I have declared that the personal project I've been working on for a little bit is now officially at "version 1.0" and I'm hoping that this week I'll recruit a few users to try it out.
It is a notes application. Very simple, it just takes plain text notes, but written as solidly as I could make it. (And designed to install on phones as a PWA, using "save to desktop".)
If anyone feels like checking it out, it's at https://mini-notes.com. Accounts are free (now and forever). Any feedback would be welcome.
Finishing up at $dayjob before taking the whole month of June off and starting a new job after that.
Also RustWeek in Utrecht. Considering attending the Hackathon on Thursday for the social aspect, as I will likely also need to spend some time on my job too
Oh, very cool! I do love a good European programming conference with associated hackathon. That seems to me to be a near-optimal life choice, at least as an occasional thing!
(Some agent-related shenanigans, since apparently this stuff isn't going away.)
This past weekend, I updated my Linux-only agent sandbox to support a basic filtering proxy server. This doesn't handle credential injection yet. But if I have time, I might work on that next weekend. (Or whenever!) No, this is not quite yet ready for anyone other than me to use. But I'm personally fond of the modular profiles:
redoubtful run --uses pi --uses llama-server pi
This creates a bwrap + pasta namespace. It also mounts a "phantom" home directory that looks my real one, but which contains only the project directory and anything specifically mapped in by the profile. Similarly, there's no network access besides what's configured, and the environment is sanitized. But the net result is that I can move seamlessly back and forth between my editor and a sandboxed agent.
Coding agents. This project has also been an exercise in trying Qwen3.6 27B as a coding agent. I actually like this particular small model, because it's clever enough to do a moderately messy refactoring or to implement a straightforward feature. But it doesn't forgive woolly thinking or vibe coding at all. Using Opus 4.7, on the other hand, allows delegating so much of the design and implementation that I quickly lose track of what's going on.
My new hypothesis: The actual bottleneck in high-quality agentic coding is the model in the developer's head. Therefore, if we take all that "Theory of Constraints" stuff seriously, we should carefully optimize our processes to help maintain the programmer's mental model. Or the team's. And Opus 4.7 has reached a point where it's actively corrosive to maintaining a mental model.
Old-school "fill in the middle" (FIM) autocompletion. I actually miss the late-2024 version of Copilot, the one that would happily suggest entire "ghost" functions. I don't especially like the new version that proposes diffs. The old style felt sort of like touch-typing, subjectively. Except faster. The new style feels more like hunt-and-peck because I need to switch my mental focus from my cursor to some proposed 3-part edit.
So I have been going through lots of "FIM completion models" and hooking them into Zed. The pleasant surprise here was Qwen3.6 35B A3B, which actually mostly works as a FIM-style autocompleter, without any fancy post-processing. I tried various experiments, like swapping PSM format to SPM, and proxying /v1/completions to /v1/chat/completions with a custom prompt and tools. But each of these made the model worse.
My next hypothesis is that some fine-tuning might help. But that would be a bunch of work to set up, even using a LoRA adapter.
Made an emoji programming language called Mojie. Complete with transpiling to Js, Python, or English, a playground, and an LSP. No AI involved. Managed to reuse most of the codebase from a stack-based language I made for a talk. I quite like it. But building it sent me down the path of "what if vibe coding produced an AST rather than actual code", so that's what I'm investigating this week. Quite excited about it, actually!
And in the meanwhile, I'm migrating to Codeberg. First migrated to Forgejo in my homelab, but next to actual Codeberg.
Recovering from a vacation. It was very fun, but I’m also glad to be home.
I haven’t had much time to code personal projects lately and I’m hopeful I’m gonna be able to find some time this week. I’m itching to code some again!
As for work, it’ll be helping out some more junior coworkers. I’m hoping I’ll be able to get them to try some more active roles in project planning. No better way to learn than to try doing something!
Working on the next iteration of my interactive multi-line search tool septum. This engineering pass is pretty close to done, next is a feature polish pass. The previous version (0.2.1) has been useful and stable enough I want to set up rolling release binaries in GitHub.
I'm experimenting with a Chinese-character incremental game.
Instead of letters or tiles, the basic units are radicals/components. The player places them on a 5x5 grid, and the game uses a small hanzi.csv dataset to detect valid characters formed by adjacent components.
The main design problem this week: if every component generates money, players may ignore character formation and just optimize raw income. So I'm trying to make completed characters act more like engines, while standalone components remain weak.
Building a P2P client for my Abject project (https://abject.world). Right now if you want to run the client on the phone you have to host the server and client yourself somewhere. But since Abject already has P2P capabilities, I am transforming the client to a p2p one so you can run it at home and still access it on your phone from anywhere in the world.
I'm adding support for “member-macros” in my WIP language/compiler. The idea is that you can declare macros not only in global/local scopes, but also as belonging to some data structure and the macro is always expanded having the original object as a parameter.
mcherm | 16 hours ago
I have declared that the personal project I've been working on for a little bit is now officially at "version 1.0" and I'm hoping that this week I'll recruit a few users to try it out.
It is a notes application. Very simple, it just takes plain text notes, but written as solidly as I could make it. (And designed to install on phones as a PWA, using "save to desktop".)
If anyone feels like checking it out, it's at https://mini-notes.com. Accounts are free (now and forever). Any feedback would be welcome.
ElaineRoberts | 13 hours ago
Love the idea, would totally use it... but the UI is very much not to my taste :(
polyfloyd | 15 hours ago
Finishing up at $dayjob before taking the whole month of June off and starting a new job after that.
Also RustWeek in Utrecht. Considering attending the Hackathon on Thursday for the social aspect, as I will likely also need to spend some time on my job too
emk | 10 hours ago
Oh, very cool! I do love a good European programming conference with associated hackathon. That seems to me to be a near-optimal life choice, at least as an occasional thing!
scubbo | 7 hours ago
Congrats!
emk | 14 hours ago
(Some agent-related shenanigans, since apparently this stuff isn't going away.)
This past weekend, I updated my Linux-only agent sandbox to support a basic filtering proxy server. This doesn't handle credential injection yet. But if I have time, I might work on that next weekend. (Or whenever!) No, this is not quite yet ready for anyone other than me to use. But I'm personally fond of the modular profiles:
This creates a
bwrap+pastanamespace. It also mounts a "phantom" home directory that looks my real one, but which contains only the project directory and anything specifically mapped in by the profile. Similarly, there's no network access besides what's configured, and the environment is sanitized. But the net result is that I can move seamlessly back and forth between my editor and a sandboxed agent.Coding agents. This project has also been an exercise in trying Qwen3.6 27B as a coding agent. I actually like this particular small model, because it's clever enough to do a moderately messy refactoring or to implement a straightforward feature. But it doesn't forgive woolly thinking or vibe coding at all. Using Opus 4.7, on the other hand, allows delegating so much of the design and implementation that I quickly lose track of what's going on.
My new hypothesis: The actual bottleneck in high-quality agentic coding is the model in the developer's head. Therefore, if we take all that "Theory of Constraints" stuff seriously, we should carefully optimize our processes to help maintain the programmer's mental model. Or the team's. And Opus 4.7 has reached a point where it's actively corrosive to maintaining a mental model.
Old-school "fill in the middle" (FIM) autocompletion. I actually miss the late-2024 version of Copilot, the one that would happily suggest entire "ghost" functions. I don't especially like the new version that proposes diffs. The old style felt sort of like touch-typing, subjectively. Except faster. The new style feels more like hunt-and-peck because I need to switch my mental focus from my cursor to some proposed 3-part edit.
So I have been going through lots of "FIM completion models" and hooking them into Zed. The pleasant surprise here was Qwen3.6 35B A3B, which actually mostly works as a FIM-style autocompleter, without any fancy post-processing. I tried various experiments, like swapping PSM format to SPM, and proxying
/v1/completionsto/v1/chat/completionswith a custom prompt and tools. But each of these made the model worse.My next hypothesis is that some fine-tuning might help. But that would be a bunch of work to set up, even using a LoRA adapter.
eeue56 | 12 hours ago
Made an emoji programming language called Mojie. Complete with transpiling to Js, Python, or English, a playground, and an LSP. No AI involved. Managed to reuse most of the codebase from a stack-based language I made for a talk. I quite like it. But building it sent me down the path of "what if vibe coding produced an AST rather than actual code", so that's what I'm investigating this week. Quite excited about it, actually!
And in the meanwhile, I'm migrating to Codeberg. First migrated to Forgejo in my homelab, but next to actual Codeberg.
nick4 | 14 hours ago
Recovering from a vacation. It was very fun, but I’m also glad to be home.
I haven’t had much time to code personal projects lately and I’m hopeful I’m gonna be able to find some time this week. I’m itching to code some again!
As for work, it’ll be helping out some more junior coworkers. I’m hoping I’ll be able to get them to try some more active roles in project planning. No better way to learn than to try doing something!
strongly-typed | 11 hours ago
Existential crisis as I experience inability to pay for my bills.
Also, working through Theorem Proving in Lean 4
slowerloris | 8 hours ago
Working on my game, and by extension, game engine. It's been a lot to learn, but struggling through new concepts has been a big part of the fun.
pyj | 5 hours ago
Working on the next iteration of my interactive multi-line search tool septum. This engineering pass is pretty close to done, next is a feature polish pass. The previous version (0.2.1) has been useful and stable enough I want to set up rolling release binaries in GitHub.
yiyu | 36 minutes ago
I'm experimenting with a Chinese-character incremental game.
Instead of letters or tiles, the basic units are radicals/components. The player places them on a 5x5 grid, and the game uses a small hanzi.csv dataset to detect valid characters formed by adjacent components.
The main design problem this week: if every component generates money, players may ignore character formation and just optimize raw income. So I'm trying to make completed characters act more like engines, while standalone components remain weak.
Prototype: https://www.chunqiuyiyu.com/hangrid/
mark | 16 hours ago
trying to get assethoard.com's Arch package signed and pushed to the aur repository.
icefox | 13 hours ago
Working, hopefully productively, and playing with sdl3.
Also playing Fire Emblem 3 Houses, if my RSI allows.
mempko | 11 hours ago
Building a P2P client for my Abject project (https://abject.world). Right now if you want to run the client on the phone you have to host the server and client yourself somewhere. But since Abject already has P2P capabilities, I am transforming the client to a p2p one so you can run it at home and still access it on your phone from anywhere in the world.
vinipsmaker | 10 hours ago
I'm adding support for “member-macros” in my WIP language/compiler. The idea is that you can declare macros not only in global/local scopes, but also as belonging to some data structure and the macro is always expanded having the original object as a parameter.
emk | 10 hours ago
Huh, for some reason this makes me think of Rust's "postfix member keyword"
.await. I'd love to see an example of what you're describing!