lemontheme

Neat! My guess would have been Lonestar as the main inspiration. Autobattlers aren't my main genre. Even the much-loved He Is Coming bores me after half an hour, but I'll give this a whirl!

I’m still figuring out codemode. I was using it in a prototype, but ended up stripping it out again, after realizing my small local LLM was using more tokens than usual. It was combining tool calls elegantly in code exactly how I hoped it would. The problem was that when any of those embedded tool calls failed (e.g. on parameter validation) the parent code execution tool call also failed. In response, the LLM kept rewriting large parts of the original code block.

Btw, Monty by the pydantic team is a joy to work with if you need a way to securely run unverified code. It’s a simplified Python dialect. You can also use it from JS, iirc.

Aerospace user of two years. I grumble every time I need to pry a lost app window out of the bottom right corner of the screen, but aerospace is by far the least of all evils for anyone who wants fast, keyboard-driven desktop switching (and tiling, obviously, although I still let most of my windows float)

Fwiw – Orion (webkit) has tree tabs

I thought openrouter just routes you to the same provider for the rest of the session, so that you keep hitting the same cache. Is that not the case?

Also, I’d love to use Deepseek directly (or any of the Chinese providers, at that). Seems only fair to pay the lab that built the model. Unfortunately, any requests to Chinese servers is deeply frowned upon here (Belgium, EU). For personal use: sure. As a token intelligence strategy for the company: absolutely fucking not.

The model will gladly generate a historically factual response. It's the Chinese providers that block it.

That’s a funny ad absurdum. But let’s be honest: if a coworker dumps a hundred lines of logs into a Slack channel without contextualization the reaction would understandably be ‘wtf am I supposed to do with this?’ Now what I’ll sometimes get is an accompanying AI-generated analysis of the logs. Out of diligence I’ll read them. To their credit, they usually mention the issue, but it’s buried down deep among so much noise. My time feels wasted. Eight paragraphs of Claudish could have been one sentence. My coworkers could have used AI to reduce it down to that one sentence and then rewrite it in their own words. They chose not to because that would have taken more time for them, not realizing the burden now shifts to me (the more expensive senior engineer).

My personal line in the sand is human-to-human communication. I’m fine with AI use on either end, but keep the channel clean. I see it as a sign of respecting each other’s time.

Btw, there are pre-genAI cases of generated text. Webshop product descriptions and sports news summaries, for instance – both of those are easily templatized. I guess the difference is there is no human name attached to them and they’re fairly short

I would not extend it to cases where I’ve made the choice to interact with AI, no. Precisely because there is no pretense. I know that I’m dealing with an intelligent system that has different strengths and weaknesses than a human. I also know that I can (usually) direct that system to dial down its verbosity.

But I still took time to thumb it out on my phone keyboard, just as I’m taking time to write this reply. The time it takes you to read this reply will likely be less than it took me to write. I accept that asymmetry because it’s how we as humans can convey that we thought something was worth saying. Long walls of AI slop flip the equation, placing the burden on the reader. It’s not about originality, not for me at least. It’s about respecting each other’s time as humans

Is that with or without the slop grenade examples included, because I can see how those might trigger the detector

If it’s not worth writing, it’s not worth reading.

That said, wouldn’t it be funny if this site, too, were AI generated?

Yeah, ticket is great! I liked it so much I ended up porting it to Python, which I find easier to read than awk/sh, and tweaking the command surface over dozens of iterations of trying to make the design more ‘self-explanatory’ for coding agents. https://github.com/a3lem/tiquette

(Excuse the vibe-coded README)

Anyway, even though I’m using it every day, I’m beginning to doubt the core idea of tracking hundreds of .md files in git. They get in the way of fuzzy file search, and if you keep closed tickets around in an archive directory, they just pile up endlessly.

Lately I’ve been considering breaking tickets out of VCS entirely or moving them to a dedicated repository just for ticket tracking.

Git bug looks like it might at least handle getting tickets out of the way

My take: if not for arxiv and huggingface, the field of ML would be nowhere near where it is today.

More to your question, I recommend something like semanticscholar to find actual relevant papers. Try to identify researchers that seem trustworthy, then explore the citation network around them.

For more hot off the press stuff, follow what gets boosted on social media.

Still doesn’t cover the truly niche stuff but it’s a start

Enjoyed this. Clear and calm argumentation style is a nice change of pace from all the hype or doom, both of which serve to boost a narrative intended for investors.

I agreed right up to the conclusion about the need for increased safety regulation. How do we ensure such regulation protects us instead of the interests of a handful of ‘all-powerful’ frontier labs?

Maybe regulation can also mean a new accountability framework. That I would agree with. ‘I didn’t mean to hack into that bank; my autonomous agent did it while price shopping’ sounds ridiculous, yet it’s not far removed from what frontier labs are getting away with in the pursuit of more attention

Hold on, isn't that just standard practice for post-training LLMs for agentic use? Give task, generate n rollouts, grade rollouts (either at termination or after each tool call)? Or is the difference that the rollouts are generated ahead of time and then graded? (Of course, then it's not really on-policy.)

I felt this article, minus the conclusion, perhaps. There is something profoundly rotten going on. Used to be that you had to keep the quiet part to yourself. Now you can espouse white supremacy and enjoy broad social capital. Everyone at the top appears to have agreed that, yeah, fascism-lite is actually kind of convenient from a capitalist perspective, and they’re enabling anyone who helps spread that ethos. Meanwhile half of our industry is at best apathetic.

‘You can use Omarchy without caring about DHH’s politics,’ explained one of my colleagues. And he’s kind of right. Death of the author – that sort of thing. But using it also lends credibility to the maker, who can point at their project’s popularity as an indirect sign of vindication of their ideology. Hence, as I see it, the only morally right choice is not to use it.

Regarding the conclusion, I wouldn’t leave. I see it as my duty to push back, even if the effect is woefully modest. We need to start reminding each other again that, no, this isn’t okay. In the office and online.

Klassif.ai | ML/AI Engineer | Hybrid (Leuven, Belgium; remote 3 days/week)

We started off naively as just another intelligent document processing SaaS. Years of experience have shown us where the real pain points and value lies. Join my AI team and be horrified by how many critical business processes rely on email. Then help me execute on our ambitious roadmap!

If you find yourself nodding along with the following statements, you may be a good fit:

- You enjoy getting your work into the hands of users

- You’re okay debugging a stray tensor in the morning and designing services in the afternoon

- You care about the field of AI and interact with its community

- You can ‘sense’ when a loss curve is doomed

- You’ve fine-tuned LLMs, even just for fun

- You’ve been thinking a lot about agents lately, for some reason

- You don’t mind the odd bit of customer support (it’s where we get our best insights!)

All credentials welcome, including recent graduates. We’ll shape the role around you. What matters is creativity and visible engagement with the field.

Who may apply: EU citizens or holders of a Belgian work permit. (Everything else is more admin than we have time for.)

https://klassif.ai/careers/ai-ml-engineer

Hah, I use the same probe when I’m unsure which provider openrouter is routing me to!

Fwiw, deepseek v4 will happily discuss it. Only chinese providers will stop it in its tracks and give a canned answer. Streamed responses sometimes start with what the model was actually generating before it got cut off.

It’s top bad, really. Sometimes the Chinese providers are the model labs themselves, like deepseek. I’d like my money to go directly to deepseek, since they did all the work. But data protection concerns aside, how do I trust a system that denies objective reality? (Kind of like how Grok will tell me that wikipedia is ‘woke’.)

Kind of a coincidence to see this post here, just after I’ve gotten done playing with Kakoune and poking around the docs.

After using Helix fulltime for two years, I’m considering a switch. Helix is fantastic OOTB and I have considerable muscle memory in it, but the pace of development is glacial. There are I think three active maintainers, and they’re completely swamped by PRs, many of which have sat unreviewed since they were openend. It’s enough to wonder why anyone would even consider contributing. That would be understandable if it weren’t for the fact that, no, Helix isn’t ‘finished’. Newer LSP features aren’t supported; DAP for python has never worked; I also miss code folding, bread crumbs, and tab name disambiguation. The steel scheme plugin branch is functional and several neat plugins are beginning to emerge. However, I switched back to main after noticing that the plugin branch adds about 1s to startup time. (I’m constantly opening and closing helix.)

Anyway, Kakoune has an attractive philosophy. There’s ‘editor for life’ potential. A really minimal core that I can shape to my needs in the moment. The scripting language looks even gnarlier than vimscript but no matter – we have coding agents for that now.

But Kakoune’s minimalism also brings downsides that make me wonder if it’s more effort than it’s worth. For example, that minimalism rules out a dependency on treesitter. But you can hardly expect to shell out for treesitter highlighting and expect good performance. With Helix, nvim, and Zed all consolidating on treesitter, I don’t see why I’d want to spend time on an already solved problem.

All in all, neovim is feeling like a better option, and as another commenter says, configuring it has become remarkably easy thanks to coding agents. And apparently native multiple cursors are on the roadmap.

Nono has been my daily driver since the start of the year. It's not a perfect sandbox -- that's for sure. For example, the default network rules let you escape via a global TMUX server. But it is extremely practical. It gives me enough guarantees to feel confident about running in YOLO mode. So far nothing has gone awry.

As a zellij sponsor I’ve been wondering the same.

I rebuild herdr from master every few days to gauge the direction it’s moving in.

What herdr gets right over zellij:

- Default keybinds don’t conflict with helix. Even the leader key mode in zellij keeps alt-bindings around. Herdr keeps things nicely tucked away.

- There’s beta support for the kitty image protocol. Three years on, we’re still waiting in zellij.

- Notifications get passed correctly through to the terminal.

As for herdr’s killer feature, agent status monitoring, there is apparently a zellij plugin.

Big win for zellij: the docs and code feel human. By contrast, herdr’s docs have all the tells of a vibe coded project. Not to say the maintainer hasn’t put care into it! But I’m getting burned out on LLM docs.

Claude is obsessed with making things land. More than once I've reminded it that it's not a pilot.

And it has teeth.

Having used it for three weeks before coming back to Kitty, my conclusions were that it’s a better terminal for people who need a terminal but don’t ‘live in’ their terminal. Ghostty just feels really polished and ‘native’ and it’s obviously faster than, say, iterm or the stock term on macOS. So Ghostty is what I would now recommend to anyone who just wants a nice terminal without the goal of turning it into a highly personalized IDE, like Kitty allows you to do.

The only feature of Ghostty that I now miss in Kitty: setting the text size of each window pane independently. Oh and the slightly smoother quake terminal, even though I don’t use that much.

Ghostty makes some choices that I personally find bizarre. For one, on macOS, cmd+, opens config in textedit instead of in $EDITOR. Another is there isn’t a key action for moving panes; you need to use the mouse. At least half of the config options are platform-specific. And weirdest of all scripting is limited and can only really be done with (get this) AppleScript.

So, like I surmised (and I could be wrong, of course), Ghostty is just a nice terminal that lives alongside your other apps; Kitty is powerful enough to practically act like a second desktop environment (that is also cross-platform).

Of course, this post is about libghostty, the underlying term emulator library. And that is an amazing contribution to the ecosystem! Zmx, herdr, soon neovim, now emacs, hopefully someday Zed – so many things can now embed an amazingly performant terminal

I can only speak for western Europe. New-builds prioritize energy efficiency, because energy is expensive and, depending on the region, involves burning fossil fuels. There also tax incentives. The lucky new builds have reversible floor heating attached to a heat pump. That can keep the temperature around 26 celsius. Still on the warmer side but not unbearable. Anyway, AC is (or at least was) seen as a last resort, hardly worth it for the ten or so actually hot days we get a year. Obviously, that counter has been steadily going up, so the mindset is changing.

If you do most of your work on a central server and treat your other devices as thin clients, then (afaik) Tmux/Zellij are the way to go – not to mention if you want to multiplayer with colleagues.

(I know Wezterm has remote session multiplexing in beta, but I’d be surprised if you can re-attach from a different machine.)

Lately I’ve been experimenting with embracing the total scriptability of my favorite terminal, Kitty. Its creator is famously outspoken about non-native multiplexers, precisely because of the pitfalls you mentioned and because they hold back the terminal ecosystem.

Couple of recent issues I’ve had: washed-out colors in Claude Code, desktop notifications not passing through, scrollback weirdness, and image protocol mismatches.

I used to treat Zellij as an OS to build my own IDE in. Now it’s Kitty.

I save layouts as a Kitty session. Important limitation: this doesn’t persist running processes. And so enter zmx. Whenever I need a process I can reattach to I wrap it in zmx.

A terminal agent can interact with the kitty cli to create splits and such. I’m still on the fence as to whether that’s safe. Probably not. Also, tmux is likely better represented in the training data. Currently experimenting with running claude inside tmux (multiplexing) inside nono (sandbox).

According to my Volvo EX30 I’m always tired. It feels like they’ve dialed down the sensitivity in recent updates, but for the first year it was beeping at me five minutes into every drive. Really easy to trigger the alert by fake yawning. I turned it into a little game. Oh, and I can’t look over my shoulder to check my blindspot. Nope: driver distracted.

My steering wheel has one macro button, which I can bind to EITHER disabling speed alerts or driver monitoring alerts but, very unfortunately, not both. So guess who’s sometimes poking at their touchscreen while they should be focusing on the road?

Not against alerts in principle but it should be a less is more kind of thing

Exactly — and it carries the point that others would try to gate

It’s the repetitive paragraph structure that raised my suspicions.

Sentence. Short sentence. Noun. Noun. Noun. Sentence to reflect. Wrap it up in a bow.

Sequences of short incomplete sentences, particularly in threes is such a tell

Not just the GBA, apparently! 3rd party handhelds like the Pocket and emulators are supported too.

From the website to the software, this looks stunning.

Tried chip music with nanoloop several years ago. Wasn't 100% for me: 1) pure chip music is too jarring for my ears; 2) it's in the name ('nanoloop'), but it pushes you toward short repeating loops. Once you've heard the same 4-bar loop a couple of hundred times, it's hard to imagine the next phrase. Feels like getting stuck. This has a 128 pattern * 16-step sequencer. If those patterns are easy to chain, then this could be really nice for longer phrases.

Can anyone comment on how much effort goes into flashing a cartridge? I've never done it. Back in the day, I just ordered nanoloop straight from the website.