for a primitive and drop it in and have it work with no effort. I don't care that this isn't a new category I care that it makes the category trivial to adopt.
Jev looks cool. I hope to try it. But so far, the most interesting thing to me about it is the reactions. They suggest a lot of folks are not aware that encoder-only classifiers exist, have existed for a long time, and are good and valuable and widely used. It makes sense. LLMs have become so powerful and widespread so fast, it's all that some people have seen. For instance, for years now, you've been able to fine-tune a ModernBERT model (or maybe LFM2.5?) to do fast cheap structured output from a language input. I expect Jev is much better, and IIUC does not need training, but the category is not new.
15
11
1
222
12,861
Juan Leni retweeted
macOS: your disk is almost full me: it's not full, it's 9 worktrees each with their own tokio macOS: that's what full means kache init. one copy, shared across worktrees. keep using cargo. github.com/kunobi-ninja/kach…
5
7
99
Juan Leni retweeted
Sam Altman: coding is no longer the safest skill to bet your future on. “Learn to program was so obviously the right thing over a recent period of time, and now it’s not.” “Become high agency.” “Get good at generating ideas.” “Be very resilient. Be very adaptable to a rapidly changing world.” “I think these are going to matter more than any specific skill.” “And I think these are all learnable.”
148
239
52
3,238
463,477