hombre_fatal

On the other hand, AI is good at following invariants and adversarially policing the system to pay back debt. The latter is too expensive for humans as a project grows.

Right now you can open Claude Code in your project and prompt it "start a workflow that fans out subagents, each with a narrow scope, to look for {simplifications, correctness by construction opportunities, <insert your goal>} ranked by impact vs confidence, and then collate the findings into result.md". If you have the tokens, do more workflows that find more issues and/or vet and polish the findings already in the file.

I do that every time my weekly usage is about the reset to spend me last tokens.

Maintenance is becoming trivial, and anyone who thinks AI dooms you to an ever growing mudball apparently hasn't considered using AI for anything other than appending more code to the mudball.

The responses to the pricing aspect of this announcement around the web disagree.

I'm not saying you can't find paid software, especially from Oracle and Microsoft, but there's a different expectation for "just-a-guy announcing his project on twitter".

You can see a similar mentality regarding Elm in general where the approachability of one guy had people in some sort of parasocial entitlement to the project that you wouldn't see if, for example, it were Google or an unknown who built Elm.

Pretty much every software problem we work on breaks down into steps that are solved in prior art.

Even if it weren't, if you're capable of explaining the context and constraints of your problem, then a modern LLM with effort=high will generally come up with a solution that's worth starting with because it's well-reasoned.

Moreover, you can start with the solution and then course-correct based on future information because refactoring is trivial with an LLM, yet human projects often ratchet into a local optimum because refactoring is too expensive.

I don't think "it's been built before" does as much work as it seems. I didn't fork a project. The LLMs reasoned about how to build the project from scratch using trade-offs that made sense for my needs, and they made reasoned, unsolicited deviations from kitty, xterm, and co, not just blindly doing what some ref impl did. Btw, it was still a lot of work because my project isn't just "kitty but swift".

Then the models went on to drive a well-reasoned incremental implementation of a system that lets me use the terminal running on my Macbook from my iPhone over tailscale with a decent scheme it came up with itself.

The point is that I don't really have to know how things work to build good software with modern models. LLMs can do things like read Linux source code and adversarially refine ideas such that the final idea is a good one. And my biggest influences on the project can be automated through the use of reusable markdown files.

Now, I'm at risk of downplaying all my years in software here, but I see the writing on the wall. It was only one year ago that I only trusted AI to do autocomplete.

With Opus 5 / Fable / gpt-5.6 you can simply ask them to fan out subagents to look for ideal architectural simplifications and rank their findings by impact vs confidence.

You can swap out "architectural simplification" with performance opportunities, bugs, correctness, etc. I get the orchestrator agent to then itemize it all into a file where I can keep track of which ones I've implemented.

The results are pretty astounding. I run these right before my weekly limits reset for each subscription and the findings will dictate the secondary tasks I get done during the week.

It's definitely token-heavy. I'm on the $200/mo Claude Code sub and the $100/mo Codex sub.

But it's pretty clear to me that software engineering is more or less solved and all you need is enough patience + tokens to get what you want. I think 20 years of engineering experience more lets me save on tokens rather than unlock things nobody else can build.

An example of the scope of one of my AI-engineered projects is a iterm2/ghostty-like terminal app that implements its own pty session, parsing, rendering. It's almost 2000 commits right now.

That said, I have a specific workflow that isn't just a blind "ok now make it so a screen can be split into panes". I have a plan phase focused on coming up with ideal invariants and such. But I'm not sure anymore how much of that is useful vs just yoloing a solution and then paying technical debt in sweeps, like garbage collection.

For personal projects, I defer almost everything day-to-day to the models, so it gives me a good feel for what the models are good at.

Even then, it's kind of a wash these days between the sota models, and we're talking about maybe a 10% performance difference or something. But every once in a while there's the experience of one model spinning its wheels on a bug/repro/issue while another model comes in and one-shots the solution.

gpt-5.6-sol uses fewer tokens and OpenAI has been more generous with usage, but quality-wise I wouldn't say they are inferior to each other.

I have a Claude Code and OpenAI subscription so that I can use Opus/Fable/gpt-5.6 as I please, and the models are often catching things the other models missed. So much that I would significantly weaken my workflow if I dropped one subscription.

My best workflow at the moment is to create the initial plan with Fable (before review/revise-cycling with other models). From my own testing it seems slightly better at arriving at high-level ideal solutions after sweeping the whole project, projecting future needs, then coming up with good trade-offs like "by construction" correctness.

While mostly subjective, maybe the closest objectivity I have here is noticing fewer revision cycles needed with Fable-initialized plans.

This is good because there are much better places to find fulfillment.

It's like when you get a new smartphone and you baby it for those first couple weeks until, bam, it falls and gets a ding in it. Then after a brief sense of loss, you suddenly see it as a tool with utility and a lifespan. I think that's where you want to be with material things and technology.

You're basically saying that before you can know if AI is useful to you and your dealings, AI must solve all of a given corporation's problems including all the human/organizational components and even the trivial ones like how serious Github is about updating their status page, something that probably came down to one guy.

It's just really weak "AI bad" snark.

I think the only thing that matters is whatever the kid finds appealing enough to do at all.

e.g. I started with Ruby on Rails. Not even Ruby. The feedback loop is what convinced me that I wanted to write software and then dive into the fundamentals.

On the other hand, norms in software right now are that suckers build and maintain software for free + "the love of the game should be enough for anyone", so it's shocking when people break the norm.

Well, it would be a mistake on your part to assume that everything in an organization, especially a large one like github, is a reflection of the capabilities of the state of the art in anything (AI, architecture, programming language, organizational management, operations, etc.).

It's clear that updating the status page as quickly as possible isn't even something github cares about.

I also think about people stuck in bad relationships they'd otherwise leave.

Then again, CAC measures end-stage calcified plaque, not the soft plaque it started as, equally clogged arteries but also able to break off to cause strokes until the body calcified it.

The time to take action is long before you have a CAC score. e.g. Start tracking ApoB regularly, see if you have genetic LPa exposure, and avoid the foods that increase your exposure while ignoring the grifters telling you it's a nonissue.

That link just says the planning stage should vet the idea concretely so that the plan focuses on a solution that won’t immediately have to pivot.

And I think plan files should focus on general ideas and invariants, not do “implementation as prose”. That way they perform as mini-ADRs that are useful historically, especially to mine why the system is the way it is.

On the other hand, fast food places around me have large digital touch-screen kiosks for ordering. You can always go up to the counter to order, but it seems like everyone prefers the kiosk if one is available. For good reasons too -- you get to see what the food looks like, you don't feel rushed, etc.

I'm not so sure the experience (of talking to fast food employees?), if you can call it that, matters all that much to people who want to grab a quick bite with their family.

To me the biggest change in restaurants is the norms/expectations of service personnel. e.g. I'll walk into a restaurant and the young hostess might stare at me until I go "Uh, for two please". Or the waiter at a restaurant might make it feel like he's only standing in for his friend and never was a waiter before in his life before I release him back to the waiter social circle going on in the corner. But I can't blame that on "data".

> Consider what someone who doesn't use AI would take away reading that quote from the article, and what harnesses were actually made for.

Or you could just communicate the point that you have in your head yourself instead of hoping I do it for you and then arrive at your conclusion when I just reached my own different, independent conclusion after making the same consideration.

Man, how is everyone so wishy washy on this subject? Why not give us the blurb you would write instead for that article and audience?

Yet one of the reasons harnesses abstract tool calls is to control what agents can do instead of just yoloing commands and inline python scripts. It's such a central feature, and agents are so hard to control, that harnesses like claude code and codex use an AI classifier to additionally auto-approve tool calls.

It's one of the things it does, especially in the context of agents doing unexpected things.

I don't get what's so painful about that description. What would you write instead, specifically? The point is that the harness doesn't completely lock the agent down.

I also don't get what's "really weird" about the article showing up on HN. Should we be completely insulated from how tech topics and which stories show up in non-tech media?

They probably just don't see it worth the extra effort. And maybe they want to keep it in sync with the web UI.

It definitely is a fraction of the work these days to maintain three different desktop apps + two different mobile apps + a web app. But it's still a lot more work than an Electron app that is merely tweaked between platforms.

Whether it's the government or powerful companies that the government fails to limit, it is still an instance of the US failing its citizens.

> with a limp

> wheelchair

You really took "cripple the model" to heart!

The only place I've been where that might happen is rural mountainous Mexico where the "gas station" is a huachicolero funneling fuel into your car from a plastic water bottle next to a construction paper sign with "GAS" scribbled on it.

Everywhere else, gas stations aren't open if there's no power because nothing works, not even the cash register.

But I'm starting to see the uni freshman libertarian vibes here having been there once myself.

Without more knowledge about the technical aspect, it could be a good thing that they're all so similar. If you tell humans to go from their kitchen to their bedroom, they all stand up and walk the same way. Nobody decides to crab walk. Maybe you'd have to cripple the model in some way to do that.

On the other hand, if you want something different with LLMs, all it takes is a few more words of creative flair in the prompt.

The US has failed its people so badly here.

I heard a family member's spam call that used Trump's voice to personally ask for a donation. It made me wonder how many people might fall for that, especially people in cognitive decline. And the call was spoofed from their area code which is a dirty way to get people to answer.

It simply shouldn't be possible to spoof a call. Or spoof an IP packet origin.

But it's not an "invalid argument" to share your experience. Just like it's not an invalid argument to say you don't like chocolate ice cream.

"Just ignore people who reach out to you" and "Just don't get burned out" aren't pointing out an invalid argument, either. It's just peanut gallery nonsense.

Open source software is easy come, easy go. Part of that is accepting it when maintainers don't feel it's worth it anymore. And now with AI you basically have to choose between no contributors or a firehose of slop issues/PRs, something that changed overnight.

I don't see how AI needs StackOverflow anymore.

It can either examine the ground truth source code to answer your question "How to expire cookies using RoR Devise gem" or it can read docs for you or it can spin up local experiments to black box examine some software. If humans had done that before posting on StackOverflow, the question never would have made it there.

Its reasoning ability is long passed hoping an example exists online for it to copy.

Sure, at the end of the day you have to do your own research and consider whether any medical professional is under- or overreacting. It's why a second opinion makes sense.

I have naturally higher blood pressure such that I'm usually 135/75 in the doctor's office. But I also run every day, I'm fit, and I have a dialed in diet. No doctor has given me flak for it.

But I also see this thinking turn into empty contrarianism that is also dangerous.

I know some keto/carnivore grifters say high blood pressure is fine and high blood cholesterol is fine and that you should actually double down on table salt and that as long as your a1c is good then nothing matters. Now people trying to navigate their numbers are even more confused.

The 120/80 benchmark isn't lunacy. Most people in the west have issues with hypertension because that's the direction it goes due to the common risk factors.

If you're at 135/85, and you're middle aged / younger, and your diet and lifestyle aren't already optimized, it's something you should probably care about, not look for the next highest rung on the ladder so that you can celebrate for being below it.

I had to temporarily use a 1080p monitor, and forcing a higher resolution was the only way to get enough info on screen to do what I needed to do productively despite the blur.