First milestone for ysr (read: wiser) is complete - it's a yrs/yjs compatible library that runs natively on disk (via LMDB). No need for separate persistence, no need to load entire doc into memory.
Currently supports update binary compatibility for maps, lists, rich text: github.com/Horusiath/ysr
The biggest difference in moving from regular employment back to contracts is that now I work from dawn to dusk, 7 days a week with no vacation on sight. And I don't know if I have work secured to the end of the year.
My kind of memes.
So it looks like CC is making their AI dumber, while the usage is evaporating noticeably faster. What subscription terms say about that?
Nothing, because subscription provides no measurable obligation to anything (CC Max is just "5x more happiness*" than CC Pro). idk how is that even legal.
"Hey, Aang. It's great that you managed to find another airbender and defrost him after 5000 years. You won't be the last airbender anymore."
Meanwhile, airbenders 5000 years ago:
Getting a lot of videos on YT from American vloggers prising Poland recently.
Folks, my friend is bookstore clerk. She earns ~1000USD/mo (net). She lives with her husband in 1BR, which is 1000USD/mo. This is considered as the best era in history of the country.
"Jaws", "The Crow" (1994), "The Doors" (1991), "Dracula" (1992) - all watched when my age was a single digit number. Don't consider it "too young" though in any negative sense (maybe except Jaws).
I tried AI on several different projects, one things is common: the learning and expertise gained in AI-driven projects is close to zero.
It's similar to reading CS papers: you have "feeling" of learning without deep understanding. You can't learn to play guitar by asking someone to play it for you
While the devil lies in the details, but a general PoC of similar idea (folder + file sync over yjs) has been made some time ago: github.com/yjs/titanic
bsky and tangled share the same auth, so bsky the app could support "rich" embeds: if a user has a tangled repo in their post, the like button could be a tangled star button instead!
SQLite has 5 storage types (null, varint, float, text, blob) + column affinity. My toy DB doesn't use varint (due to FlatBuffer-style row layout), but I can try to bit-pack storage and affinity types together to skip using separate column type affinity altogether.
So what's the problem? It lies and skips important details for its own convenience all the time. But it communicates in a way that makes you want to believe it.
The worst part of discussing things with AI is that it may be dumb, but sounds intelligent and communicates in more comprehensive manner than most humans (esp. given time constraints).
I don't think I ever had this level of conversation and action plan when discussing the design with real person:
Reading a shitton of code generated by AI can be overwhelming and encourages blind acceptance - especially since the writing looks professional.
But the fact that it does sound convincing (it's the only thing that AI is actually optimised for) doesn't mean it knows what it's doing.
Tbh. something that Claude is pretty good at is figuring out Rust borrow checker errors.
“I was lost away from home in a bizarre territory where people made plans that didn’t make sense with the aplomb of a drunk LLM.”
isolveproblems.substack.com/p/how-micros...
Follow up questions:
- Can using AI to commit warcrimes be considered as "for entertainment purposes"?
- Are Anthropic stakeholders ($600bln valuation) aware that it cannot be used by users to make money?
After all, we wouldn't download a car and we don't want to violate ToS.
Do these companies boasting how they are doing 100% AI-gen their code to make millions of dollars actually read ToS for their tools?:
- www.microsoft.com/en-us/micros...
- www.anthropic.com/legal/consum...
this is an incredibly thoughtful post
So I'm vibing a toy SQL database in C#. Uses SQLite dialect, but very different design decisions. It's slower (and probably always will be) than SQLite on most benchmarks, but there are a lot more things I want to implement.
gist.github.com/Horusiath/21...
A little pearl I've just found - article (+ working Scala example) explaining how cost-based SQL query optimiser works: github.com/tuannh982/qu...
alternate title: "single-page apps are good actually"
I'm still confounded by people claiming AI "is just a tool" and comparing it to compilers, formatters or IntelliSense code gen. Is this how people are coping with it?
Isn't that like most of them?
a graph of axios's 25 transitive dependencies
25 transitive dependencies to get what the browser already gives you for free npmgraph.js.org?q=axios@1.14.0
Tbh. one of the reasons why I'm using AI on 3 different projects right now is a feeling, that "free lunch" (actually $110/mo) will soon be over and we'll get dragged into ass-f**king phase next.
But it may come sooner than I expected: notes.zachmanson.com/copilot-edit...
Another argument for using a local-first VCS is that with the current trend, by the end of 2026 my PC will have a higher uptime than GH or Claude.
