Cursor CEO Warns Vibe Coding Builds 'Shaky Foundations' That Eventually Crumble (fortune.com) 54
Michael Truell, the 25-year-old CEO and cofounder of Cursor, is drawing a sharp distinction between careful AI-assisted development and the more hands-off approach commonly known as "vibe coding." Speaking at a conference, Truell described vibe coding as a method where users "close your eyes and you don't look at the code at all and you just ask the AI to go build the thing for you." He compared it to constructing a house by putting up four walls and a roof without understanding the underlying wiring or floorboards. The approach might work for quickly mocking up a game or website, but more advanced projects face real risks.
"If you close your eyes and you don't look at the code and you have AIs build things with shaky foundations as you add another floor, and another floor, and another floor, and another floor, things start to kind of crumble," Truell said. Truell and three fellow MIT graduates created Cursor in 2022. The tool embeds AI directly into the integrated development environment and uses the context of existing code to predict the next line, generate functions, and debug errors. The difference, as Truell frames it, is that programmers stay engaged with what's happening under the hood rather than flying blind.
"If you close your eyes and you don't look at the code and you have AIs build things with shaky foundations as you add another floor, and another floor, and another floor, and another floor, things start to kind of crumble," Truell said. Truell and three fellow MIT graduates created Cursor in 2022. The tool embeds AI directly into the integrated development environment and uses the context of existing code to predict the next line, generate functions, and debug errors. The difference, as Truell frames it, is that programmers stay engaged with what's happening under the hood rather than flying blind.
Well that's unexpected... (Score:4, Insightful)
Are they insulating themselves from liability?
Re:Well that's unexpected... (Score:4, Insightful)
Either that, his company just happens to sell a code management & analysis tool that allegedly reduces the chance of AI code-slop gumming up an app. The fire insurance sales-person likes to remind customers of how hot and dry the weather feels.
Re:Well that's unexpected... (Score:4, Interesting)
"AI companies" were never the people to hype "vibe coding" as a serious software engineering approach. The people doing that were bloggers, Youtubers, and other people not very knowledgeable about software engineering making money off the new hip trend.
Remember that the term "vibe coding" was thought up by Karpathy: very smart guy, but not primarily a software engineer by his own admission. And even then he never claimed more than that vibe coding was great for hobbyists working on small projects.
7 year Gartner hype cycle (Score:2)
The people who have to parrot, promote and endlessly talk about the latest shiny and great technology need a new thing to focus one every 3 years to make money.
Look at the Gartner Group's 7 year technology hype cycle year over year for the very low success rate of the most hyped technologies.
One bright point about the hype shift to AI is that the endless promotion of "low code" products and robotic process automation products has nearly stopped. Too many bad products with little hope of long term viability
Re:Well that's unexpected... (Score:5, Interesting)
Are they insulating themselves from liability?
That's an angle I don't see mentioned very often when it comes to AI in the work world. The AI companies don't want to assume liability for anything their creation does (like giving away all the snacks in a vending machine for free, or deleting a bunch of corporate data in a fever dream), but then they try and sell AI as a replacement for humans doing many jobs.
Would a company hire a human who preemptively refuses to be held responsibility for their actions on the job?
Re: (Score:2)
Re: (Score:2)
Are they insulating themselves from liability?
That sounds quite plausible to me. Maybe they finally got some competent legal advice. They sure did not have that before.
Re: (Score:3, Funny)
Re: (Score:1)
Stop hallucinating, nobody believes you post on slashdot and have a "girlfriend". Even chatgpt-zero looks down on you.
I refuse to use AI coding tools... (Score:2)
I don't need AI coding tools to write code for me, I am perfectly capable of doing that myself. Plus there isn't an AI coding tool on the planet that can do the stuff I do with reverse engineering proprietary file formats, interacting with obscure dead game engines and working with proprietary secret code that no AI would have ever seen before.
Re:I refuse to use AI coding tools... (Score:5, Insightful)
I don't need AI coding tools to write code for me, I am perfectly capable of doing that myself.
Yes, but you can't do it as fast as you can using AI coding tools, according to your boss. He doesn't have any actual numbers to back up this assertion, and he wouldn't be looking at the impact on time in the QA/testing phase either. But he's sure he's right after talking to that guy he played golf with last month, and is willing to stake your job on it.
Re: (Score:2)
They save a lot of time on boilerplate / web design. That said, Gemini 3 (the current leaderboard champ) tried to talk me into doing something the other day that a junior dev might not have been skeptical enough of to catch, and would have ended with catastrophic downtime due to common exploits.
Re: (Score:3)
I am not surprised. The current hype-AI is incapable of writing reliably secure code.
Re: (Score:3)
They save a lot of time on boilerplate / web design.
How much new boilerplate do you need in a project which you can't pull from the git repo of your older projects?
I don't see any non-trivial savings that can justify even the $20/mo for the cheapest "copilot" account.
Re: (Score:2)
I asked Claude to improve the design on a Vue component and it did. It didn't do any magic, it just looked up what CSS library we were using and it used it as intended (accent colors, primary colors, etc). and threw in some icons.
Could I have done it? Sure. But as a "mostly backend" developer, it would have taken me a lot of time to read the docs and examples. And another good chunk of time to choose the right icons.
Re: (Score:2)
So, its not like AI is a "shortcut", instead what it is, is an exploit generation system, loose code system, and a cpu time waster.
Re: (Score:2)
Definitely don't use LLM-generated code without scrutiny, but it's not bad as a starting point or as a source for potential approaches. It's also not bad for code review: I asked gpt-oss to adapt some (not yet tested) code in a certain way, and it noticed a cut-and-paste error in addition to adapting the code. I ended up not using that adaptation, but the big report was helpful.
One can -- and I think should -- be skeptical of lots of things about LLMs, from business models and environmental impacts to qua
Re: I refuse to use AI coding tools... (Score:3)
Re: (Score:2)
That's absolutist in a rather silly way. A number of statistical "machine learning" techniques are correct by construction -- if you ask the right question. Kernel methods like support vector machines work very well. If they exhibit a "wrong" answer then it means the original problem included an incorrect assumption or omitted important data -- and any technique is liable to give a wrong answer in those cases.
Re: (Score:2)
Dont use any output from any machine learning model without checking: It is statistical models, which can do predictions better than random, but often are completely wrong. Always verify.
That won't scale. Our world runs on plenty of systems that use statistical models to automate decision making. We prioritize how often we verify these systems. We already can't verify every single decision made by statistical models, but when it is something as benign as your Google search results we are fine with only verifying a small sample to do quality control. The same will be true of most code written by AI coding tools in the near future.
Re: (Score:2)
That won't scale.
You know what won't scale? Capitalism. We waste too much time, energy, and effort duplicating work because for legal and profit reasons it has to be done over here. That's why we're now having this LLM spasm. They're out of ways to squeeze another nickel out of every dollar.
Re: (Score:3)
That won't scale.
You know what won't scale? Capitalism. We waste too much time, energy, and effort duplicating work because for legal and profit reasons it has to be done over here. That's why we're now having this LLM spasm. They're out of ways to squeeze another nickel out of every dollar.
Capitalism is very wasteful, but so far the waste of capitalism has always outperformed more centrally controlled economies. That said, I agree it likely won't scale forever. The size and power of large consolidated global companies, the easy movement of wealth by extremely wealthy families, and how easy it is for the wealthy to wield populism as a weapon to maintain political power have arguably already pushed capitalism to its limits.
This LLM "spasm" is capitalism at its best. Massive funding into hundred
Re: (Score:2)
I don't need AI coding tools to write code for me, I am perfectly capable of doing that myself.
I said that I didn't need OpenGL or DirectX in the 90s because I was quite happy fine tuning my own graphics libraries with C and x86 assembly language. I also didn't want to learn early game engines for similar reasons. But I obviously couldn't keep up with progress by using the last generation's tools. I assume the same happened for early software developers when early compilers first started gaining popularity.
I don't think AI is going to replace software developers, but I doubt there will be many develo
Re: (Score:2)
You have a place for a co worker? That really sounds interesting!
Re: (Score:3)
Actually, I'll be AI *could* help you reverse-engineer the proprietary file formats and obscure/dead games. Of all things AI is good for, it's *best* at deciphering patterns.
Re: (Score:3)
I take the exact opposite view. I want to know what this new tools is all about. I want to know what it can do and what it can't. I want to know the difference between the hype and the reality. Unless you use it with a researcher's eye, you'll never know.
Because of my extensive study of AI tools, I'm able to speak to the crazy concerns and fears of people who don't understand what it is or exactly what kind of good or bad things could happen.
Knowledge is power.
Did you see that Microsoft? (Score:2)
Sure your coders will be really engaged with 1 million lines of Rust code per month per dev.
Do comments count as code? (Score:2)
Just asking for a friend who's on that Microsoft team.
How will AI replace developers? (Score:4, Insightful)
Did all those companies firing staff due to AI knowledge this? We're they told by any shiny new AI company about these any many other limitations?
Sounds like an "oopsie" moment for one of those companies that laid off 20%+ of their developers because AI was going to replace them all.
Re: (Score:3)
That means until we have AGI. For which we do not even know whether it is possible and, if possible, we will not get anytime soon. May be 100 years, 1000 years or "never".
I really do not understand why so many people are willing to believe AI is all-powerful. There must be a widespread mental defect somewhere that can explain this disconnected belief.
Re: (Score:2)
I really do not understand why so many people are willing to believe AI is all-powerful. There must be a widespread mental defect somewhere that can explain this disconnected belief.
"Our program who art in memory, Hello be thy name..."
Re: (Score:3)
Someone cleverer than me speculated that most people hear about "AI" and immediately think of the AI that they have seen in movies. Those AIs, being fictional, are basically omniscient and (with a few notable exceptions) flawless. Those people then assume that real AI is just as capable as the fictional AI.
They will continue believing so until their interaction with AI goes off the rails in a way that affects them personally.
One would think that the regular news stories about lawyers getting spanked for cit
Re: (Score:2)
Makes a lot of sense to me. Of course, it would be a complete intellectual fail on the side of those practicing this approach. No fact-checking used whatsoever. But that would not be a surprise.
cursor ai (Score:1)
it cost 20 or so amonth and allows me to gereate a framework and do simple crud(kinda). it does save me a lot of time.
Now i have yet to see it produce any code that actaully worked with out fixes.
Duh (Score:4, Interesting)
That is really all that is to say here. Well, I would like to see some research into why so many people massively overestimate what LLM-type AI can do. I mean, I took one look and was not impressed. Why do people seem to think that AI needs to be regarded as all-knowing, all-powerful until the converse is proven? And sometimes not even with that proof? I really do not get it.
Re: Duh (Score:2)
Re: (Score:3)
That was tried and has failed several times now. One problem: Creating specs is much _harder_ than writing code. Another one is that code synthesized from specs is not very good.
That approach is not going to work without AGI. Which we do not have and may never have.
Re: Duh (Score:3)
If you're inexperienced, you can't see this this is a trap. If you are under a certain age you're not even capable of hard boiled analysis because you're trained to accept that "there's an app for that
Re: (Score:2)
Hmm. Makes sense to me. I have never understood gambling either. The only thing I ever found in it is "you do it for the feeling of expectation", but even that wears off on me after a single experiment. I do understand that some people ("the victim personality") are strongly affected by this and that many people get this effect to a weaker degree. I essentially do not. Probably why hypes do not work on me.
I can intellectually follow the argument and reasoning chain and I can see from observation how this wo
Re: Duh (Score:2)
Because from the point of view of a layman, it's all WOWs and OMGs.
But if you're an experienced professional, all you see is a steaming pile of shit. All the issues you see daily in junior devs' MRs coming back at uou at once and at scale.
Re: (Score:2)
But if you're an experienced professional, all you see is a steaming pile of shit. All the issues you see daily in junior devs' MRs coming back at uou at once and at scale.
Indeed. But without the learning that some of the juniors have and without the possibility to fire the other ones.
Re: (Score:2)
https://www.theguardian.com/commentisfree/2025/dec/26/ai-dark-ages-enlightenment
excerpts
The German philosopher Immanuel Kant famously defined the Enlightenment as "man's emergence from his self-imposed immaturity." Immaturity, he wrote, "is the inability to use one's understanding wi
Re: (Score:2)
Nice find! Makes a ton of sense, given what can be observed in the world.
But again, I guess I tick differently than the mainstream. Mental laziness is not something I do, I find it boring and stressful. When my mind works, I relax. I have been accused of "being unable to relax", which is very much not true, but given that most people seem to relax by switching their mind off, the accusation makes some sense. But for me, my mind just comfortably runs and hands me insights, that is not something that takes ef
DOGE rewrote the entire SSA Codebase with AI (Score:5, Informative)
Anybody remeber this?
DOGE To Rewrite SSA Codebase In 'Months'
[posted on slashdot 2025-03-29]
https://developers.slashdot.or... [slashdot.org]
They must have finished it by now. So how did it go? Thought so.
Re: (Score:2)
Hard to tell because the SSA went online, and attempts to use the phone you end up with an AI nightmare. Couple that with people who get SS benefits and they're just trapped in an endless loop trying to figure out what's happening with their benefits.
Which is a way to cut down use of SS I guess. The people are just unable to use a computer to file the paperwork to get benefits, they can't talk to anyone at a now0closed office nor get anyone on the phone.
Re: unfortunate new business opportunity (Score:1)
Re: (Score:2)
Yep. That whole statement was excessively stupid and disconnected back when it was made. Some people could see it immediately. Now we have proof. Of course, the usual idiots will find tons of stupid excuses ...
Critical domain thinking (Score:3)
Golden rule of AI:
You should never use AI in the domain you yourself have no knowledge or experience in. You should only use it in areas where you have enough expertise to validate its output.
Just because you use AI won't make you a surgeon or a software engineer.
Re: (Score:3)
Indeed. But that means people using AI competently have to actually have more and better skills than people doing it directly. And in some cases (software) we already know it makes them slower on top of that.
That does not quite sounds like the grand promises the AI pushers have made.
have fun in the unemployment line (Score:2)
Grumpy old people, stay off my lawn, I walked to school in the snow here at Slashdot these days.
Embrace tech or get left behind.
AI alone can't sustain code long term (YET), that doesn't mean it won't be used to get a project off the ground, get funding and then hire a couple sr people who embrace AI.
Think there is a bubble, you don't understand the market. It will pivot before it ever pops. If it did pop that means firing humans and really replacing them with cheaper labor (India, etc.) and still leverage A
AI is like spray paint (Score:2)
Coding with AI is like using a spray can to paint a canvas. It covers a large area with amazing speed but isn't great for fine detail work. Sometimes you have to get in there with a small brush and do the detail work yourself. The trick is to know when to use which tool.