Software Bonkers

https://craigmod.com/essays/software_bonkers/

by razin • 15 days ago

34 21 15 days ago

craigmod.com

xqb64• 15 days ago

Lost me at "Claude Code".

xyzsparetimexyz2• 15 days ago

Bit late to the party mate. Everyone cycled into llm psychosis in December and had cycled out by February or so

wffurr• 15 days ago

"cycled out" what. The tech world is still completely insane.

groby_b• 15 days ago

Yes, but we've traded LLM psychosis for agent psychosis, followed by loop psychosis. (Or was it the other way round? Also, we did dark software factories already, right?)

Turns out that if you try to be "15 minutes into everyone else's future", you get better and more frequent trips than any psychedelics could provide.

natbennett2• 15 days ago

This was posted in March 2026.

tonyedgecombe• 15 days ago

It’s all invalid then because the new models are soooo much better.

xyzsparetimexyz2• 15 days ago

Oh true. The submission should've been dated then. Shame on OP.

seblon• 15 days ago

Maybe not the most interesting article on HN, but I need to agree. Accounting in general is way too completex. For a Non business case (no tax involved), I decided to use hledger, but my "UI" to it is just an LLM. It was nerver so simple to make "okay enough" book keeping.

natbennett2• 15 days ago

The phenomenon this describes is the most important effect of LLM coding tools, and this is my favorite description of it.

danshipt• 15 days ago

As impressive as it sounds, I don’t think this is representative of the AI panorama. It’s difficult for non tech people to build tailored software just like that, and unless things change drastically, I don’t see the majority of the population building their own software just like they use mobile phones

0gs4• 15 days ago

yeah, i disagree. i think i am essentially ~non-technical and it is not difficult to build one-user applications at all. it's not even difficult to build "real" software!

klibertp2• 15 days ago

Yesterday, Codex helped me resurrect a project[1] from 2013, where the code was published, but trapped inside an exotic format that can only be read by a single application. That application had ~14 breaking releases (rewrite everything!) over the past 13 years. Needless to say, no current version of the app can read the 2013 file, and no current system can run version 2.0 (it's now at version 14) of the app in 2026. Codex helped me create a containerized environment that pulled dependencies from that era, connected the container to my host GUI session, and successfully exported the code to normal-ish text files. It took half an hour (of Codex time, not mine) to check changelogs and try different version combinations. I was quite happy with the results.

Today, I wanted to capture screenshots every 5 seconds on a Windows machine. Codex created a dotNet project that compiled to a (self-contained!) bundle. It worked, of course, and I even ended up using it. The problem was that the bundle was 167 MB in size, held together by lots of XML config, and implemented the "RBGA-bitmap to PNG file" dumping from scratch; the whole thing was close to 1.5k lines of code. 30 LOC of PowerShell would be a perfectly acceptable solution here. I was... well, not impressed, and I would never show this solution to anyone, but it at least works and I didn't need to brush off my PS skills as a result, which is kind of good enough in this specific case.

As a programmer, I can immediately recognize when the LLM output is something nobody should ever see. It doesn't matter if nobody's going to see it anyway, I agree - yes, "one-user applications" fall in that category. However, when you mention <<"real" software>>, that changes the perspective. How would you know whether your software is "real"(-ish) or not? "Hey Claude, are we production ready enough right now?" That won't work.

TL;DR: requirements for code quality are hard to estimate, and code quality is something many professional programmers struggle to consistently recognize (not to mention achieve). Until the models get good enough to make good code a baseline, "real" software will remain hard to write, and better left to people who can recognize bad code quickly. That doesn't mean only professional programmers: plenty of hobbyist coders are "non-technical" in general. But at least for now, you still need to be a coder yourself to get consistently good results from the AI.

[1] https://pleiad.cl/research/software/gradualtalk

nylonstrung2• 15 days ago

I think this is true if you're the kind of "non-technical" person who reads HN which is still the minority

The smallest amount of framing and architectural forethought pays massive dividends but I imagine the person who says "build me an accounting app" while being apathetic to what language and stack it uses like apps such as Loveable imagine will still get bad results

klibertp2• 15 days ago

I think the end goal the author describes:

I can imagine soon-to-arrive interfaces where you just drag and drop components while narrating your desires with your voice, the models able to perform the “brainstorming,” “planning,” and “work” — operations that can take ten minutes or longer today — in mere seconds tomorrow.

Is impossible without ASI and more. This very vision failed to materialize in the past (4th-generation languages?) and still routinely fails, not only with models/agents, but also with humans and entire teams of humans on the other side. For a model to be that useful, ASI is the first enabling factor, but it also needs to develop mind-reading hardware and software and convince people that it's not breaching their privacy. I don't think this is going to happen. Not this millennium, at least.

There is a good case for prototypes, MVPs, and personal mods. I can imagine more and more users making use of the freedom to modify the open-source code. Technically, it was always possible, but for a normal user, it was not a realistic choice (learn to code (long and hard) or hire a programmer (expensive and inconvenient)). Even for me - a programmer by trade - fixing bugs in random pieces of software I might use (or not) once in a while was something I very rarely had the spare time and energy to do. The capabilities of the current crop of AI models/harnesses make this WAY easier: cheap on a subscription and requiring much less of my time. But that's for personal code modification. Pushing vibe-coded changes to anywhere outside of my machine is still something I wouldn't do, because I can see how mediocre the output code is. Unless the models can always write code as good as the top 20% of human-written code, their output will remain a liability. It's OK if I'm the only user; it's wrong to push such a problem to others. The issue here is that non-programmers cannot recognize when the code is good enough, which is why I used the word "always" - otherwise, pushing LLM code to others is a coin toss whether it's helpful or detrimental (for the project, the maintainers, and other users).

0gs4• 15 days ago

ARE we still in the dorks-only phase? i am certainly a gigantic loser, but i am not sure i am the sort of dork being invoked and i have been doing and thinking about the exact same kinda stuff. i think the genie is out of the bottle or whatever

andrekandre• 15 days ago

reading articles like this, i really wish people would before headlines before a group of paragraphs so i can scan the progression/points and see if its worth reading... a wall of paragraphs and a title like "software bonkers" doesn't help

just my 2c

nylonstrung2• 15 days ago

I think the overarching trend is "Buy v. Build" has reversed as hard as Roe v. Wade

If we haven't already crossed this point, the time that goes into software procurement, implementation, hand-off with the vendor, talking to support, getting customization will be less than just making something turnkey that solves exactly your problems

But we're definitely at the point already where building something quickly with AI is a already much more fun and rewarding use of time for any semi-technical person

calldacopsidgaf• 15 days ago

okay

Software Bonkers

Related Stories