Eight years of wanting, three months of building with AI
lalitm.com
lalitm.com
This experience is familiar to every serious software engineer who has used AI code gen and then reviewed the output:
> But when I reviewed the codebase in detail in late January, the downside was obvious: the codebase was complete spaghetti14. I didn’t understand large parts of the Python source extraction pipeline, functions were scattered in random files without a clear shape, and a few files had grown to several thousand lines. It was extremely fragile; it solved the immediate problem but it was never going to cope with my larger vision,
Some people never get to the part where they review the code. They go straight to their LinkedIn or blog and start writing (or having ChatGPT write) posts about how manual coding is dead and they’re done writing code by hand forever.
Some people review the code and declare it unusable garbage, then also go to their social media and post how AI coding is completely useless and they’re not going to use it for anything.
This blog post shows the journey that anyone not in one of those two vocal minorities is going through right now: A realization that AI coding tools can be a large accelerator but you need to learn how to use them correctly in your workflow and you need to remain involved in the code. It’s not as clickbaity as the extreme takes that get posted all the time. It’s a little disappointing to read the part where they said hard work was still required. It is a realistic and balanced take on the state of AI coding, though.
One thing I will add: I actually don’t think it’s wrong to start out building a vibe coded spaghetti mess for a project like this… provided you see it as a prototype you’re going to learn from and then throw away. A throwaway prototype is immensely useful because it helps you figure out what you want to build in the first place, before you step down a level and focus on closely guiding the agent to actually build it.
The author’s mistake was that he thought the horrible prototype would evolve into the real thing. Of course it could not. But I suspect that the author’s final results when he did start afresh and build with closer attention to architecture were much better because he has learned more about the requirements for what he wanted to build from that first attempt.
I often see criticism towards projects that are AI-driven that assumes that codebase is crystalized in time, when in fact humans can keep iterating with AI on it until it is better. We don't expect an AI-less project to be perfect in 0.1.0, so why expect that from AI? I know the answer is that the marketing and Twitter/LinkedIn slop makes those claims, but it's more useful to see past the hype and investigate how to use these tools which are invariably here to stay
That's a big leap of faith and... kinda contradicts the article as I understood it.
My experience is entirely opposite (and matches my understanding of the article): vibing from the start makes you take orders of magnitude more time to perfect. AI is a multiplier as an assistant, but a divisor as an engineer.
1. Autocomplete. Pretty simple; you only accept auto-completes you actually want, as you manually write code.
2. Software engineering design and implementation workflow. The AI makes a plan, with tasks. It commits those plans to files. It starts sub-agents to tackle the tasks. The subagents create tests to validate the code, then writes code to pass the tests. The subagents finish their tasks, and the AI agent does a review of the work to see if it's accurate. Multiple passes find more bugs and fix them in a loop, until there is nothing left to fix.
I'm amazed that nobody thinks the latter is a real thing that works, when Claude fucking Code has been produced this way for like 6 months. There's tens of thousands of people using this completely vibe-coded software. It's not a hoax.
also Claude Code is notoriously poorly built, so I wouldn't tout it as SOTA
Get the best programmer in the world. Have them write the most perfect source code in the world. In 10 years, it has to be completely rewritten. Why? The designer chose some advanced design that is conceptually superior, but did not survive the normal and constant churn of advancing technology. Compare that to some junior sysadmin writing a solution in Perl 5.x. It works 30 years later. Everyone would say the Perl solution was of inferior quality, yet it provides 3x more value.
but I'm not judging Claude Code by how it looks. I kinda like the aesthetics. I'm talking about how slow, resource hungry and finnicky/flickery it is. it's objectively sloppy
And people can look at the results (illegally) because that whole bunch of code has been leaked. Let's just say it's not looking good. These are the folks who actually made and trained Claude to begin with, they know the model more than anyone else, and the code is still absolute garbage tier by sensible human-written code quality standards.
(Anthropic, of course, believes that advances in AI capability over the next few years will so radically reshape society that there's no point worrying about the long term.)
Previously, takes were necessarily shallower or not as insightful ("worked with caveats for me, ymmv") - there just wasn't enough data - although a few have posted fairly balanced takes (@mitsuhiko for example).
I don't think we've seen the last of hypers and doomers though.
Ironically this itself is one of the hyper/doomer takes.
There seems to be a concencus among people I follow/read, that somewhere around that time was an inflexion point in coding LLMs, and this matches my personal experience.
(My comment was in the context of LLMs being used to generate non-throwaway code, not general GenAI use - apologies if that was unclear.)
LLMs have been a very useful tool to generate non-throwaway code for at least 2 years. It's a question of Overton Window, of mainstream acceptance, that has made you and the people you're talking about come to a "consensus" that there was an agentic LLM coding step change around that time. There was indeed an inflection point, but it was a social one: a window shift, where it became socially acceptable among mainstream SWEs to hold this sentiment. It had already been the case for a long time, but for those who understood this, expressing that was still not deemed acceptable. Doing so would brand one as a hyper, at least in the particular circles you were in.
You were in reality part of the doomers, just as much as the current doomers are still doomers.
I was in YC W24 (exactly two years ago) with an AI coding startup ("agentic" and "vibe" buzzwords weren't invented yet). LLMs were useful as smarter autocomplete, but were struggling (and us with them!) on anything nontrivial.
Comparing that state to the past few months and saying it's just a social change is a bit of a stretch :)
Source: Multiple successful prodiction applications with >90% LLM generated code already in Late 2024/Early 2025. Not slop, not note taking apps. Nothing ground-breaking but real-world valuable usecases. Both by myself and by others.
Aider came out May 2023. Augment Code (not very well known but best in class for a long time, hilariously poor at marketing/sales completely bungling this advantage) was April 2024.
I’ve been driving Claude as my primary coding interface the last three months at my job. Other than a different domain, I feel like I could have written this exact article.
The project I’m on started as a vibe-coded prototype that quickly got promoted to a production service we sell.
I’ve had to build the mental model after the fact, while refactoring and ripping out large chunks of nonsense or dead code.
But the product wouldn’t exist without that quick and dirty prototype, and I can use Claude as a goddamned chainsaw to clean up.
On Friday, I finally added a type checker pre-commit hook and fixed the 90 existing errors (properly, no type ignores) in ~2 hours. I tried full-agentic first, and it failed miserably, then I went through error by error with Claude, we tightened up some exiting types, fixed some clunky abstractions, and got a nice, clean result.
AI-assisted coding is amazing, but IMO for production code there’s no substitute for human review and guidance.
Personally, I think it's just the natural flow when you're starting out. If he keeps going, his opinion is going to change and as he gets to know it better, he'll likely go more and more towards vibecoding again.
It's hard to say why, but you get better at it. Even if it's really hard to really put into words why
have you ever learned a skill? Like carving, singing, playing guitar, playing a video game, anything?
It's easy to get better at it without understanding why you're better at it. As a matter of fact, very very few people master the discipline enough to be able to grasp the reason for why they're actually better
Most people just come up with random shit which may or may not be related. Which I just abstained from.
This is something everyone who cares about improving in a skill does regularly - examine their improvement, the reasons behind it, and how to add to them. That’s the basis of self-driven learning.
And that's not really explainable without exploring specific examples. And now we're in thousands of words of explanation territory, hence my decision to say it's hard to put it into words.
For instance, if I say “I noticed I run better in my blue shoes than my red shoes” I did not learn anything. If I examine my shoes and notice that my blue shoes have a cushioned sole, while my red shoes are flat, I can combine that with thinking about how I run and learn that cushioned soles cause less fatigue to the muscles in my feet and ankles.
The reason the difference matters is because if I don’t do the learning step, when buy another pair of blue shoes but they’re flat soled, I’m back to square one.
Back to the real scenario, if you hold on to your ungrounded intuition re what tricks and phrasing work without understanding why, you may find those don’t work at all on a new model version or when forced to change to a different product due to price, insolvency, etc.
Pursued far enough, any line of thought will reach something non-deterministic - or, simply, That's The Way It Is - however unsatisfying that is to those of us who crave straightforward answers. Like it or not, our ground truth as human beings ultimately rests on intuition. (Feel free to say, "No, it's physics", or "No, it's maths", but I'll ask you if you're doing those calculations in your head as you run!)
If you want to say "god is responsible for creating the precipitation cycle", sure. But we don't disregard understanding that exists to substitute intuition.
Yeah, "God did it" is the first of those answer layers at which some people stop interrogating the world around them, just like "that's just the way I am" is where some people stop developing their self-understanding. Neither of those answers advance civilization / ourselves any further than the status quo. They're terrible answers! Everyone should be digging deeper.
However, I would not use the word "understanding" in opposition to "intuition". Someone who can generate a ballistics chart understands trajectories, but so does someone who can reliably put a basketball through a hoop or a bullet on target. I would set "analysis" against "intuition" (or "instinct", if you prefer), but they're not in opposition: instead, they reinforce each other. We're all familiar with the scientists and mathematicians who ride a hunch to a ground-breaking discovery, which is then validated by exhaustive analysis. From the other direction, athletes and musicians analyze their technique in minute detail, and practice incessantly, in order to ingrain analytical insights into instinct. (Or, if you prefer a less physical example, programmers study algorithms so that they can intuit which to apply to a particular problem.)
My point - badly expressed in my earlier comment - is that as humans we exist moment-by-moment, and as such react, in each moment, by intuition. As important as analysis is, we cannot live in analytical mode: it lags too much! Furthermore, approximately none of us will ever make a groundbreaking discovery in any field, far less in all of the areas to which we can (and should!) direct our analytical energy. At some point we have to stop (even if we are a groundbreaking genius in one area, we'll have to in all of the others), and accept the answer that satisfys our purpose or exhausts our motivation.
You are just making stuff up or regurgitating material from a pop science book.
Please explain walking to me so that I can explain it to a person who forgot how to walk such that he can walk after the explanation.
Mind you, I don't think the process of improvement in those dimensions is fundamentally different, just much less direct and not easily (or perhaps even at all) articulable.
Its not random that AI happens to be built by the very same people that turned internet forums into the most addictive communication technology ever.
I think "more and more" is doing some very heavy lifting here. On the surface it reads like "a lot" to many people, I think, which is why this is hard to read without cringing a bit. Read like that it comes off as "It's very addictive and eventually you get lulled into accepting nonsense again, except I haven't realized that's what's happening".
But the truth is that this comment really relies entirely on what "more and more" means here.
Then use ideation to architect, dive into details and tell the AI exactly what your choices are, how certain methods should be called, how logging and observability should be setup, what language to use, type checking, coding style (configure ruthless linting and formatting before you write a single line of code), what testing methodology, framework, unit, integration, e2e. Database, changes you will handle migrations, as much as possible so the AI is as confined as possible to how you would do it.
Then, create a plan file, have it manage it like a task list, and implement in parts, before starting it needs to present you a plan, in it you will notice it will make mistakes, misunderstand some things that you may me didn’t clarify before, or it will just forget. You add to AGENTS.md or whatever, make changes to the ai’s plan, tell it to update the plan.md and when satisfied, proceed.
After done, review the code. You will notice there is always something to fix. Hardcoded variables, a sql migration with seed data that should actually not be a migration, just generally crazy stuff.
The worst is that the AI is always very loose on requirements. You will notice all its fields are nullable, records have little to no validation, you report an error when testing and it tried to solve it with an brittle async solution, like LISTEN/NOTIFY or a callback instead of doing the architecturally correct solution. Things that at scale are hell to debug, especially if you did not write the code.
If you do this and iterate you will gradually end up with a solid harness and you will need to review less.
Then port it to other projects.
Sounds like a solid way to make crud web apps though.
Holier than thou “yeah well I work on stuff that doesn’t use databases, checkmate!” doesn’t really land - data still gets moved around somehow, and often over a network!
For that I usually get it reviewed by LLMs first, before reviewing it myself.
Same model, but clean session, different models from different providers. And multiple (at least 2) automated rounds of review -> triage by the implementing session -> addressing + reasons for deferring / ignoring deferred / ignored feedbacks -> review -> triage by the implementing session -> …
Works wonders.
Committing the initial spec / plan also helps the reviewers compare the actual implementation to what was planned. Didn’t expect it, but it’s worked nicely.
And that's also a decent barometer for what it's good at. The more amount of critical assumptions AI needs to make, the less likely it is to make good ones.
For instance, when building a heat map, I don't have to get specific at all because the amount of consequential assumptions it needs to make is slim. I don't care or can change the colors, or the label placement.
But it's not all bad news. TIL about Parameters<T>.
So you do agree? If you are having to review and correct then it's not really the LLM writing it anymore. I have little doubt that you can write good Typescript, but that's not what I said. I said LLMs cannot write good Typescript and it seems you agree given your purported actions towards it. Which is quite unlike some other languages where LLMs write good code all the time — no hand holding necessary.
What else would you need to correct? I've never had trouble with LLMs generating basic syntax in any language. Architecture is exactly the aspect of language where LLMs seem to like to go to crazytown when in Typescript. It seems you've noticed too if the ideas in that area have had to come all from you.
Exactly. You can write good Typescript, no doubt, but LLMs cannot. This is not like some other languages where LLM generated code is actually consistently good without needing to become the author.
If you set up restrictive linters and don't explicitly prohibit agents from adding inline allows, most LOC will be allow comments.
Based on this learning, I've decided to prohibit any inline allows. And then agents started doing very questionable things to satisfy clippy.
Recent example:
- Claude set up a test support module so that it could reuse things. Since this was not used in all tests, rust complained about dead_code. Instead of making it work, claude decided to remove test support module and just... blow up each test.
If you enable thinking summaries, you'll always see agent saying something like: "I need to be pragmatic", which is the right choice 50% of the time.
This should be done on day one with a company-wide skill or project template that defines hard limits and processes for the Agent.
Strict linters, formatters and code quality checks are essential to de-slopify the code as much as possible.
That doesn't fix bad design though, that's still on humans.
What’s really happening is that you’re all of those people in the beginning. Those people are you as you go through the experience. You’re excited after seeing it do the impossible and in later instances you’re critical of the imperfections. It’s like the stages of grief, a sort of Kübler-Ross model for AI.
Professional software engineers like many of us have a big blind spot when it comes to AI coding, and that's a fixation on code quality.
It makes sense to focus on code quality. We're not wrong. After all, we've spent our entire careers in the code. Bad code quality slows us down and makes things slow/insecure/unreliable/etc for end users.
However, code quality is becoming less and less relevant in the age of AI coding, and to ignore that is to have our heads stuck in the sand. Just because we don't like it doesn't mean it's not true.
There are two forces contributing to this: (1) more people coding smaller apps, and (2) improvements in coding models and agentic tools.
We are increasingly moving toward a world where people who aren't sophisticated programmers are "building" their own apps with a user base of just one person. In many cases, these apps are simple and effective and come without the bloat that larger software suites have subjected users to for years. The code is simple, and even when it's not, nobody will ever have to maintain it, so it doesn't matter. Some apps will be unreliable, some will get hacked, some will be slow and inefficient, and it won't matter. This trend will continue to grow.
At the same time, technology is improving, and the AI is increasingly good at designing and architecting software. We are in the very earliest months of AI actually being somewhat competent at this. It's unlikely that it will plateau and stop improving. And even when it finally does, if such a point comes, there will still be many years of improvements in tooling, as humanity's ability to make effective use of a technology always lags far behind the invention of the technology itself.
So I'm right there with you in being annoyed by all the hype and exaggerated claims. But the "truth" about AI-assisted coding is changing every year, every quarter, every month. It's only trending in one direction. And it isn't going to stop.
Strong disagree. I just watched a team spend weeks trying to make a piece of code work with AI because the vibe coded was spaghetti garbage that even the AI couldn’t tell what needed to be done and was basically playing ineffective whackamole - it would fix the bug you ask it by reintroducing an old bug or introducing a new bug because no one understood what was happening. And humans couldn’t even step in like normal because no one understood what’s going on.
In 1998, I'm sure there were newspaper companies who failed at transitioning online, didn't get any web traffic, had unreliable servers crashed, etc. This says very little about what life would be like for the newspaper industry in 1999, 2000, 2005, 2010, and beyond.
It is absolutely the case that virtual reality technology will only get better over time. Maybe it'll take 5, or 10, or 20, or 40 years, but it's almost a certainty that we'll eventually see better AR/VR tech in the future than we have in the past.
Would you bet against that? You'd be crazy to imo.
Whether what they're using in 20 years is produced by the company formerly known as Facebook or not is a whole different question.
See https://simonwillison.net/guides/agentic-engineering-pattern...
All else being equal, ofc you'd rather have good code than bad code.
But millions of non-developers who can suddenly build simple software for themselves (or use AI assistants that generate simple ephemeral software on the fly) aren't going to care about the underlying code quality so long as it works, which it does.
And hundreds of thousands of software engineers, who can suddenly build singlehandedly what in the past took a team of 5-10 to build, are going to be okay with the tradeoff of getting massive speed boosts but with quality not quite as high as if they were to build everything themselves. Same for software engineers who are now churning through their list of side project ideas, which has mostly sat dormant for the past 10 years.
Previously they lived in a world where the only axes were, "Build more things vs have good quality" or "Build ambitious things vs have good quality." Now they have a new axis which is, "Allow AI to help vs have total personal control of quality," which is quite similar to the pre-existing axis of, "Hire employees vs have total personal control of quality." But way cheaper and more accessible. Of course some will take advantage. Which means, de facto, a world where code quality (or at least personal control of code quality) is on average less important and prioritized than it used to be.
We're seeing all of this happen right now. People are making these choices in large numbers today.
Guns, wheels, cars, ships, batteries, televisions, the internet, smartphones, airplanes, refrigeration, electric lighting, semiconductors, GPS, solar panels, antibiotics, printing presses, steam engines, radio, etc. The pattern is obvious, the forces are clear and well-studied.
If there is (1) a big gap between current capabilities and theoretical limits, (2) huge incentives for those who to improve things, (3) no alternative tech that will replace or outcompete it, (4) broad social acceptance and adoption, and (5) no chance of the tech being lost or forgotten, then technological improvement is basically a guarantee.
These are all obviously true of AI coding.
It isn't even a good job of cherry picking: we never got mainstream supersonic passenger aircraft after the Concorde because aerospace technology hasn't advanced far enough to make it economically viable and the decrease in progress and massively increasing costs in semiconductors for cutting edge processes is very well known.
There's no broad social acceptance of supersonic flight because it creates incredibly loud sonic booms that the public doesn't want to deal with. And despite that, it's still a bad counterexample, as companies continue to innovate in this area e.g. Boom Supersonic.
At best you can say, "It's taking longer than expected," but my point was never that it will happen on any specific schedule. It took 400 years for guns to advance from the primitive fire lances in China to weapons with lock mechanisms in the 1400s. Those long time frames only prove my point even more strongly. Progress WILL happen, when there is appetite and acceptance and incentive and room to grow, and time is no obstacle. It's one of the more certain things in human history, and the forces behind it have been well studies.
Just as certain: the people and jobs who are obsoleted by these new technologies often remain in denial until they are forgotten.
We're obviously talking about 1-10 years here, not 100-1000 years.
AI will get better at making good maintainable and explainable code because that’s what it takes to actually solve problems tractably. But saying “code quality doesn’t matter because AI” is definitely not true both experientially and as a prediction. Will AI do a better job in the future? Sure. But because their code quality improves not because it’s less important.
Where we're headed is toward a world where a ton of software is ephemeral, apps literally created by AI out of thin air for a single use, and then gone.
Which is to say, not at all.
Original wiring done by a professional, later changes by “vibe electrician” homeowners.
Every circuit might be a custom job, but they all accumulate into something a SWE calls “technical debt”.
Don’t like how the toaster and the microwave are on the same circuit even though they are in different parts of the kitchen? You’re lucky if you can even follow the wiring back to the circuit box to see how it was done. The electrical box is so much of a mess where would you even run a new circuit?
That’s the future we’re looking at.
Not all software is meant to be some permanent building block upon which other software sits.
When new technology arrives that makes earlier ways of doing things obsolete, the consistent pattern throughout history has been that existing experts and professionals significantly underestimate the changes to come, in large part because (a) they don't like those changes, and (b) they're too used to various constraints and priorities that used to be important but no longer are. In other words, they're judging the new tech the lens of an older world, rather than through the lens of a newer world created by the new tech.
It can hold a single flower pot, will fall over if a kid tries to climb on it, is made from wood that's delicious to dogs and most likely won't last 25 years because it wasn't finished properly.
But the person who made it needed a table for a flower pot, doesn't have kids or pets and will happily build a new one in 10 years when the old one breaks down from bad joins.
Not every piece of shitty software has a massive attack surface the will immediately kill people if it operates wrong.
Is my personal cheap imitation of Hazel[0] I vibe-coded in two evenings perfectly bug free and does it replicate every feature of Hazel perfectly? Of course not, but it does the exact 5 things I need it to do and saved be the upgrade price to Hazel 6.
The death of newspapers is quite the spectacle too. No one seems to understand how bad it is... the youngest generation can't even seem to recognize that anything is missing. We've effectively amateurized journalism so that only grifters and talentless hacks want to attempt it, and only in tiny little soundbites on Twitter or other social media (and they're quickly finding out how it might be more lucrative to do propaganda for foreign governments or MLM charlatanism). When the death of the software industry is complete, it too will have been completely amateurized, the youngest generation will not even appreciate that people used to make it for a living, and the few amateurs doing it will start to comprehend how much more lucrative it will be to just make poorly disguised malware.
It's the opposite, code quality is becoming more and more relevant. Before now you could only neglect quality for so long before the time to implement any change became so long as to completely stall out a project.
That's still true, the only thing AI has changed is it's let you charge further and further into technical debt before you see the problems. But now instead of the problems being a gradual ramp up it's a cliff, the moment you hit the point where the current crop of models can't operate on it effectively any more you're completely lost.
> We are in the very earliest months of AI actually being somewhat competent at this. It's unlikely that it will plateau and stop improving.
We hit the plateau on model improvement a few years back. We've only continued to see any improvement at all because of the exponential increase of money poured into it.
> It's only trending in one direction. And it isn't going to stop.
Sure it can. When the bubble pops there will be a question: is using an agent cost effective? Even if you think it is at $200/month/user, we'll see how that holds up once the cost skyrockets after OpenAI and Anthropic run out of money to burn and their investors want some returns.
Think about it this way: If your job survived the popularity of offshoring to engineers paid 10% of your salary, why would AI tooling kill it?
What you're missing is that fewer and fewer projects are going to need a ton of technical depth.
I have friends who'd never written a line of code in their lives who now use multiple simple vibe-coded apps at work daily.
> We hit the plateau on model improvement a few years back. We've only continued to see any improvement at all because of the exponential increase of money poured into it.
The genie is out of the bottle. Humanity is not going to stop pouring more and more money into AI.
> Sure it can. When the bubble pops there will be a question: is using an agent cost effective? Even if you think it is at $200/month/user, we'll see how that holds up once the cost skyrockets after OpenAI and Anthropic run out of money to burn and their investors want some returns.
The AI bubble isn't going to pop. This is like saying the internet bubble is going to pop in 1999. Maybe you will be right about short term economic trends, but the underlying technology is here to stay and will only trend in one direction: better, cheaper, faster, more available, more widely adopted, etc.
Again it's the opposite. A landscape of vibe coded micro apps is a landscape of buggy, vulnerable, points of failure. When you buy a product, software or hardware, you do more than buy the functionality you buy the assurance it will work. AI does not change this. Vibe code an app to automate your lightbulbs all you like, but nobody is going to be paying millions of dollars a year on vibe coded slop apps and apps like that is what keeps the tech industry afloat.
> Humanity is not going to stop pouring more and more money into AI.
There's no more money to pour into it. Even if you did, we're out of GPU capacity and we're running low on the power and infrastructure to run these giant data centres, and it takes decades to bring new fabs or power plants online. It is physically impossible to continue this level of growth in AI investment. Every company that's invested into AI has done so on the promise of increased improvement, but the moment that stops being true everything shifts.
> The AI bubble isn't going to pop. This is like saying the internet bubble is going to pop in 1999.
The internet bubble did pop. What happened after is an assessment of how much the tech is actually worth, and the future we have now 26 years later bears little resemblance to the hype in 1999. What makes you think this will be different?
Once the hype fades, the long-term unsuitability for large projects becomes obvious, and token costs increase by ten or one hundred times, are businesses really going to pay thousands of dollars a month on agent subscriptions to vibe code little apps here and there?
This is what everyone says when technology democratizes something that was previously reserved for a small number of experts.
When the printing press was invented, scribes complained that it would lead to a flood of poorly written, untrustworthy information. And you know what? It did. And nobody cares.
When the web was new, the news media complained about the same thing. A landscape of poorly researched error-ridden microblogs with spelling mistakes and inaccurate information. And you know what? They were right. That's exactly what the internet led to. And now that's the world we live in, and 90% of those news media companies are dead or irrelevant.
And here you are continuing the tradition of discussing a new landscape of buggy, vulnerable products. And the same thing will happen and already is happening. People don't care. When you democratize technology and you give people the ability to do something useful they never could do before without having to spend years becoming an expert, they do it en masse, and they accept the tradeoffs. This has happened time and time again.
> The internet bubble did pop... the future we have now 26 years later bears little resemblance to the hype in 1999. What makes you think this will be different?
You cut out the part where I said it only popped economically, but the technology continued to improve. And the situation we have now is even better than the hype in 1999:
They predicted video on demand over the internet. They predicted the expansion of broadband. They predicted the dominance of e-commerce. They predicted incumbents being disrupted. All of this happened. Look at the most valuable companies on earth right now.
If anything, their predictions were understated. They didn't predict mobile, or social media. They thought that people would never trust SaaS because it's insecure. They didn't predict Netflix dominating Hollywood. The internet ate MORE than they thought it would.
What part of renting your ability to do your job is "democratizing"? The current state of AI is the literal opposite. Same for local models that require thousands of dollars of GPUs to run.
Over the past 20 years software engineering has become something that just about anyone can do with little more than a shitty laptop, the time and effort, and an internet connection. How is a world where that ability is rented out to only those that can pay "democratic"?
> When the printing press was invented, scribes complained that it would lead to a flood of poorly written, untrustworthy information. And you know what? It did. And nobody cares.
A bad book is just a bad book. If a novel is $10 at the airport and it's complete garbage then I'm out $10 and a couple of hours. As you say, who cares. A bad vibe coded app and you've leaked your email inbox and bank account and you're out way more than $10. The risk profile from AI is way higher.
Same is even more true for businesses. The cost of a cyberattack or a outage is measured in the millions of dollars. It's a simple maths, the cost of the risk of compromise far oughtweights the cost of cheaper upfront software.
> You cut out the part where I said it only popped economically, but the technology continued to improve.
The improvement in AI models requires billions of dollars a year in hardware, infrastructure, end energy. Do you think that investors will continue to pour that level of investment into improving AI models for a payout that might only come ten to fifteen years down the road? Once the economic bubble pops, the models we have are the end of the road.
"Renting your ability to do your job"?
I think you're misunderstanding the definition of democratization. This has nothing to do with programmers. It has nothing to do with people's jobs. Democratizing is defined as "the process of making technology, information, or power accessible, available, or appealing to everyone, rather than just experts or elites."
In other words, democratizing is not about people who who have jobs as programmers. It's about the people who don't know how to code, who are not software engineers, who are suddenly gaining the ability to produce software.
Three years ago, you could not pay money to produce software yourself. You either had to learn and develop expertise yourself, or hire someone else. Today, any random person can sit down and build a custom to-do list app for herself, for free, almost instantly, with no experience.
> The improvement in AI models requires billions of dollars a year in hardware, infrastructure, end energy. Do you think that investors will continue to pour that level of investment into improving AI models for a payout that might only come ten to fifteen years down the road? Once the economic bubble pops, the models we have are the end of the road.
10-15 year payouts? Uhhh. Maybe you don't know any AI investors, but the payout is coming NOW. Many tens of thousands of already gotten insanely rich, three years ago, and two years ago, and last year, and this year. If you think investors won't be motivated, and there aren't people currently in line to throw their money into the ring, you're extremely uninformed about investor sentiment and returns lol.
You can predict that the music will stop. That's fair. But to say that investors are worried about long payout times is factually inaccurate. The money is coming in faster and harder than ever.
Your definition only supports my point. The transfer of skill from something you learn to something you pay to do is the exact and complete opposite of your stated definition. It turns the activity from something that requires you to learn it to one that only those that can afford to pay can do.
It is quite literally making this technology, information, and power available to only the elite.
> Uhhh. Maybe you don't know any AI investors, but the payout is coming NOW.
What payout? Zero AI companies are profitable. If you're invested in one of these companies you could be a billionaire on paper, but until it's liquid it's meaningless. There's plenty of investors who stand to make a lot of money if these big companies exit, but there's no guarantee that will happen.
The only people making money at the moment are either taking cash salaries from AI labs or speculating on Nvidia stock. Neither of which have much do with the tech itself and everything to do with the hype.
I don't know what to say to you. More people are coding now with AI than ever coded before. If your argument was true, then that would just mean that there are more elites than ever. Obviously that's not what's happening.
> What payout? Zero AI companies are profitable.
Because they're reinvesting profits into continued R&D, not because their current products are unprofitable. You're failing to understand basic high-growth business models.
> If you're invested in one of these companies you could be a billionaire on paper, but until it's liquid it's meaningless.
Plenty of AI companies have exited, and plenty of other AI companies offer tender offers where shareholders have been able to sell their shares to new investors. Again, it sounds like you just aren't really educated on what's happening. Plenty of people are millionaires in real life, not just on paper. You're massively incorrect about the payout landscape that investors are considering.
> The only people making money at the moment are either taking cash salaries from AI labs or speculating on Nvidia stock.
No, founders, early-stage investors, and employees with stock have cashed out in many cases. Again, it just feels like you're not aware of what's happening on the ground.
> Neither of which have much do with the tech itself and everything to do with the hype.
That's a very different argument. If you want to say that the investment is unsound, then fine, that's your opinion, but trying to say that investors have no appetite because they have to wait 10 to 15 years for a payout is incredibly incorrect.
I don't know how I can explain this any more clearly.
If you need AI to create software, and the cost of AI is $200/month, then only people who can afford $200/month can create software.
Costs will increase. The current cost is substituted by investor funding. Sell at a loss to get people hooked on the product and then raise the price to make money, a "high-growth business model" as you say.
The cost to make a competitor to Anthropic or OpenAI is tens or hundreds of billions of dollars upfront. There will be few competitors and minimal market pressure to reduce prices, even if the unit costs of inference are low.
$200/month is already out of reach of the majority of the population. Increases from here means only a small percentage of the richest people can afford it.
I don't know what definition of "elite" you're using but, "technology limited so that only a small percentage of the population can afford it" is... an elite group.
This is fun and all, but I think we've reached the end of the productive discussion to be had and I don't have much more to say. Charitably, we're leaving in completely different realities. I just hope when the bubble pops the fall isn't too hard for you.
1. You can build small applications with the $20/month sub, much more with the $100/month. Competition and technology improvements will inevitably improve the price to value ratio.
2. Cable sports subscriptions are in a similar price range. Expensive, but not exclusive to “the elites”.
This is the median income. If it's a struggle for someone on this income then it's worse for half of all Americans, and American incomes are higher than most of the rest of the world.
[0]: https://en.wikipedia.org/wiki/Per_capita_personal_income_in_...
Personally, I think millions more people having the ability to create some subset of software is an incredible shift.
This is an absurd claim. There are many things the majority of the population spends money on that cost more than this.
You need to take a step back and look at the economic reality of the majority of Americans today. Many live paycheck-to-paycheck, even those with "middle class" incomes. For many a $200 one-off bill is debilitating, yet alone a recurring subscription. If you don't know that, you have a dangerously narrow view of the economy.
Your entire hypothetical is based on "ifs" that aren't true. Nothing in this sentence is true. You don't need AI to create software, the cost of AI development is much less than $200/month on average, and many more people can afford AI dev than programming bootcamps or classes or degrees.
> Costs will increase. The current cost is substituted by investor funding. Sell at a loss to get people hooked on the product and then raise the price to make money, a "high-growth business model" as you say.
Inference is already profitable at current pricing. Most funding goes toward R&D for new model training, not inference.
Also, inference costs dropped over 280x between Nov 2022 and Oct 2024. Inference will continue to get cheaper as we develop more specialized hardware and efficient models.
This is not Uber, subsidizing the cost of human drivers. This is real tech, chips and servers and software. Costs fall over time, not rise. Innovation does not go backwards.
And I'm not being condescending about normal people. Developers often don't think about the possibility of making software that does a particular thing until they actually see software that does that thing. And they're going to also going to prefer to buy than vibe code unless the program is small and insignificant.
I myself have run an online community for early-stage startup founders for over a decade. The number of ambitious people who would love to build something but don't know how to code and in the last year or two have started cranking out applications is tremendous. That number is far higher than the number of software engineers who existed before.
Ok, so another fundamental proposition is monetary resources are needed to fund said technology improvement.
Whats wrong with LLMs? They require immense monetary resources.
Is that a problem for now? No because lots of private money is flowing in and Google et al have the blessing of their shareholders to pump up the amount of cash flows going into LLM based projects.
Could all this stop? Absolutely, many are already fearing the returns will not come. What happens then? No more huge technology leaps.
1. lots of room for progress, i.e. the theoretical ceiling dwarfed the current capabilities
2. strong incentives to continue development, i.e. monetary or military success
3. no obviously better competitors/alternatives
4. social/cultural tolerance from the public
Literally hasn't happened. Even if you can find 1 or 2 examples, they are dwarfed by the hundreds of counter examples. But more than likely, you won't find any examples, or you'll just find something recent where progress is ongoing.
Useful technology with room to improve almost always improves, as people find ways to make it better and cheaper. AI costs have already fallen dramatically since LLMs first burst on the scene a few years back, yet demand is higher than ever, as consumers and businesses are willing to pay top dollar for smarter and better models.
1. As I said before, we've long since reached diminishing returns on models. We simply don't have enough compute or training data left to make them dramatically better.
2. This is only true if it actually pans out, which is still an unknown question.
3. Just... not using it? It has to justify its existence. If it's not of benefit vs. the cost then why bother.
4. The public hates AI. The proliferation of "AI slop" makes people despise the technology wholesale.
2. Sure, depends on #1. But the incentive is undeniable.
3. It has. Do you think people are using Claude Code in incredible numbers for no reason?
4. The public and businesses are adopting AI en masse. It's incredibly useful. Demand is skyrocketing. I don't think you could show that negative public sentiment has been sufficient to stop this, any more than negative sentiment about TVs, headphones, bicycles, etc (which was significant).
With the exception of #1, I feel like you're arguing that things won't happen, where the numbers show they've already have happened and are accelerating.
Spaghetti code is still spaghetti code. Something that should be a small change ends up touching multiple parts of the codebase. Not only does this increase costs, it just compounds the next time you need to change this feature.
I don't see why this would be a reality that anyone wants. Why would you want an agent going in circles, burning money and eventually finding the answer, if simpler code could get it there faster and cheaper?
Maybe one day it'll change. Maybe there will be a new AI technology which shakes up the whole way we do it. But if the architecture of LLMs stays as it is, I don't see why you wouldn't want to make efficient use of the context window.
I said that (a) apps are getting simpler and smaller in scope and so their code quality matters less, and (b) AI is getting better at writing good code.
Are you trying to imply that having more things means that each of them will be smaller? There are more people than there were 500 years ago - are they smaller, or larger?
Also, the printing press did lead to much longer works. There are many continuous book series that have run for decades, with dozens of volumes and millions of words. This is a direct result of the printing press. Just as there are television shows that have run with continuous plots for thousands of hours. This is a consequence of video recording and production technologies; you couldn't do that with stage plays.
You seem to be trying to slip "smaller in scope" into your statement without backing, even though I'd insist that applications individuals wrote being "smaller in scope" was a obvious consequence of the tooling available. I can't know everything, so I have to keep the languages and techniques limited to the ones that I do know, and I can't write fast enough to make things huge. The problems I choose to tackle are based on those restrictions.
Those are the exact things that LLMs are meant to change.
The same is true of most things that get democratized. Look at video. TikTok, YouTube, YouTube shorts.
Look at all the apps people are building are building for themselves with AI. They are typically not building Microsoft Word.
Of course there will be some apps that are bigger and more ambitious than ever. I myself am currently building an app that's bigger an more ambitious than I would have tried to build without AI. I'm well aware of this use case.
But as many have pointed out, AI is worse at these than at smaller apps. And pretending that these are the only apps that matter is what's leading developers imo to over-value the importance of code quality. What's happening right now that's invisible to most professional engineers is an explosion in the number of time, bespoke personal applications being quickly built by non-developers are that are going to chip away at people's reasons to buy and use large, bloated, professional software with hundreds of thousands of users.
The apps those people were making before LLMs became ubiquitous were no apps. So by definition they are now larger and more ambitious.
We don't need more slop apps, we already have that and have for years.
But just as many are creating One More Habit Tracker or Todo App, so many that Apple had to change their review guidelines to block the surge of low-tier app slop.
Internally people are creating bespoke tools for themselves to fix issues in their daily workflows that would've either been a 100k€ software project that lasts for 6 months or required an expensive SaaS system with 420 extra features they didn't need - and the price to match.
Think about what happened to writing when we went from scribes to the printing press, and from the printing press to the web. Books and essays didn't get bigger. We just got more people writing.
> However, code quality is becoming less and less relevant in the age of AI coding, and to ignore that is to have our heads stuck in the sand. Just because we don't like it doesn't mean it's not true.
> [...]
> We are increasingly moving toward a world where people who aren't sophisticated programmers are "building" their own apps with a user base of just one person. In many cases, these apps are simple and effective and come without the bloat that larger software suites have subjected users to for years. The code is simple, and even when it's not, nobody will ever have to maintain it, so it doesn't matter. Some apps will be unreliable, some will get hacked, some will be slow and inefficient, and it won't matter. This trend will continue to grow.
But I don't see how it can be used as an argument for "code quality is becoming less and less relevant".
If AI is producing 10 times more lines that are necessary to achieve the goal, that's more resources used. With the prices of RAM and SSD skyrocketing, I don't see it as a positive for regular users. If they need to buy a new computer to run their vibecoded app, are they really reaping the benefits?
But what's more concerning to me is: where do we draw the line?
Let's say it's fine to have a garbage vibecoded app running only on its "creator" computer. Even if it gobbles gigabytes of RAM and is absolutely not secured. Good.
But then, if "code quality is becoming less and less relevant", does this also applies to public/professional apps?
In our modern societies we HAVE to use dozens of software everyday, whether we want it or not, whether we actually directly interact with them or not.
Are you okay with your power company cutting power because their vibecoded monitoring software mistakenly thought you didn't paid your bills?
Are you okay with an autonomous car driving over your kid because its vibecoded software didn't saw them?
Are you okay with cops coming to your door at 5AM because a vibecoded tool reported you as a terrorist?
Personally, I'm not.
People can produce all the trash they want on their own hardware. But I don't want my life to be ruled by software that were not given the required quality controls they must have had.
I mean, I agree, but you could say this at any point in time throughout history. An engineer from the 1960s engineer could scoff at the web and the explosion in the number of progress and the decline in efficiency of the average program.
An artist from the 1700s would scoff at the lack of training and precision of the average artist/designer from today, because the explosion in numbers has certain translated to a decline in the average quality of art.
A film producer from the 1940s would scoff at the lack of quality of the average YouTuber's videography skills. But we still have millions of YouTubers and they're racking up trillions of views.
Etc.
To me, the chief lesson is that when we democratize technology and put it in the hands of more people, the tradeoff in quality is something that society is ready to accept. Whether this is depressing (bc less quality) or empowering (bc more people) is a matter of perspective.
We're entering a world where FAR more people will be able to casually create and edit the software they want to see. It's going to be a messier world for sure. And that bothers us as engineers. But just because something bothers us doesn't mean it bothers the rest of the world.
> But then, if "code quality is becoming less and less relevant", does this also applies to public/professional apps?
No, I think these will always have a higher bar for reliability and security. But even in our pre-vibe coded era, how many massive brandname companies have had outages and hacks and shitty UIs? Our tolerance for these things is quite high.
Of course the bigger more visible and important applications will be the slowest to adopt risky tech and will have more guardrails up. That's a good thing.
But it's still just a matter of time, especially as the tools improve and get better at writing code that's less wasteful, more secure, etc. And as our skills improve, and we get better at using AI.
I'm curious about software that's actively used but nobody maintains it. If it's a personal anecdote, that's fine as well
Strongly disagree with this thesis, and in fact I'd go completely the opposite: code quality is more important than ever thanks to AI.
LLM-assisted coding is most successful in codebases with attributes strongly associated with high code quality: predictable patterns, well-named variables, use of a type system, no global mutable state, very low mutability in general, etc.
I'm using AI on a pretty shitty legacy area of a Python codebase right now (like, literally right now, Claude is running while I type this) and it's struggling for the same reason a human would struggle. What are the columns in this DataFrame? Who knows, because the dataframe is getting mutated depending on the function calls! Oh yeah and someone thought they could be "clever" and assemble function names via strings and dynamically call them to save a few lines of code, awesome! An LLM is going to struggle deciphering this disasterpiece, same as anyone.
Meanwhile for newer areas of the code with strict typing and a sensible architecture, Claude will usually just one-shot whatever I ask.
edit: I see most replies are saying basically the same thing here, which is an indicator.
That's all very true, but what you're missing is that the proportion of codebases that need this is shrinking relative to the total number of codebases. There's an incredible proliferation of very small, bespoke, simple, AI-coded apps, that are nonetheless quite useful. Most are being created by people who have never written a line of code in their life, who will do no maintenance, and who will not give two craps how the code looks, any more than the average YouTuber cares about the aperture of their lens or the average forum commenter care about the style of their prose.
We don't see these apps because we're professional software engineers working on the other stuff. But we're rapidly approaching a world where more and more software is created by non-professionals.
I agree that there will be more small, single-use utilities, but you seem to believe that this will decrease the number or importance of traditional long-lived codebases, which doesn't make sense. The fact that Jane Q. Notadeveloper can vibe code an app for tracking household chores is great, but it does not change the fact that she needs to use her operating system (a massive codebase) to open Google Chrome (a massive codebase) and go to her bank's website (a massive codebase) to transfer money to her landlord for rent (a process which involves many massive software systems interacting with each other, hopefully none of which are vibe coded).
The average YouTuber not caring about the aperture of their lens is an apt comparison: the median YouTube video has 35 views[0]. These people likely do not care about their camera or audio setup, it's true. The question is, how is that relevant to the actual professional YouTubers, MrBeast et al, who actually do care about their AV setup?
[0] https://www.intotheminds.com/blog/en/research-youtube-stats/
It takes a long time for humanity to adjust to a new technology. First, the technology needs to improve for years. Then it needs to be adopted and reach near ubiquity. And then the slower-moving parts of society need to converge and rearrange around it. For example, the web was quite ready for apps like Airbnb in the mid 90s, but the adoption+culture+infra was not.
In 5, maybe 10, certainly 15 years, I don't think as many people are going to want to learn, browse, and click through a gazillion complex websites and apps and flows when they can easily just tell their assistant to do most of it. Google already correctly realizes this as an existential threat, as do many SaaS companies.
AI assistants are already good enough to create ephemeral applications on the fly in response to certain questions. And we're in the very, very early days of people building businesses and infra meant to be consumed by LLMs.
And how do you think their assistant will interact with external systems? If I tell my AI assistant "pay my rent" or "book my flight" do you think it's going to ephemerally vibe code something on the banks' and airlines' servers to make this happen?
You're only thinking of the tip of the iceberg which is the last mile of client-facing software. 90%+ of software development is the rest of the iceberg, unseen beneath the surface.
I agree there will be more of this but again, that does not preclude the existence of more of the big backend systems existing.
People want convenience, not a way to generate an application that creates convenience.
Your example with the dataframes is completely unstructured mutation typical of a dynamic language and its sensibilities.
I know from experience that none of the modern models (even cheap ones) have issues dealing with global or near-global state and mutating it, even navigating mutexes/mutices, conds, and so on.
It actually becomes more and more relevant. AI constantly needs to reread its own code and fit it into its limited context, in order to take it as a reference for writing out new stuff. This means that every single code smell, and every instance of needless code bloat, actually becomes a grievous hazard to further progress. Arguably, you should in fact be quite obsessed about refactoring and cleaning up what the AI has come up with, even more so than if you were coding purely for humans.
Now I’m being told code quality doesn’t matter at all.
I completely agree. Just going through the beginner & hobbyist forums, the change from "can you help me with code to do X" to "I used ChatGPT/Claude/Copilot to write code to do X" happened with absolutely startling speed, and it's not slowing down. There was clearly a pent-up demand here that wasn't being met otherwise.
People are using AI to get code written. They have no idea what code quality is and only care that what they built works.
AFAICT, every time technology has allowed non-technical people to do more, it's opened up new opportunities for programmers. I don't expect this to be any different, I just want to know where the opportunities are.
But that's boring nerd shit and LLMs didn't change who thinks boring nerd shit is boring or cool.
I use LLMs in my every day work. I’m also a strong critic of LLMs and absolutely loathe the hype cycle around them.
I have done some really cool things with copilot and Claude and I keep sharing them to within my working circle because I simply don’t want to interact that much with people who aren’t grounded on the subject.
I started using Copilot at work because that's what the company policy was. It's a pretty strict environment, but it's perfectly serviceable and gets a lot of fresh, vetted updates. IDE integration with vs code was a huge plus for me.
Claude code is definitely a messier, buggier frontend for the LLM. It's clunkier to navigate and it has much more primitive context management tools. IDE integration is clunky with vs code, too.
However, if you want to take advantage of the Anthropic subscription services, I've found Claude Code is the way to go... Simply because Anthropic works hard to lock you into their ecosystem if you want the sweet discounts. I'm greedy, so I bit the bullet for all of the LLM coding stuff I do in my personal life.
I have found in general that for the type of work I do (senior to staff level engineering, 90-10 research to programming) that Claude Opus is the only model really worth my time - but I just really like the Copilot CLI tooling.
I do use LLMs to learn about new subjects but we already only bill 10% for "coding" and that's inflating it to cover other parts.
I can't imagine that slopping it up would be a great decision. Having alien code that no one ever understood between a bug report and a solution. Anthropic isn't going to give us money for our lost contracts, is it?
> I can't imagine that slopping it up would be a great decision. Having alien code that no one ever understood between a bug report and a solution. Anthropic isn't going to give us money for our lost contracts, is it?
Absolutely, that's a real concern. The only time I will let it loose on something is a throwaway project to test something, or a small tool that I know I can write deterministic tests for.
On codebases of any significant size, I'm using it more like a custom domain Stackoverflow search engine.
SWEs spend 20% of the time writing code for exactly the same reason brick-layers spend 20% of their time laying bricks
- A lot of research. Libraries documentation, best practice, sample solutions, code history,... That could be easily 60% of the time. Even when you're familiar with the project, you're always checking other parts of the codebase and your notes.
- Communication. Most projects involve a team and there's a dependency graph between your work. There may be also a project manager dictating things and support that wants your input on some cases.
- Thinking. Code is just the written version of a solution. The latter needs to exists first. So you spend a lot of time wrangling with the problem and trying to balance tradeoffs. It also involves a lot of the other points.
Coding is a breeze compared to the others. And if you have setup a good environment, it's even enjoyable.
Some people do find it unfun, saying it deprives them of the happy "flow" of banging out code. Reaching "flow" when prompting LLMs arguably requires a somewhat deeper understanding of them as a proper technical tool, as opposed to a complete black box, or worse, a crystal ball.
I kinda like how you can just use it for anything you like. I have bazillion personal projects, I can now get help with, polish up, simplify, or build UI for, and it's nice. Anything from reverse engineering, to data extraction, to playing with FPGAs, is just so much less tedious and I can focus on the fun parts.
I recently had to rewrite a part of such a prototype that had 15 years of development on it, which was a massive headache. One of the most useful things I used LLMs for was asking it to compare the rewritten functionality with the old one, and find potential differences. While I was busy refactoring and redesigning the underlying architecture, I then sometimes was pinged by the LLM to investigate a potential difference. It sometimes included false positives, but it did help me spot small details that otherwise would have taken quite a while of debugging.
I completely agree that this is the case right now, but I do wonder how long it will remain the case.
If it generates the slop version in a week but it takes me 3 more weeks to clean it up, could I have I just done it right the first time myself in 4 weeks instead? How much money have I wasted in tokens?
Soooooo....
As one who hasn't taken the plunge yet -- I'm basically retired, but have a couple of projects I might want to use AI for -- "time" is not always fungible with, or a good proxy for, either "effort" or "motivation"
> How much money have I wasted in tokens?
This, of course, may be a legitimate concern.
> If it generates the slop version in a week but it takes me 3 more weeks to clean it up, could I have I just done it right the first time myself in 4 weeks instead?
This likewise may be a legitimate concern, but sometimes the motivation for cleaning up a basically working piece of code is easier to find that the motivation for staring at a blank screen and trying to write that first function.
Cleaning up agent slop code by hand is also a miserable experience and makes me hate my job. I do it already because at $DAYJOB because my boss thinks “investing” in third worlders for pennies on the dollar and just giving them a Claude subscription will be better than investing in technical excellence and leadership. The ROI on this strategy is questionable at best, at least at my current job. Code Review by humans is still the bottleneck and delivering proper working features has not accelerated because they require much more iteration because of slop.
Would much rather spend the time making my own artisanal tradslop instead if it’s gonna take me the same amount of time anyway - at least it’s more enjoyable.
As I said, I'm retired, and so I've never had to clean up AI slop at $DAYJOB.
Since the whole AI thing would be a learning experience for me, it would include trying to toilet train the AI itself, as others have intimated can be done in some cases, rather than dealing with a bunch of already-checked-into-the-repo-slop.
And that may be a losing proposition. I don't know; haven't tried it yet.
> Would much rather spend the time making my own artisanal tradslop instead if it’s gonna take me the same amount of time anyway - at least it’s more enjoyable.
Although I haven't had the AI experience you describe, I have had a similar experience with coworkers who moved fast and broke all kinds of shit. That was similarly no fun. It's like trying to work on your wife's minivan, but she won't pull over and let you properly fix it.
Given sufficient time, I enjoy polishing/perfecting/refactoring code. My final output often looks radically different from my prototype. It is clear to me that I would hate the situation you describe. It is not clear to me that starting with prompted slop and wrangling it into submission would be much less enjoyable to me than writing my own slop and then wrangling it into submission.
> especially when I find the people I’d be giving money to so reprehensible.
This is a bit of a concern, but I'm pretty sure that, at the moment, every token you burn costs them more than you.
I’ve been sounding the alarm in my own circles about the lack of junior roles now because of AI - which will lead to a shortage of seniors in just a few years - but there is something even more sinister: juniors no longer improve enough to be intermediates and seniors, and worse…seniors and intermediates have regressed to juniors through laziness and cognitive offloading.
Like if I’m just sending code review to a middle man prompter - why not just skip the middle man? I’m already wrangling a handful of AI agents myself every day, so what is even the point of this extra person anyway? I don’t want to replace people with AI but if the person is so lazy that even I would probably prefer just doing the prompting myself then why shouldn’t I replace them with AI?
My problem, if and when I get started, would be tangential to this. It is clear that communication with LLMs is changing so rapidly that there may not be any universal long-lived lessons to be learned from optimizing your interactions with a particular model.
I know that one-shotting things is probably not best, but determining how far to take it and when to cut over and finish it myself is something that I want to learn, but perhaps not too well.
My skills are an eclectic mix of high- and low- level. I know exactly what, for example, a frequency analyzer can do for me, but controlling the $400K frequency analyzer is often best left to the guy who lives and breathes it.
Likewise, my debugging skills are exceptional, but I am not as proficient with any particular debugger as are people who live in the debugger daily because they write terrible code. My debugging skills are mostly predicated on a big part of your daily life -- reading code.
(To be fair, I have known a very few people who live in the debugger because they are dealing with intractable problems caused by other people, but those are the rarities. I, myself, used to live in the debugger a lot when I was writing graphics drivers for the mostly undocumented Windows 3.1.)
Which brings us to your reports and/or co-workers. These people have always existed. They pride themselves on and partly base their value on and derive their value from the tools they think they know inside-out.
In truth, they don't know the tools, but they are intimately familiar with the controls of the tool, like a child who knows how to make a smartphone do exactly what their parent needs it to do.
So, as long as it's a tool you need, but it's too painful for you to control directly, these people are useful. In your case, you already have cause to use the LLM directly on a regular basis, so, as you point out, the value of these people is diminishing and maybe already negative.
> why shouldn’t I replace them with AI?
You probably should. Or, at a minimum, if possible, you should restructure things so that the people who are doing things that you are already proficient at are doing them for someone else who isn't as proficient at the tools, and you can get out of that loop.
One reason I am not yet completely insane is that I realized about 40 years ago that the place I hated most being was inside someone else's debug loop. Because most people are objectively stupid, and this goes double for people who need you in that loop. So I always work to structure my responsibilities and work setup to avoid this. If I find a bug in an internal supplier's code, I create an MVCE and hand it over to them. If an internal customer claims to find a bug in my code and doesn't provide an MVCE, I figure out what they are attempting to do, create my own MVCE for their function, and either fix it if it was really my problem, or hand it back to them, and ask them to expand on it until it breaks and get back to me.
Reflecting on this, I realize that I am probably not too likely to succumb to interminable prompting loops, because that wouldn't feel much different to what I nave avoided most of my life. On the few occasions over the last four decades where being involved in someone else's debug loop was completely unavoidable, the most useful thing I brought to the table when they were out of ideas and ready to throw a lot of effort at trying random things was a series of questions like "What are you going to learn from that? What will your decision points be?"
And I'm not much of a gambler, so I won't be spending too many tokens hoping "the next time, for sure!"
The only comparison I can come up with is 3D printers, but even that's not as ridiculously fast and easy as AI coding. An average person can ask an agent to write a program, in any popular language, and it'll do it, and it'll work. We still need people intelligent enough to steer the agent, but you do not need to edit a single line of code anymore.
In both cases, you feel super productive all the time, because you are constantly putting in instructions and getting massive amounts of output, and this feels like constant & fast progress. It's scary how easy it is to waste time on LLMs while not even realizing you are wasting time.
2. Make it pretty
3. Make it fast.
If you have something that's pretty and fast but doesn't work, is that useful?
But if you have something that works, but is ugly and slow - you can still use it and fix it while you dogfood it. Maybe figure out other ways to improve it beyond making the code look pretty.
The AI’s are more than capable of producing a mountain of docs from which to rebuild, sanely. They’re really not that capable - without a lot of human pain - of making a shit codebase good.
There is something at this point kind of surreal in the fact that you know everyday there will be this exact blog post and these exact comments.
Like, its been literal years and years and yall are still talking about the thing thats supposed to do other things. What are we even doing anymore? Is this dead internet? It boggles the mind we are still at this level of discourse frankly.
Love 'em hate 'em I don't care yall need to freaking get a grip! Like for the love god read a book, paint a picture! Do something else! This blog is just a journey to snooze town and we all must at some level know that. This feels like literal brain virus.
Is there evidence these groups are a minority? I mean, the OP sounds like they are taking the right approach but I suspect it requires both skill/experience and an open mind to take their approach.
Just because an approach has good use-cases doesn't mean those are going predominate.
Being completely methodical about development really helps. obra/superpowers, for example, gets close but I think it overindexes on testing and doesn’t go far enough with design document templates, planning, code style guides, code reviews, and more.
Being methodical about it takes more time, but prevents a good bit of the tech debt.
Planning modes help, but they are similarly not methodical enough.
But, the errors that are described - no architecture adhesion, lack of comprehension, random files, etc. are a matter of not leveling up the sophistication of use further, not a gap in those tools.
As an example. Very clearly laying out your architecture principles, guidance, how code should look on disk, theory on imports, etc. And then - objectively analyzing any proposed change against those principles, converges toward sane and understandable.
We've been calling it adversarial testing across a number of dimensions - architecture, security, accessibility, among other things. Every pr gets automatically reviewed and scored based on these perspectives. If an adversary doesn't OK the PR, it doesn't get merged.
If AI needs to re-write everything from scratch everytime you make a design change, that may have some obvious inefficiencies and limitations to it but if it can also do that in a few hours or a week, is it really that bad comapred to months of stalling and excuses from devs trying to understand the work of someone before them who wasnt given enough time to make well documented clean code to begin with?
Like it is undoubtedly worse for hobby projects to rely on the AI output 100%, but im actually not so sure for commercial products. It'll be the same type of spagetti garbage everywhere. There will be patterns in its nonsense that over years people will start to get accustomed to. Youll have people specialize in cleaning up AI generated code, and itll more or less be a relatively consistent process compared to picking up random developer spaghetti
maybe this is a hot take though
Scary AF
I appreciate the balanced takes and also the notion that one can use these AI tools to build software with principled use.
However, what I am still failing to see is concrete evidence that this is all faster and cheaper than just a human learning and doing everything themself or with a small team. The cat is out of the bag, so to speak, but I think it's still correct to question these things. I am putting in a _lot_ of work to reach a principled status quo with these tools, and it is still quite unclear whether it's actually improvement versus just a side quest to wrangle tools that everyone else is abusing.
This is my experience. Tests are perhaps the most challenging part of working with AI.
What’s especially awful is any refactor of existing shit code that does not have tests to begin with, and the feature is confusing or inappropriately and unknowingly used multiple places elsewhere.
AI will write test cases that the logic works at all (fine), but the behavior esp what’s covered in an integration test is just not covered at all.
I don’t have a great answer to this yet, especially because this has been most painful to me in a React app, where I don’t know testing best practices. But I’ve been eyeing up behavior driven development paired with spec driven development (AI) as a potential answer here.
Curious if anyone has an approach or framework for generating good tests
Pull out as many pure functions as possible and exhaustively test the input and output mappings.
And finally, how do you address spec drift?
The tricky part of unit tests is coming up with creative mocks and ways to simulate various situations based on the input data, w/o touching the actual code.
For integration tests, it's massaging the test data and inputs to hit every edge case of an endpoint.
For e2e tests, it's massaging the data, finding selectors that aren't going to break every time the html is changed, and trying to winnow down to the important things to test - since exhaustive e2e tests need hours to run and are a full-time job to maintain. You want to test all the main flows, but also stuff like handling a back-end system failure - which doesn't get tested in smoke tests or normal user operations.
That's a ton of creativity for AI to handle. You pretty much have to tell it every test and how to build it.
If some component doest benefit from being extensively tested, then it's still the same today. The difference is now it's so easy to generate something, no matter how useless it is. Worse part is, no one cares. Test passes, it doesn't affect production, line coverage increases, managers think the software is more tested, developers just let a prompt do everything. It's all just testing theatre.
I think E2E is the more important than ever. AI is pretty good at getting the local behaviour correct. So unit tests are of less value. Same can't be said for the system as a whole. The best part is, AI is actually pretty good at writing E2E tests. Ofc, given that you already know what you want to test
This is a great article. I’ve been trying to see how layered AI use can bridge this gap but the current models do seem to be lacking in the ambiguous design phase. They are amazing at the local execution phase.
Part of me thinks this is a reflection of software engineering as a whole. Most people are bad at design. Everyone usually gets better with repetition and experience. However, as there is never a right answer just a spectrum of tradeoffs, it seems difficult for the current models to replicate that part of the human process.
In one of the cases, I was searching for a way to extract a bunch of code that 5-6 queries had in common. Whatever this thing was, its parameters would have to include an array/tuple of IDs, and a parameter that would alter the table being selected from, neither of which is allowed in a clickhouse parameterized view. I could write a normal view for this, but performance would’ve been atrocious given ClickHouse’s ok-but-not-great query optimizer.
I asked AI for alternatives, and to discuss the pros and cons of each. I brought up specific scenarios and asked it how it thought the code would work. I asked it to bring what it knew about SQL’s relational algebra to find the an elegant solution.
It finally suggested a template (we’re using Go) to include another sql file, where the parameter is a _named relation_. It can be a CTE or a table, but it doesn’t matter as long as it has the right columns. Aside from poor tooling that doesn’t find things like typos, it’s been a huge win, much better than the duplication. And we have lots of tests that run against the real database to catch those typos.
Maybe this kind of thing exists out there already (if it does, tell me!) but I probably wouldn’t have found it.
This could likely be extracted much easier now from the new code, but imagine API docs or a mapping of the logical ruleset with interwoven commentary - other devtools could be built easily, bug analysis could be done on the structure of rules independent of code, optimizations could be determined on an architectural level, etc.
LLMs need humans to know what to build. If generating code becomes easy, codifying a flexible context or understanding becomes the goal that amplifies what can be generated without effort.
1) All-knowing oracle which is lightly prompted and develops whole applications from requirements specification to deployable artifacts. Superficial, little to no review of the code before running and committing.
2) An additional tool next to their already established toolset to be used inside or alongside their IDE. Each line gets read and reviewed. The tool needs to defend their choices and manual rework is common for anything from improving documentation to naming things all the way to architectural changes.
Obviously anything in between as well being viable. 1) seems like a crazy dead-end to me if you are looking to build a sustainable service or a fulfilling career.
> In theory, you can try to preserve this context by keeping specs and docs up to date. But there’s a reason we didn’t do this before AI: capturing implicit design decisions exhaustively is incredibly expensive and time-consuming to write down. AI can help draft these docs, but because there’s no way to automatically verify that it accurately captured what matters, a human still has to manually audit the result. And that’s still time-consuming.
I agree that it's time consuming and we don't have a good solution yet, but my guess is that a huge part of the next 3 years of iteration in the craft of Software Engineering is going to be creating tools and practices to make this possible. Especially as AIs get better at the actual writing of the code, the key failure mode for agentic coding is going to be the intent gap between what you asked for and what you wanted.
Oof, this hit very close to home. My workplace recently got, as a special promotion, unlimited access to a coding agents with free access to all the frontier models, for a limited period of time. I find it extremely hard to end my workday when I get into the "one more prompt" mindset, easily clocking 12-hour workdays without noticing.
I now have several projects going in languages that I've never used. I have a side project in Rust, and two Go projects. I have a few decades experience with backend development in Java, Kotlin (last ten years) and occasionally python. And some limited experience with a few other languages. I know how to structurer backend projects, what to look for, what needs testing, etc.
A lot of people would insist you need to review everything the AI generates. And that's very sensible. Except AI now generates code faster than I can review it. Our ability to review is now the bottleneck. And when stuff kind of works (evidenced by manual and automated testing), what's the right point to just say it's good enough? There are no easy answers here. But you do need to think about what an acceptable level of due diligence is. Vibe coding is basically the equivalent of blindly throwing something at the wall and seeing what sticks. Agentic engineering is on the opposite side of the spectrum.
I actually emphasize a lot of quality attributes in my prompts. The importance of good design, high cohesiveness, low coupling, SOLID principles, etc. Just asking for potential refactoring with an eye on that usually yields a few good opportunities. And then all you need to do is say "sounds good, lets do it". I get a little kick out of doing variations on silly prompts like that. "Make it so" is my favorite. Once you have a good plan, it doesn't really matter what you type.
I also ask critical questions about edge cases, testing the non happy path, hardening, concurrency, latency, throughput, etc. If you don't, AIs kind of default to taking short cuts, only focus on the happy path, or hallucinate that it's all fine, etc. But this doesn't necessarily require detailed reviews to find out. You can make the AI review code and produce detailed lists of everything that is wrong or could be improved. If there's something to be found, it will find it if you prompt it right.
There's an art to this. But I suspect that that too is going to be less work. A lot of this stuff boils down to evolving guardrails to do things right that otherwise go wrong. What if AIs start doing these things right by default? I think this is just going to get better and better.
I know not everybody is quite ready for this yet. But I'm working from the point of view that I won't be manually programming much professionally anymore.
So, I now pick stuff I know AIs supposedly do well (like Go) with good solid tool and library ecosystems. I can read it well enough; it's not a hard language and I've seen plenty other languages. But I'm clearly not going to micro manage a Go code base any time soon. The first time I did this, it was an experiment. I wanted to see how far I could push the notion. I actually gave it some thought and then I realized that if I was going to do this manually I would pick what I always pick. But I just wasn't planning to do this manually and it wasn't optimal for the situation. It just wasn't a valid choice anymore.
Then I repeated the experiment again on a bigger thing and I found that I could have a high level discussion about architectural choices well enough that it did not really slow me down much. The opposite actually. I just ask critical questions. I try to make sure to stick with mainstream stuff and not get boxed into unnecessary complexity. A few decades in this industry has given me a nose for that.
My lack of familiarity with the code base is so far not proving to be any issue. Early days, I know. But I'm generating an order of magnitude more code than I'll ever be able to review already and this is only going to escalate from here on. I don't see a reason for me to slow down. To be effective, I need to engineer at a macro level. I simply can't afford to micro manage code bases anymore. That means orchestrating good guard rails, tests, specifications, etc. and making sure those cover everything I care about. Precisely because I don't want to have to open an editor and start fixing things manually.
As for Rust, that was me not thinking about my prompt too hard and it had implemented something half decent by the time I realized so I just went with it. To be clear, this one is just a side project. So, I let it go (out of curiosity) and it seems to be fine as well. Apparently, I can do Rust now too. It's actually not a bad choice objectively and so far so good. The thing is, I can change my mind and redo the whole thing from scratch and it would not be that expensive if I had to.
I didn't have to review the code for understanding what Claude did, I reviewed it for verifying that it did what it had been told.
It's also nuts to me that he had to go back in later to build in tests and validation. The second there is an input able to be processed, you bet I have tests covering it. The second a UI is being rendered, I have Playwright taking screenshots (or gtksnapshot for my linux desktop tools).
I think people who are seeing issues at the integration phase of building complex apps are having that happen because they're not keeping the limited context in mind, and preempting those issues by telling their tools exactly how to bridge those gaps themselves.
Really enjoy your style.
I agree with you in theory but in my opinion, it doesn't work so well when you don't even know what exactly you are looking for at the start. Yes I knew I wanted a formatter, linter, parse but which language should those be written in, should they be one project or many, how the pieces should fit together, none of that was clear to me.
As I pointed out in the article, in these sort of "greenfield projects" I work a lot better with concrete prototypes and code in front of me I can dissect instead of trying to endlessly play with designs in my head.
> It's also nuts to me that he had to go back in later to build in tests and validation.
I think this is a little misleading. Yes I did do some testing retroactively (i.e. the upstream validation testing) but I was using TDD + verifying outputs immediately, even during the vibe coding phase. The problem as I point out is that this is not enough. Even when I had unit tests written at the same time as they code, they had lots of holes and over time, I kept hitting SQL statements which failed which the testing did not cover.
I'd really recommend separating prototyping work like that out into a pre-design phase. Do the prototypes and figure out the direction for the actual project, but then come back in with a clean repo and design docs built off the prototypes, for claude to work from. I started out using claude to refactor my old projects (or even my codex ones) before I realized it worked better starting fresh.
I think sometimes it silently decides that certain pieces of code or design are absolute constraints, and won't actually remove or change them unless you explicitly tell it to. Usually I run into this towards the end of implementation, when I'll see something I don't expect to and have to tell it to rip it out.
One example recently was an entire messaging queue (nats jetstream) docker image definition that was sitting in the deploy files unused, but claude didn't ever mention or care about it as it worked on those files; it just silently left it sitting there.
Another example was an auth-bypass setting I built in for local testing during prototyping, being not just left alone by Claude but actually propagated into other areas of the application (e.g. API) without asking.
90 percent of the things users want either A) dont exist or B) are impossible to find, install and run without being deeply technical.
These things dont need to scale, they dont need to be well designed. They are for the most part targeted, single user, single purpose, artifacts. They are migration scripts between services, they are quick and dirty tools that make bad UI and workflows less manual and more managable.
These are the use cases I am seeing from people OUTSIDE the tech sphere adopt AI coding for. It is what "non techies" are using things like open claw for. I have people who in the past would have been told "No, I will not fix your computer" talk to me excitedly about running cron jobs.
Not everything needs to be snap on quality, the bulk of end users are going to be happy with harbor freight quality because it is better than NO tools at all.
But it does a good job of countering the narrative you often see on LinkedIn, and to some extent on HN as well, where AI is portrayed as all-capable of developing enterprise software. If you spend any time in discussions hyping AI, you will have seen plenty of confident claims that traditional coding is dead and that AI will replace it soon. Posts like this is useful because it shows a more grounded reality.
> 90 percent of the things users want either A) dont exist or B) are impossible to find, install and run without being deeply technical. These things dont need to scale, they dont need to be well designed. They are for the most part targeted, single user, single purpose, artifacts.
Yes, that is a particular niche where AI can be applied effectively. But many AI proponents go much further and argue that AI is already capable of delivering complex, production-grade systems. They say, you don't need engineers anymore. They say, you only need product owners who can write down the spec. From what I have seen, that claim does not hold up and this article supports that view.
Many users may not be interested in scalability and maintainability... But for a number of us, including the OP and myself, the real question is whether AI can handle situations where scalability, maintainability and sound design DO actually matter. The OP does a good job of understanding this.
> But when I reviewed the codebase in detail in late January, the downside was obvious: the codebase was complete spaghetti...It was extremely fragile; it solved the immediate problem but it was never going to cope with my larger vision...I decided to throw away everything and start from scratch
This part was interesting to me as it lines up with Fred Brooks "throw one away" philosophy: "In most projects, the first system built is barely usable. Hence plan to throw one away; you will, anyhow."
As indicated by the experience, AI tools provide a much faster way of getting to that initial throw-away version. That's their bread and butter for where they shine.
Expecting AI tools to go directly to production quality is a fool's errand. This is the right way to use AI - get a quick implementation, see how it works and learn from it but then refactor and be opinionated about the design. It's similar to TDD's Red, Green, Refactor: write a failing test, get the test passing ASAP without worrying about code quality, refactor to make the code better and reliable.
In time, after this hype cycle has died down, we'll come to realize that this is the best way to make use of AI tools over the long run.
> When I had energy, I could write precise, well-scoped prompts and be genuinely productive. But when I was tired, my prompts became vague, the output got worse
This part also echoes my experience - when I know well what I want, I'm able to write more specific specifications and guide along the AI output. When I'm not as clear, the output is worse and I need to spend a lot more time figuring it out or re-prompting.
Everybody remembers that soundbite but nobody remembers that he changed his mind about it later and switched to advocating iterative refinement.
If we are all honest, it seems to be the case - most of the time:
- Refactoring (Sometimes starting again.. this is rarely starting from scratch as there would have been some insights and personal design decisions garnered from the previous experience) - Specificity (It is heavily influenced by energy which also is different depending time of day or on the individual level)
At the end of the day, it takes taste + experience of the user to make anything of notable complexity(architecture) with AI.(For now and the nearest future at least).
I find reading articles as this gives me a renewed sense of agency as a technologist and my growing list of passions.
A solid thank you to Lalit Maganti for sharing and the better HN community. I found a lot to steal reuse from the material/banter.
When I ported pikchr (also from the SQLite project) to Go, I first ported lemon, then the grammar, then supporting code.
I always meant to do the same for its SQL parser, but pikchr grammar is orders of magnitude simpler.
> Unfortunately, unlike many other languages, SQLite has no formal specification describing how it should be parsed. It doesn’t expose a stable API for its parser either. In fact, quite uniquely, in its implementation it doesn’t even build a parse tree at all9! The only reasonable approach left in my opinion is to carefully extract the relevant parts of SQLite’s source code and adapt it to build the parser I wanted
Did they made a proper problem research in the first place?
To be clear when I say "formal specification", I'm not just talking about the formal grammar rules but also how those interpreted in practice. Something closer to the ECMAScript specification (https://ecma-international.org/publications-and-standards/st...).
[1] https://github.com/LalitMaganti/syntaqlite/blob/93638c68f9a0...
The problem comes from how SQLite's upstream parse.y works. Becuase it doesn't actually generate the parse tree, instead generating the bytecode directly, the intepretation of any node labelled "id" or "nm" is buried inside the source code behind many layers of functions. You can see for yourself by looking at SQLite's parse.y [2]
[1] https://github.com/LalitMaganti/syntaqlite/tree/main/syntaql... [2] https://sqlite.org/src/file?name=src/parse.y&ci=trunk
Also, nice work: this makes the world just a little nicer!
It is really good for getting up to speed with frameworks and techniques though, like they mentioned.
I have several Open Source projects and wanted to refactor them for a decade. A week ago I sat down with Google Gemini and completely refactored three of my libraries. It has been an amazing experience.
What’s a game changer for me is the feedback loop. I can quickly validate or invalidate ideas, and land at an API I would enjoy to use.
I think my vibe-coding success also has to do with the problem being not that “novel” and prior art exists.
Nevertheless still impressive.
Vision, taste and good judgment are going to be the key skills for software developers from now on.
It also reduces my hesitation to get started with something I don't know the answer well enough yet. Time 'wasted' on vibe-coding felt less painful than time 'wasted' on heads-down manual coding down a rabbit hole.
- first used copy and paste in and out of Grok
- started using CLI tools e.g. Claude and OpenCode
- move up to using 3 and sometimes 4 agents at the same time
- considered going to the agents managing agents
- have settled on having LLMs build tools that are both deterministic, usable by humans and the agent, and also faster (b/c there is less "back and forth")
Honestly, it feels a LOT like when Kubernetes came out. e.g. you stopped running containers on a box using Docker Compose plus scripts/configs etc. Instead gave a large part of the operation to an "agent" (in this case k8s) that managed all of the details you didn't need to care about anymore.
I've also realized that while the LLMs can crank out code at a very high rate, someone still needs to make sure everything is running, debug issues etc. You could set up agents to monitor what the agents do but then you still end up with someone needing to keep an eye on everything. If anything, you need MORE people b/c now you can just keep spinning up new components etc.
Also, was in a discussion with one of the best developers I've ever worked with. It came down to the following point:
"Programming is rapidly becoming a hobby. Software engineering is becoming more important than ever."
> I paid for that with a total rewrite.
With so much waste and not a single example of the "brilliant at giving you the right answer to a specific technical question"
> The takeaway for me is simple: AI is an incredible force multiplier
Seems more like a feel multiplier, rather than force.
> 500 tests, many of which I felt I could reuse
Indeed, feeling is the only saving grace for a mountain of random unreviewed tests
In my opinion, "giving me a better understanding for the architecture of the project" is reasonable technical compensation.
> Indeed, feeling is the only saving grace for a mountain of random unreviewed tests
I think I say a line or two above that this was after a review of the codebase so I did review these tests.
> giving me better understanding
Examples of that would also be nice (I don't doubt the personal feel that waste was justified)
> JOURNAL before: ...
> JOURNAL after: ... > was wrong here, learned this
Essentially ~all of the tests were found to be useful but in a more "smoke test" capacity i.e. they provided good "basic" coverage but it was clear that it was also not sufficient.
Which is why in the rewrite: 1) I built a TCL driver that run the upstream SQLite tests and verified we accepted or rejected the SQL in the same way as SQLite.
2) I wrote a test runner which checked for "idempotence" i.e. run the formatter over all the SQL from all the other types of tests then verify that the AST was identical in the input and output.
3) I also wrote a script which ran the formatter over the PerfettoSQL standard library [1], a real world SQLite-based codebase that I knew and deeply understood so I could go through each file and manually check the output.
> Examples of that would also be nice (I don't doubt the personal feel that waste was justified)
Some things learned concretely:
1) C was not going to work for the higher level parts of the project, even the formatter was not pleasant to read or write in C, the validator was much worse
2) Doing the SQLite source extraction in the same language meant that I could ship a really cool feature where the syntaqlite CLI could "generate dialect extensions" without people needing to download a separate script, run their own extraction on the SQLite source code, or worse yet, need to fork syntaqlite. This actually makes it technically possible for people in the web playground to dynamically build extensions to SQLite (though I haven't ended up plumbing that feature through yet)
3) Having a DSL [2] for extensions of SQLite (that e.g. PerfettoSQL could use) was the correct way to go rather than using YAML/JSON/XML etc becaue of how much clarity it provided and how AI took a lot of the annoyance of maintaining a DSL away.
4) I need to invest much more in testing from the start and also more testing where the correctness can be "proved" in some way (e.g. idempotence testing or SQLite upstream testing as described above)
[1] https://github.com/google/perfetto/tree/main/src/trace_proce... [2] https://docs.syntaqlite.com/v0.2.15/guides/custom-dialects/
BorgCfg had exactly the same situation.
mpvl (borgcfg original author, author of https://cuelang.org/) and others had tried to refine bcl while bcl itself is underspecified.
Eventually, the team built a drop-in replacement of bcl and specced out the language almost entirely.
The biggest lesson to me was that engineering never has any short cut.
Nowhere is this more obvious in my current projects than with CRUD interface building. It will go nuts building these elaborate labyrinths and I’m sitting there baffled, bemused, foolishly hoping that THIS time it would recognise that a single SQL query is all that’s needed. It knows how to write complex SQL if you insist, but it never wants to.
But even with those frustrations, damn it is a lot faster than writing it all myself.
Most of my questions are "in one sentence respond: long rambling context and question"
Then three weeks later you're tracing some control flow that makes no sense and nobody knows why it's structured that way. Not you, not the model. I've been treating it like code from a contractor now, review every line same as a junior dev's PR. Gets tedious but the alternative is worse.
I like this a lot. It suggests that AI use may sometimes incentivize people to get better at metacognition rather than worse. (It won't in cases where the output is good enough and you don't care.)
This precisely captures my experience with AI tools. When I understand the domain very deeply, AI feels like magic. I can tell it exactly how I want something implemented and it just appears in 30 seconds. When I don't understand something very well, however, I get easily misled by bogus design choices that I've delegated to the AI. It's so easy for me to spend 4 hours drafting some prototype in an almost dreamlike state of productive bliss, only for it to crash apart when I discover some fundamental bug in the thing I've vibecoded.
Ideally: local; offline.
Or do I have to wrestle it for 250 hours before it coughs up the dough? Last time I tried, the AI systems struggled with some of the most basic C code.
It seemed fine with Python, but then my cat can do that.
> Unfortunately, unlike many other languages,
what
> SQLite has no formal specification describing how it should be parsed.
sqlitebrowser.org is cool but it's not the sort of developer tools I'm talking about. As I clarify in the side notes, I'm looking for a formatter, linter, LSP, not an IDE.
> https://sqlite.org/syntax.html
As I replied to some other comment, I'm very aware that there is a syntax diagram but that really only tells half the story. If you actually look at those diagrams into detail, or you look into the the actual parse.y grammar (https://sqlite.org/src/file?name=src/parse.y&ci=trunk), you'll find that they're missing a lot of information which is required for you to actually interpret the SQL into an AST.
When I say "formal specification", I'm not just talking about the formal grammar rules but also how those interpreted in practice. Something closer to the ECMAScript specification (https://ecma-international.org/publications-and-standards/st...).
Unfortunately, AI seems to be divisive. I hope we will find our way back eventually. I believe the lessons from this era will reverberate for a long time and all sides stand to learn something.
As for me, I can’t help but notice there is a distinct group of developers that does not get it. I know because they are my colleagues. They are good people and not unintelligent, but they are set in their ways. I can imagine management forcing them to use AI, which at the moment is not the case, because they are such laggards. Even I sometimes want to “confront” them about their entire day wasted on something even the free ChatGPT would have handled adequately in a minute or two. It’s sad to see actually.
We are not doing important things and we ourselves are not geniuses. We know that or at least I know that. I worry for the “regular” developer, the one that is of average intellect like me. Lacking some kind of (social) moat I fear many of us will not be able to ride this one out into retirement.
I am a technologist. But I am seriously concerned about the ecological consequences of the training and usage of AI. To me, the true laggards are those, who have not understood yet, that climate change requires a prudent use of our resources.
I don't mind people having fun or being productive with AI. But I do mind it when AI is presented as the only way of doing things.
Only an AI would bother to create a throwaway account to post such a shallow comment that is mostly fearmongering to push people to use AI.
The counter here would be, what if AI could be made efficient? Suddenly OK then? Is it truly about the resources?
Walking to the nearest farm with my horse is much, much more sustainable than maintaining a sprawling toxic civilizational level infrastructure so I can go into my car to the supermarket. I get your point, but nearly every aspect of our world is filled to the brim with mind boggling complexity and corresponding resource usage and we get used to it.
"AI is an incredible force multiplier for implementation, but it’s a dangerous substitute for design."
Seconded!
Expanding a thought beyond 280 characters and publishing it somewhere other than the X outrage machine is something we should be encouraging.