Please create an account to participate in the Slashdot moderation system

 



Forgot your password?
typodupeerror
×
AI

A Quarter of Startups in YC's Current Cohort Have Codebases That Are Almost Entirely AI-Generated (techcrunch.com) 53

A quarter of startups in Y Combinator's Winter 2025 batch have 95% of their codebases generated by AI, YC managing partner Jared Friedman said. "Every one of these people is highly technical, completely capable of building their own products from scratch. A year ago, they would have built their product from scratch -- but now 95% of it is built by an AI," Friedman said.

YC CEO Garry Tan warned that AI-generated code may face challenges at scale and developers need classical coding skills to sustain products. He predicted: "This isn't a fad. This is the dominant way to code."

A Quarter of Startups in YC's Current Cohort Have Codebases That Are Almost Entirely AI-Generated

Comments Filter:
  • Are they... (Score:4, Funny)

    by Barny ( 103770 ) on Thursday March 06, 2025 @10:11AM (#65214595) Journal

    Offering bug bounties?

  • New white color tech jobs are going to be gone in 10 years.

    • No, not gone. They'll be different, and probably reduced in number.

    • by tiqui ( 1024021 ) on Thursday March 06, 2025 @10:50AM (#65214719)

      It's becoming more and more common to see people spelling out words and phrases in ways that advertise that they have not READ them but have heard (and misunderstood) them. People who are well read do not make these errors, but people who have spent most of their lives absorbing info [poorly] by watching TV shows and movies do.

      It's NOT "white color", the proper term is "white collar" and it's a reference to the color of the fabric in a the shirt a man wears when he goes to work. In decades past, average guys who had no college degree [or perhaps simply preferred to work with their hands] and worked in factories, maintenance shops, the trades, etc wore blue shirts (and hence these jobs were referred to as "blue collar" jobs). The college educated men who worked in offices, or places like engineering areas wore traditional suits which generally included coats, ties, and white shirts [which would easily by ruined by dirty work]. These jobs were referred to as "white collar" jobs - again, a reference to the color of the fabric of the shirt worn (specifically, the fabric visible around the neck even if a person is wearing a shop apron or a coat).

      This error is similar to the error made by persons who say somebody "could care less" when trying to indicate some person has no regard for something or someone. The proper expression is "couldn't care less". These sorts of errors are becoming more common and the folks making them do not realize that they are indicating to those around them that they either do not read, or they lack proper reading comprehension skills. You may not think this matters, but you never know when the person interviewing you for your next job may take note...

      • There's something known as a common vernacular. If you want to communicate with other people, you use a common vernacular.

        I couldn't care less if you disagree with my use of the expression "white collar."

        • Your vernacular makes you sound racist, as if you believed those particular jobs were only for people with a particular skin color.

          You don't just appear uneducated, but racist too.

          • I think if I asked 100 people a question about white collar vs. blue collar jobs, maybe 1 (or less) would suggest that I was racist.

            Please just replace "white collar" with "college educated men who worked in offices or from home that traditionally wear hoodies" and return the relevant conversation.

            • Sigh, only in America the colour of a collar can be considered racist.
            • Yes, but if you spelled it "white color" as you originally posted, that figure would probably jump from 1 to 25-35.

              Spelling matters.

              • My dude, it was a typo.

          • This whole, "Ewww yer RACIST!" argument is getting old AF. It's almost as if those that claim such are actually just idiots.

        • "other people" is a broad term, my friend. Let's say that "If you want to communicate with educated people..." or let's say "If you want to communicate with uneducated people..."

          • by HiThere ( 15173 )

            Doesn't matter. "I could care less" is understood by everyone to mean a *claim* that the speaker couldn't care less. And also that they consider the context too unimportant for formal speech. It has nothing to do with literacy (though the "color"/"collar" could, but probably just indicates that they haven't run across the expressing in literary works, but only in speech. From a quick perusal of my memory I haven't run across it in text in the last couple of decades...which may tell you about my literar

        • by AvitarX ( 172628 )

          The comment is on your missing "l" in collar.

          You wrote "white color" and the. Someone replied thinking you meant literally white color rather than dealing with auto correct or a typo.

          Unless you meant to type "white color"

        • Go back and reread your post. Maybe you didn't notice, but it says "white color" and not "white collar". I'd be more likely to attribute this to an autocorrect error from typing on a phone keyboard myself, but for all intensive purposes getting into an internet slap fight doesn't do much to convince me that the person who responded to you was far off the mark. But what can I say, it's a doggy dog world and some people are always looking for an escape goat.
      • And sometimes people say "I'm so hungry I could eat a horse," but then if you presented them with a horse they wouldn't eat it! Thank you for fighting the good fight my dear autistic gentleman.

        • by taustin ( 171655 )

          My mom bought horse meat from the butcher shop when my parents were in Iceland.

          Some people will, in fact, eat a horse.

  • I haven't tried many AIs but many people have said that they can't do complete code for anything more complex than a basic standalone app. How is it that so much was done with AI then? Don't understand the gap here.
    • by mysidia ( 191772 )

      I believe you prompt AIs the same way a developer organizes tasks mentally. Piece by piece,
      and that's how you can use AI code. You still need a human builder, and the AI is not a true intelligence behind the design of your software.
      AI cannot tell you what a good design is.

      It's not that you can tell the AI "Write me a program that does (complicated list of things)" and get a viable result..

      You tell the AI "create a function for this file A that takes X input and does Y"

      Then you the human have to sort

      • So they are promoting thousands of times and linking thousands of bids of code together in a way that works? That sounds painful.
        • by mysidia ( 191772 )

          It's not that painful. You go into the Copilot editor and start typing a function, then it suggests code for you.

          The code it suggests considers the context of where the code is being added and the existing content of your program.

          • And 80% of the time that code is horribly wrong when working on anything more sophisticated that a college homework assignment.

        • Not only painful but not much different than how normal development works without AI.
          It might speed things up slightly if it doesn't introduce hidden bugs but the human is still the architect
          who is breaking it down into small parts and making those small parts work together and work
          with external 3rd party APIs and libraries.

      • How do I try this? Say I need to make a dart class that extends from another dart class, implements the parent methods, and then makes a few async library streams that other classes can connect to. Wouldn't something that simple take a long time? Is there any way to try it for free?
        • by Ksevio ( 865461 )

          You can just create a free account on https://claude.ai/ [claude.ai] and ask it to do those things. Might need to upload the parent class and define connections better but it should have some results in under a minute.

          Sometimes it's better to do it in pieces so you can have a function that takes x inputs and produces y outputs, or for classes have it implement the base class and then add functionalities rather than doing the whole thing at once. Pretty much the same as you would instruct a human junior developer

      • He didn't really build a teleporter, he just put together the pieces.

  • so... (Score:5, Insightful)

    by tiqui ( 1024021 ) on Thursday March 06, 2025 @10:34AM (#65214643)

    A quarter of startups in Y Combinator's Winter 2025 batch are a scam. These startups are staffed by lazy people whose ideas are so simple an AI chatbot can code them, yet the staff has no idea how they work and could not fix/maintain them to save their own lives. There's also, arguably, no IP at these companies, given that their product was not the creation of a human mind. Bad investment; any other group of jerks can come along and ask another AI to code an equivalent thing and suddenly have a product that's the same or better.

    Caveat emptor.

    • A quarter of startups in Y Combinator's Winter 2025 batch are a scam. These startups are staffed by lazy people whose ideas are so simple an AI chatbot can code them, yet the staff has no idea how they work and could not fix/maintain them to save their own lives. There's also, arguably, no IP at these companies, given that their product was not the creation of a human mind. Bad investment; any other group of jerks can come along and ask another AI to code an equivalent thing and suddenly have a product that's the same or better.

      Caveat emptor.

      Or YC has been targeted by a somewhat resourceful grifter that has a middling grasp of coding, and got ahold of an AI they spam all day long with ideas, each one being a separate startup in desperate need of VC cash. Most startups fail, so it won't even be all that surprising when most of these fail, and nobody will question where the money went when someone says, "Well, not ever idea is a winner."

  • by dskoll ( 99328 ) on Thursday March 06, 2025 @10:36AM (#65214651) Homepage

    A more interesting number will be how many are successful 5 years from now.

    • by gweihir ( 88907 )

      As that number might well be zero, it would not make for good "news".

    • by rsilvergun ( 571051 ) on Thursday March 06, 2025 @11:23AM (#65214839)
      The question isn't how many are successful because most startups fail. The question is how much of the ideas and technology they prototyped using AI go on to become useful products.

      Think of it like a 3D printer. Drastically changes and reduces the cost of and the man hours for prototyping and getting things off the ground.

      If we still had a functional capitalist system with competition this would be a great thing but we pissed that away years ago while we focused on moral panics. So the complete lack of antitrust law enforcement means it's not going to help increase competition
      • by dskoll ( 99328 )

        Well sure, but comparing those startups that used mostly AI with those that didn't would still give us some useful data because it is comparing apples to apples.

        • I don't think it would. I think there's too much noise in that data to really get anywhere. It makes a direct comparison basically impossible. e.g. there is no way to compare apples to apples because you're not talking apples here is too much difference between individual startups.

          Because of that you've got to compare results from successful startups because then you've got a better baseline to work off of, which is the say success. You basically take all the variables that could create a successful or
  • by Petersko ( 564140 ) on Thursday March 06, 2025 @10:37AM (#65214657)

    Remember back in the 90s when for a brief period anybody who could grasp XML and build a rudimentary web page could find work as a web designer? While coding seems somehow different, it just has a longer arc. Commodity programmers are on the block. If your output is mostly CRUD... buckle up. Add on top of that the growing reluctance to hire new grads because of generational attitude differences, and the river has run dry for the entry level.

    I saw the writing on the wall and I planned my escape. But lots of people are getting caught flat footed. It's an inconvenience to change your career at 21. It's a catastrophe if you're 50 with dependents.

  • by Xylantiel ( 177496 ) on Thursday March 06, 2025 @11:04AM (#65214761)
    What we call "AI" now seems to be largely a copyright washing-machine. It takes code or other work licensed in a particular way, washes it of its license and context, and re-creates it. Of course that is cheaper than actually respecting licenses. Don't want to use a particular library because of its license? Just ask a code agent for a similar thing and it will just regurgitate the library with its license removed. Or just skip the first question altogether because the entity running the LLM wants you dependent on them as the middle-man instead of having the original code. Tada, 95% LLM-generated code that can't ever be reused.
  • by gillbates ( 106458 ) on Thursday March 06, 2025 @11:19AM (#65214821) Homepage Journal

    Today AI works in much the same way a compiler translates high level language into assembly language (or its object-code equivalent).

    The only difference is that AI translates English into python, rather than C/C++ into assembly. You still have to be smarter than what you're working with in order to use it, and unlike compilers, the code produced by AI is very frequently buggy.

    More than two decades ago, I wrote a macro expander for my assembly language course which would take the psuedocode provided in the assignment and translate it into assembly language. Using a computer to avoid work has been the mark of a professional programmer for more than half a century now, and I doubt that programmers of all people will stop doing it "to save our jobs." There will still be a need for programmers in the future, but the positions will be fewer and harder to find.

  • And a dangerous one. Most of that stuff will have to be thrown away.

    • You misunderstand what a fad is. NFC's were a fad. Pet rocks were a fad.

      LLM's are useful, to good programmers, and terrible ones, and even non coders. There will be plenty of stupid uses, but they provide too much general utility, with nearly no learning curve, to go away, until they are replaced by something "better". Which in this case means more utility, with less user learning curve.
  • I wonder if we could predict future success of this approach by trying to understand if previous startups used StackOverflow very heavily with great success.
  • by quantaman ( 517394 ) on Thursday March 06, 2025 @12:04PM (#65214963)

    ChatGPT / CoPilot are decent at generating snippets and speeding up development, but there's no way in hell they're generating 95% of a complex code base.

    Do I suck that much at prompts? Are the dedicated LLM coding platforms getting that advanced? Or are they fudging how much of their code AI built so they sound more efficient and scalable?

  • Very glad to see that the enshitification of nearly everything is automated now.
  • They're not innovating, they're just copying. They're effectively doing nothing more complex than installing wordpress.

SCCS, the source motel! Programs check in and never check out! -- Ken Thompson

Working...