Forgot your password?
Close
typodupeerror
Programming

Ruby on Rails Creator Says AI Coding Tools Still Can't Match Most Junior Programmers (youtube.com) 44

Posted by msmash from the flickering-in-the-dark dept.
AI still can't produce code as well as most junior programmers he's worked with, David Heinemeier Hansson, the creator of Ruby on Rails and co-founder of 37 Signals, said on a recent podcast [video link], which is why he continues to write most of his code by hand. Hansson compared AI's current coding capabilities to "a flickering light bulb" -- total darkness punctuated by moments of clarity before going pitch black again.

At his company, humans wrote 95% of the code for Fizzy, 37 Signals' Kanban-inspired organization product, he said. The team experimented with AI-powered features, but those ended up on the cutting room floor. "I'm not feeling that we're falling behind at 37 Signals in terms of our ability to produce, in terms of our ability to launch things or improve the products," Hansson said.

Hansson said he remains skeptical of claims that businesses can fire half their programmers and still move faster. Despite his measured skepticism, Hansson said he marvels at the scale of bets the U.S. economy is placing on AI reaching AGI. "The entire American economy right now is one big bet that that's going to happen," he said.
This discussion has been archived. No new comments can be posted.

Ruby on Rails Creator Says AI Coding Tools Still Can't Match Most Junior Programmers More

Ruby on Rails Creator Says AI Coding Tools Still Can't Match Most Junior Programmers

Comments Filter:

  • Managers vs trenches (Score:5, Insightful)

    by Shaiku ( 1045292 ) on Friday January 16, 2026 @12:17PM (#65929288)

    Those on the ground know the giant grift that AI coding is right now. I'm not sure management will ever figure it out, though, and we're all going to suffer in the meantime.

    • Re:Managers vs trenches (Score:5, Insightful)

      by kurkosdr ( 2378710 ) on Friday January 16, 2026 @12:38PM (#65929330)
      MBAs don't understand that "coding" is a term we software developers invented to make the job look less scary (and risky) to MBAs. In reality, coding is by far the easiest part of software development. Software development isn't "coding", software development is the design of systems to automate complex tasks (as much as possible), it involves a thorough understanding of the problem at hand (aka domain knowledge), complex organizing of components, and managing trade-offs.

      Unless it's some crappy website, software developers don't "code", they design.

      • software developers don't "code", they design.

        We used to call people who design software "software architects" or at least admit that their job was more than "coding."

        One software design is mostly* done, you want a coder with a good knowledge of the language/toolset (and in some cases, the underlying hardware) to turn the design into efficient, maintainable code.

        * setting aside that the architects and coders typically talk to each other through most of the process - you don't want a a software design that is theoretically beautiful but in practice a mi

      • Re: (Score:2)

        by narcc ( 412956 )

        I'd agree, but experience has shown that most developers just 'code', building increasingly absurd "architectures" to justify larger teams while delivering surprisingly little.

        An anecdote, as told to me by a member of his organizations technology steering committee: A request to add a new payment method to their website was finally rejected after seven months when their web team (six developers and a non-technical 'scrum master') determined that integration was not possible (after 6 months!) claiming that

        • Still, at least those people understand what they have built. Even if you get AI to spit out a codebase that covers your needs, it can't make changes to it (to cover a new need, for example), you can only tell it to spit out an entirely new codebase (that may or may not still cover your other needs the previous codebase did). The example that you mentioned (of developers failing to make a change) is an outlier, most software undergoes a lot of changes and additions over its lifetime.

          Also, let's not forge
      • Worth mentioning that the vast majority of programming isn't writing new fresh code (although that is the most fun). After the first month (or day), you are integrating your code into an existing product, and that is a different skillset.

      • Re: (Score:2)

        by gweihir ( 88907 )

        Indeed. One reason data models and interfaces matter so much. Code is minor in comparison. Still, if you fill the thing later with bad code, it will eventually collapse.

        My take is one reason we have all this "LLM Coder" stupidity is that a lot of code produced shortly before LLMs became a thing was really bad and LLMs can give the appearance to be able to perform on the level of a really bad coder. The MBAs do not understand that this is essentially building a house of cards and that it will come crashing d

    • Grift huh? I've been doing this for 30 years. I don't code anymore, I am using claude, and if you use it correctly, you'll see it's not a grift.

  • > "The entire American economy right now is one big bet that that's going to happen,"

    So what was "the entire American economy" before "AI" started to happen? Is that just all gone now? So the American economy is based now on nothing more than "AI"? This entire line of thinking is ridiculous. No the "entire American economy" is not based on "AI". There's far more to an economy than the latest tech buzzword.

    • I was going to comment something along these lines, but the part that bothers me is not the whole US economy being AI, but that it "is one big bet that that's going to happen".

      What does he mean? That if the market bets high enough it will happen, just because? By his logic then, if I go to a casino and go all in, I canâ(TM)t lose?

  • Of course they can't. They can't make decent decisions about how to put the blocks of code together that is cohesive, maintainable, and understandable. They WRITE code. They don't make it work. But in my hands... I can get weeks' worth of work done in days, and days' worth of work in hours. So you are missing the pont.

    • Re: (Score:2, Insightful)

      by ihadafivedigituid ( 8391795 )
      You make a good point: AI doesn't need to "match junior programmers" if it makes the junior programmers redundant because the senior programmers are way more effective and don't need their help anymore.

      This seems to be what's actually happening, judging by recent hiring patterns.

      • Re:Yeah! But -- when I use them .. (Score:5, Interesting)

        by Ceiu ( 8681133 ) on Friday January 16, 2026 @12:50PM (#65929362)

        > if it makes the junior programmers redundant
        > don't need their help anymore.

        A true senior developer never "needed" the help of a junior dev to begin with. A junior developer's role is not to be some senior's personal secretary to take mundane boilerplate tasks the senior doesn't want to do. The goal is to eventually grow them into a dev capable of taking their own complex, senior-level tasks so that more work can actually be done or the senior can move on to other projects/jobs/retire.

        If you see junior devs as some kind of burden to be replaced by AI, or whatever other automation or tooling, you're doing it wrong.

        • I agree -- totally. And I see my role as leading the adoption of these tools and demonstrating what they can do - NOT replacing junior engineers. Unfortunately, you always run into the management types who see you doing this and say, "Well, why do I need those guys anymore?" And it is up to me to say - "Because, imagine what WE COULD DO if we all were accelerated!" But you must have vision and must not have your nose to the keyboard.
        • Have you ever had a "junior engineer" hired at your company? Everywhere I've worked, they always wanted at least mid-level engineers.

          I HAVE worked with interns over the years, and that was more a matter of teaching them than anything else. They didn't make me more productive.

          • Re: (Score:1)

            by Ceiu ( 8681133 )

            lol they're not supposed to. You're supposed to make *them* more productive through training, knowledge sharing, and generally shaping them into what you want to see of the next generation.

            Mentoring takes time and effort, and no one starts as a senior.

    • In my experience the code that AI produces is way overkill too.

      For example, I recently had Copilot write, what should be, a small script to enable an option on all mailboxes, companywide. Something that should just be a quick for loop, checking for a parameter and setting it to something if it matched a specific value. Almost a one-liner, if I was writing it.

      The thing spit out no less than 300 lines of code with all kinds of edge case checks and safety guardrails and logging. Ok, fine, this is probably how

      • I thought that way until I started using Claude. It really does it right. And I really don't believe that AI is going to get much more intelligent than it is right now with the present LLM models. The advancement in tooling is still in its infancy. So we are going to see big changes as these tools become better and more ubiquitous. Claude just makes it easier and works with more context.

    • Re: (Score:2)

      by gweihir ( 88907 )

      But in my hands... I can get weeks' worth of work done in days, and days' worth of work in hours.

      That is called delusion. What you can do is far better mock-ups or low-reliability, no-security prototypes. These have some value for demos and experiments, but longer-term they cannot stand in for production code. May takes a while to notice though.

  • Whatever the creator of RoR thinks is bad, is probably the opposite.

    Yeah, vibe coding is chancy at best. But there are much better ways to use AI tools than lazily asking it to write everything--or even anything at all. LLMs are great for debugging, code reviews, and other things that don't actually generate code.

    • Re: (Score:2)

      by dfghjk ( 711126 )

      So you're claiming that you agree with a programmer you claim is an idiot? What does that make you?

      "But there are much better ways to use AI tools than lazily asking it to write everything--or even anything at all. LLMs are great for debugging, code reviews, and other things that don't actually generate code."

      Are they though? Or can you simply not tell the difference?

      • So you're claiming that you agree with a programmer you claim is an idiot? What does that make you?

        You lost me there. Pretty sure I disagree with him, and in this case I disagree in a fundamental sense because it sounds like he's using LLMs in a stupid way and claiming they don't really work.

        "Hansson said he remains skeptical of claims that businesses can fire half their programmers and still move faster. "

        The Mythical Man Month probably applies to LLMs: nine women can't have a baby in one month. Actually building new things depends on figuring out what to build more often than cranking out code, s

      • Re: (Score:2)

        by gweihir ( 88907 )

        So you're claiming that you agree with a programmer you claim is an idiot? What does that make you?

        Nicely spotted. For many people, "smart" has not ever entered the building. They typically are convinced they have it all figured out though. Dunning-Kruger Effect at work.

  • The proper use of current AI tools is.. (Score:5, Insightful)

    by MpVpRb ( 1423381 ) on Friday January 16, 2026 @12:38PM (#65929326)

    ..as assistants to experts who know how and where to use them and know how to review the results.
    Hypemongers make the pitch to investors that AI will allow companies to avoid paying people, usually by vastly overpromising results.
    Creating novel, robust, maintainable, complex systems is hard, regardless of the power of the tools used.
    Posts are appearing from expert programmers who are impressed by AI results.
    The hypemongers use this as evidence to support their claims that someone who knows nothing can create complex code using a simple text prompt and then release it without understanding how it works.
    The tech is advancing rapidly and is becoming more useful. It would be better for everyone if commentators simply reported the facts, not the wild speculation.

  • What do you use them to do is make programmers 20 or 30% more productive so you don't have to hire. The Salesforce story posted here describes that.

    Automation isn't all or nothing. The PlayStation 5 is made by about five people who feed components into machines because there's one step where it's hard to make a machine to move the components from machine a to machine b. But you still have literally five people making millions of consoles.

    Automation has basically wrecked the middle class and there's
    • You are 100% on the mark. I don't understand how people can't see this.

    • Re: (Score:2)

      by gweihir ( 88907 )

      Agreed. But from available actual evidence (i.e. not subjective testimony, which is always very unreliable), it is currently making coders slower, increases stress levels and introduces technological debt on architecture level. It may also do significant seeding of security vulnerabilities.

      The problem is that programmer productivity is really hard and basically impossible to measure. And so is quality of the code produced. This gets even harder when longer-term effects are concerned, like an architecture su

  • Hopefully future jr. programmers will be doing things that 2025 jr. programmers only wish they had time to do.

    Remember, in the early days of programmable computers, a programmer turned a math professor's instructions into machine code, which was far "simpler" (by today's standards) than modern "programming."

  • There is no reasonable expectation for this to get much better. Also, junior programmers are generally pretty bad.

    That said, yes, the future of the US economy is currently being bet on LLM-style AI being able to pretty soon being able on the level of at least a somewhat senior programmer. That bet will be lost. This effect get drowned out by the other factors currently destroying the US economy though.

  • First of all, I'm guessing he probably writes most of his stuff in Ruby. It's a shame. He could probably write much better systems in a proper language.

    Next, he apparently doesn't know how to use LLMs for coding and he insists on making nonsensical observations about tools he doesn't understand.

    LLM is the name of the programming language. As with any programming language, you can learn to code it in a few minutes. But if you want to claim proficiency, it takes 12-18 months of daily usage.

    LLM is not a replac

Slashdot Top Deals

16.5 feet in the Twilight Zone = 1 Rod Serling

Close