Where is all the new, good software?
39 points by gerikson
39 points by gerikson
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.
Whether the quality of the software (by whatever measures that entails) matters most in the success of a given product has arguably always depended on the market. Look at MS products, for example.
However, for some customers quality matters more and they might choose a different product. I believe that's separate from whether AI is able to or intended to improve quality or just quantity.
AI tools let us focus more on features (fast and loose) or quality (more rigorous tests, time spent experimenting and refining, etc). All comes down to what the team / business wielding it chooses to focus their tokens efforts into.
Also: I think we're still, as an industry, figuring out how to best utilize AI for either of the above outcomes, so even if a company is trying to use AI to improve quality, they just may not have quite cracked that nut yet.
This would mean that AI coding could be effective, but still lead to overall worse software and generally poor outcomes.
Yeah, partly because it's now cheaper to produce it in the first place, and then cheaper to fix later on. There is less risk in creating worse software.
In the post, answers 5 and 6 are specific examples that are not about new customer driven software.
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.
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.
This feels similar to the move from bespoke clothing craftsmanship to industrial factories producing thowaway clothes.
Which, too, has environmental and societal repercussions, not unlike those associated with the use of LLMs.
I always thought the claim of "if LLM's are so great, where is all the new software?" to be a little funny. Pre-LLMs, we would look at a company's employee growth, like Twitter having 5k employees, and people would be like "how is it that this little webapp has so many people working on it, but no new features?" And the common refrain would be "It's scale! So much of the software written at twitter isn't to add a new feature, but to maintain infra, support moderation, etc".
Now we have a similar situation where companies have "grown" (in the sense that overall throughput has been non-trivially increased), and people still ask the same question, and I feel like the answer is still the same -- code is disproportionally allocated towards tooling, infra, etc, just like it always has been. Google Docs has barely changed in 10 years and it'll continue to barely change.
There has been a ton of cool new software that has come from LLMs. One category of work is game recomps and decomps, where people rebuild games (like the well-known Pokemon gen1recomp) to be self-sustaining executables with modding support. These efforts are almost entirely LLM-driven and have been very fun.
Personally I'm less interested in using LLMs to accelerate development of new features, and more interested them in using them to build a feature in the same amount of time, but with increased robustness and reliability.
The implied question behind "where is all the new software?" is often "how does this new technology justify its costs?" Growing the market with new, better software is the most obvious route to this in many peoples minds, but equally they could be asking "where is all the economically measurable productivity?"
That might all be in internal tooling, infra, etc, but it is not acceptable for this technology to barely change anything from an externally observable economic perspective, it has to be radically impactful in measurable ways for all this AI debt to be justified.
Tangentially, I have historically seen plenty of people complain about forks and rewrites when existing software became difficult to maintain. Do we want new software or better maintenance of existing software? If using LLMs can ease the maintenance burden, we should have less need for new software projects where we could instead improve existing ones.
You have to get the maintainers to agree to have their project improved by LLMs, which is not always easy, even when they aren't radically opposed to the technology.
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.
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
these are conjectural
A fancy way to say pulled out of thin air. Without any data, this post is completely meaningless
Give us the data then. Mention one software that got you impressed lately so we can all see it.
Some anecdata that supports what the original article says. People build more software to scratch their own itch - here's a few I built:
Each of these tools have a user base (and exposure) of <5 but they solve real problems real people have - which to me is the definition of "good software"
I suspect you're agreeing with each other but I'll bite. Strata has absolutely blown my socks off. Outrageous performance and attention to benchmarking, easy to use, built in very little time.
(Obviously it's a bit meta to link to AI-related software. I would like to have a different example, but the mundane explanation is that I already have favourite software for my other needs so I expect it will take some time before I'm compelled to use a "new, good" replacement.)
It is 100% meta.
I can give one example. I'm impressed by scriptC, typescript-to-native compiler vibecoded by vercel. It is not great yet but the fact it took only 2 months to get this far is amazing.
By the way, where the edge of meta lies? Does making dictation more convenient count, if I use a public-weights NN in the middle of the pipeline?
I want dictation, and it is much easier to set up fragment-by-fragment recognition than true streaming. All software I know for cutting a recording into fragments on silence uses absolute total levels. With any kind of background noise it is … meh.
I was clearly not getting around to learning ALSA interface Right Now™, so I asked a local LLM to capture, build a spectrum histogram via FFT, and compare to the background noise profile bucket-by-bucket. So much easier to set thresholds than total volume dB thresholds, it's not even funny.
It does do what I wanted it to do, it is truly vipe-coded slop, I skimmed it and slopped (or sometimes did by hand) a few fixes where I thought it's too much even for slop, some places it industriously made better than I would lazily do. But it is all to better feed data to Whisper. Is it ML-meta?
The idea is trivial, the implementation is small-ish, but nobody that I can easily find published code for it, and I am lazy, so here we are. I might put it into some publically reachable landfill, but maybe I should properly check first that it does not do anything weird with unrelated files. For own use, I sandbox it, so I wouldn't notice.
It is great but using software to solve our own issues isn't something as revolutionary as some people and corporations are painting. Going from a whole day researching to doing it in a couple hours is an optimization but does it really change everything?
Going from 0 to 1 was revolutionary like Excel/spreadsheets. Having custom software is mostly incremental because of the amount of automation we already got from "normal" software.
Is there any billion-dollar for-profit company marketing services to individuals which is not fraudulent in one way or another?
Honestly, my reading of the code was purely optional. I did not want to give a full testing setup to the model, live loudspeakers included. It mocked something to test on generated data. I still needed to test ALSA part, of course. But if I optimised for low involvement, I would probably go quite a bit below 2 hours active time (passive background time was a lot but I don't care much).
I'd say that even for me, who has written more than half a dozen languages on a job (sometimes I could pick tools on whimsy), this shifts quite a bit what annoyances can be automated away within the effort I am willing to spare. And if I use something weaker/faster for generation, debugging is more fun than typing the slog.
I think there will be a brief moment when local models will, full-vibecoding-mode, allow power users to get nice highly personal tools. Then this will get succesfully swept under the carpet, like all the previous options for quickly getting a custom tool.
Oh, and speaking of framing — if I wold you what software has ever impressed me ever, and got absolutely a part of the toolset I use or informed the design of a replacement — you wouldn't be impressed by most entries in my list.
crosspoint-reader, (and its forks: crossink, inx etc.) and the whole ecosystem around tiny e-readers has blown my mind.
All of the sloppy recompilation/emulation projects of older console exclusive games seem to be in generally impressive state. OTOH the content is the point, not the code.
Mention one software
Well, to toot my own horn, or rather that of my agent's which has churned out more research-grade software than I ever dreamed of writing myself; I like to think https://nina.family is pretty cool. I need to polish it up and do some introductory videos because it's just an absolute pile of fun stuff. It takes a while to load which is my next order of business.
Requires JavaScript with no <noscript> so stuck on an “Loading …” that didn’t work. Enabled JavaScript, page took like 5 minutes to “load” on pretty quick internet, then still a UI with a blank page—at least I thought but the text was so tiny & so little contrast that I didn’t even see it. Every click was super laggy in LibreWolf that I straight up thought it was broken. What are we tooting about—as I can’t even tell what it does? 😅
I took them as conversation starters / hypothetical answers. Not actual proposed answers to the question.
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.
Good list. A few more:
A major limit for new projects is still creativity and agency, things that agents (somewhat ironically) cannot produce on their own. So the number of people who create did not explode as much as the productivity of those who were already creating, and there's a harder limit to that
One year (since it became good) is not enough time for many kinds of new projects. Even with the speedup agents offer, good software takes time to build. The parallelization of agents suffers from quickly diminishing returns when used on the same project, especially when its architecture is still growing.
Never heard of these claims. Who said this? I have found software to have become worse and filled with bugs and security vulnerabilities (see how many incidents have been happening, look at Microsoft (GitHub, Windows, Teams, ...). I have never heard a software "expert" (let's call him that I guess) say that software has actually been becoming better. If anything, it has been getting larger (which, I don't know whether that's good or bad to be honest).
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?
Well, I'd say that some of the claims are not just that AI will make you more productive -- but that AI will "replace" your job and write software autonomously.
That's a different kind of claim than "Rust is a productive systems programming language"
C++ and Java were pretty big improvements on the state of the art. As a case, modern operating systems are better in almost every dimension from their ancestors.
Biggest modern operating system is mainly built in C though, with a smattering of Rust.
Linux? Yes, for sure. Though they sure don't use a decade old version of C or Rust. And Linus has been unambiguous about productivity gains of tools like git and syzkaller amongst others.
But, when I said "operating system", I meant far more than just a kernel.
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.
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!
I've been vibecoding a piece of software for my own use, and it's... good and not good at the same time.
It works pretty well. It solves a real need. It's novel – no existing solutions solve my exact problem, or else I would have opted to use those. It's been fast to develop and ~ instant to add new features as I discover myself needing them. The design is sound. So it's all good in those senses.
It's not good in that ... it's slop. It's ~200 KB of very LLM-y code that's hard to follow at times. It feels overengineered – I think the implementation can be simplified to the benefit of both the user and code quality.
I'll be open-sourcing it anyway shortly, and my plan is to then take my time to study the code carefully and make it closer to what I would have written manually. For now, my emotions about it are a quantum superposition of elation and grudge. Part of the elation is due to the fact that I can, and will, dogfood it to study itself!
It's ~200 KB of very LLM-y code that's hard to follow at times ... my plan is to then take my time to study the code carefully and make it closer to what I would have written manually.
And why do you care? Why not let the LLM just do the maintenance? Isn't (one of) the point(s) of using an LLM to move up the abstraction chain and not have to worry about the fiddly details of programming?
Because blindly letting them do the maintenance will sometimes rewrite unpredictable parts of code, and thus invalidate unknown parts of accumulated testing-in-production. Not knowing what has subtly changed and for what reason is a bit of commercial-desktop-OS experience all over again…
I’ve cared about code – about how cleanly it expresses my thoughts and the problem at hand – all my life. I still do. How could I not care?
Not to mention that AI software is practically banned from traditional distribution channels. Lobsters, Reddit, language-specific forums, codeberg, they treat AI-coding as a plague.
You have to slap a "vibecoded" or "ai-generated" tag (interpreted as "slop") to anything you build with AI, regardless of software quality and engineering depth. If you don't, a moderator will rush to do it.
Hopefully this is changing.
Not to mention that AI software is practically banned from traditional distribution channels.
[...]
Lobsters, Reddit, language-specific forums, codeberg, they treat AI-coding as a plague.
Of these, only codeberg is a "distribution channel", and it's very new, not traditional. The others are discussion fora. And they still allow discussion of slop, but the moderators (in some cases) and/or the other participants in the discussion insist that it be labeled accurately.
Github, which is the dominant distribution channel for source code, requires no such disclosure.
Hopefully this is changing.
Yeah, I hope so too. Disclosure of LLM usage should be required everywhere, and failure to disclose should be a bannable offense, instead of just having a moderator add a tag or something.
but the moderators (in some cases) and/or the other participants in the discussion insist that it be labeled accurately.
No, they insist that things be labeled inaccurately, using loaded terms designed to kill discussion before it happens. Attempts at accuracy are met with violence and ignorance.
Please take it down a notch. Nobody is getting beaten up over tags on Lobsters. Forum tagging is nowhere near that important.
If you don't like the existence of the tag, it would be a very short user script to hide it from view. If you don't like the existence of differing opinions, there are plenty of other sites available.
What's innacurate? People want LLM provenance to be labeled. You may quibble with the term (e.g. "vibecoding") chosen by the community, but it's the term the community settled on for describing the provenance, and it really boils down to identifying LLM provenance at the end of the day.
Attempts at accuracy are met with violence
Please show me an example of violence directed toward someone who wants accuracy in the description of software provenance! I've not witnessed that.
There is no "ai-generated" tag.
If you don't, a moderator will rush to do it.
Most tag changes to "vibecoding" are made by users: https://lobste.rs/moderations?moderator=%28All%29&what%5Bstories%5D=stories
Fair, but "vibecoding" is even more degrading.
I get it that each forum and each community have the right to set their rules, but then no one will share what they worked on for months there, if they have to slap some insulting label to it.
It's like asking every indie project to be tagged "amateur hour", and then wonder "where are the good indie projects".
I treat vibecoding as a neutral tag and decide whether to engage with the post based on the content. The tag has an unfortunate name (llm-usage would be better) but that doesn’t force me to interpret it either negatively or positively. I think folks who are active on here have figured that out by now.
Is what you're describing also the same thing we've run into with overly self-promotional accounts on the site?
A point I would add to this -- attackers are now empowered to hack automatically, intelligently, and persistently -- I have so much cool software I built which I want to show the world, but am worried about people exploiting any resources given to them for user-generated content to host malware or steal other user's data. When someone can just point opencode+glm at my endpoints and iterate until they get something, that makes it feel riskier to expose more complex projects.