Software teams are capturing a lot of the speedup in the form of more leisure or quasi-leisure time. (To greatly simplify: software teams are offering 2x speedups to their companies. They’re then using tools that allow them to work 10x as efficiently, and taking nice long coffee breaks.)

Hard agree. LLMs make it easier to work through a problem while doing chores at home. It doesn’t speed up the review/QA pass and I don’t see that really changing until corpos are okay with not having any human accountability. So long as SWEs are on the hook for incidents in production, we are going to spend the time to QA.

score 1 · canonical

these are conjectural

A fancy way to say pulled out of thin air. Without any data, this post is completely meaningless

score 1 · canonical

I think there’s a pretty large percentage of software produced by AI that is purpose-built for the person who prompted it. Like little dashboards, widgets, or whatever that visualize data or do something really specific to their job/workflow that they aren’t sharing with others.

This is kind of wasteful, but when you only make it for yourself it’s probably a lot less work/overhead/burden than making a tool for everyone

score 1 · canonical

citation needed

score 1 · canonical

We’re making more software instead of making software better

That’s the key line, I think. More of the same isn’t progress. LLMs double down on the current paradigm: a separate codebase for every application, which then has to be maintained. A large codebase is a liability, not an asset, and LLMs grow it faster.

What we need is a new paradigm: generic, data-driven platforms where the application is data, run by one well-tested engine.

score 1 · canonical

None of these questions will have answers until the bubble bursts, the investment scams stop messing with the price signals of cloud hosted LLMs, and we can figure out what “AI” actually costs so we can figure out where it’s worth the trouble and where it’s not.

score 1 · canonical

Software teams are capturing a lot of the speedup in the form of more leisure or quasi-leisure time. (To greatly simplify: software teams are offering 2x speedups to their companies. They’re then using tools that allow them to work 10x as efficiently, and taking nice long coffee breaks.)

There’s an important flip side to this, which is the actual mental model overhead.

Unless you’re doing slop, you’ll need to think about your code. That 2-3x speedup might actually be the limit of your cognition, at least in matters of direct human-on-keyboard work for your employer. Those coffee breaks are necessary, and sometimes downright essential.

For the rest of the post - pretty much (a bit of) all of the above. I’ve conjectured that we’ll have far more smaller software and abandon SaaS that isn’t exceptional, and it seems to be happening. I’ve seen a fair few overblown SaaS companies needing to scale down.

I, personally, have coded (assisted with LLMs) a few personal services. I’m not overly motivated to publish them widely or spend marketing/sales budget, but it was fun and brings me utility every day. Can’t complain about the impact of LLM software on myself. Services I use daily, though, don’t seem as improved as I’d enjoy.

score 1 · canonical

AI doesn’t work well with humans, and I believe that’s an early manifestation of a misaligned AI.

AI is very good at churn out a large amount of code that achieves some predefined goal. This is very evident by looking at the number of game porting/decompilation projects, and software reverse engineering projects that sprung up overnight like mushrooms. The problem is the code they write is horrendous, and this sentiment is echoed by mathematicians who have read the math proofs generated by AI as well[1].

Direct consequence of this is that it’s basically impossible to integrate AI into a traditional software engineering process. The amount of code AI churns out is huge, the quality of the code in terms of human readability and maintainability is so terrible, code review just cannot catch up. [2]

And fundamentally this is because AI just does not understand what a human wants. In other word, this is an AI alignment problem.

[1]: Tristan Buckmaster, talking about his work on the Euler Equation: I can say the first LLM generated proof Levent sent me was the most horrendous I have ever read; we verified it on Lean on August 22nd. Since this point, we have been working around the clock to understand this proof and turn it into something readable.

[2]: https://arstechnica.com/ai/2026/10/ai-coding-agents-generate-more-code-but-not-more-software/ A recent study of actual coding practices across hundreds of firms finds that human code review forms a significant “bottleneck” for the overall efficiency of AI coding tools, resulting in “little evidence that firms increase software output or reduce employment” by using them.

score 1 · canonical

This has been a pretty longstanding problem of evil style paradox with programming productivity claims long before AI. If e.g. something like TDD, or Rust, or functional programming, or whatever lets us in some sense build better and more correct software, why is the software world running on decades old versions of C++ and Java?

score 1 · canonical

The problem with posts like this is they are just claims with no evidence.

I don’t like to take sides on what is essentially reality. AI exists and people use it. For them, it is useful. That’s not really a debate. If the benefits outweigh the negatives on the grand scheme of things, becomes a philosophical question.

On the specific case of software being produced here and now, it is worth noticing that the problems people attribute to AI, were already manifesting themselves in industrial scale before the advent of AI coding tools.

Reinventing the wheel everywhere, software riddled with bugs every other click, UIs and workflows driven by trends and how slick they look rather than how effective they are at a specific task. Corporate Godzilla sized distributed systems that often could be called down to a $100 server. Gigabyte size binaries that pack the same functionality than a 200kB program written in the 1990s.

On the face of all this, we blame AI… I don’t buy it.

score 1 · canonical

I think LLMs help the industry to speed up the race towards the bottom in terms of acceptable quality and innovation. Despite the opinions and feelings of tech enthusiasts and (some) software engineers,

a) people tend to accept very bad software products and learn to use it, b) people often have no choice, literally (crappy government app) or virtually (e.g. too much of a hassle to switch banks)

Today it’s possible to produce low-quality barely functional software, cheap and fast, and… it’s okay. Business-wise, I mean. It’s ok. It works. Relax, nerd.

This feels similar to the move from bespoke clothing craftsmanship to industrial factories producing thowaway clothes.

score 1 · canonical

Some of these may be true in the fullness of time.

However, I suspect the question is wrong. “Where is all the good new software?” belies an assumption that the point of the software economy is to produce utility for the consumer. AI may justify itself via a reduction in the labour expense of existing entrenched companies, where the utility of what they produce is almost a secondary concern and can even be allowed to fall so long as their entrenchment can bear it.

This would mean that AI coding could be effective, but still lead to overall worse software and generally poor outcomes.

If true, this is a societal conflict of interest that needs to be resolved, probably by means other than market forces.

score 2 · canonical

We are in fact seeing software improvement, in the form of fewer bugs and better performance.

In general, I feel like there are more bugs now. More software behaves erratically or in subtly wrong ways and my hunch is that it’s due to one of these “bug fixes and performance improvements” from LLMs, but lack of tests or (whatever) caused a regression.

Often when raising that something is broken, the answer is usually “Claude can fix it” (work around). Which…if we build a culture of workarounds for workarounds then finding and fixing the root of problems becomes socially harder and harder. LLMs are good at finding bugs, they’re not always good at gauging the relative importance of which are trivial and which are important.

You can write high rigor software aided by LLMs, but it is not the norm or the default for most users.

score 2 · canonical

Without LLMs it would be simply impossible to use OpenBSD as a router for me ;-)

Now not only I do, I can for the first time connect back home without VPNs just by automating firewalls whitelisting and automatically updating SRV records :-)

Oh, I have a network flow monitoring system running on 1GB of ram RPi3. This view required 4-16GB of RAM. Ofc I can’t visualize 7 days at low resolution yet, although with prom federation enabled (as of late) might be possible :-)

I mean, it has been fun fun fun!

score 1 · canonical