I'm done using AI
35 points by untrusem
35 points by untrusem
Upon reflection, I've come to believe that quality software is not about how many features something has or how quickly it is built but rather how cared for it is by the designers, developers, and product people who work on it.
Fantastic quote. I feel we are all rushing to add more and more bloat to everything faster, rather than treating it with the care and attention it deserves. I hope that people who think like this will have some impact reversing the trend.
I empathize with this a lot. There have been times over the last 6 months where I've been almost brought to tears because of the sudden shift in expectations for my career. It's maddening. But at the same time, the business and market do not care about my mental health and in the majority of cases do not care about who made a product. Hell, a lot of the time the quality of the code itself is inconsequential to the business, especially in startups (where I happen to be employed). I was hugely skeptical of AI coding even 8 months ago and now I'm incredibly fearful for my place in the future of this industry. I think the dread of being left behind is founded. I wish this genie could be put back in the bottle.
While I agree that care is important, I can’t help wondering why we can’t have both care and velocity with AI?
Empirical evidence suggests we cannot because it is not valued by the people who sign the checks. Just look at the legal system for more evidence, we cannot get corporations to stop literally killing people because the cost of lawsuits is less than their profits: "price of doing business".
Empirical evidence suggests we cannot
Your comment here sort of hand-waves around who "we" refers to in this context and presumes we're discussing in a context where people are signing checks. The conversation is presumably more general than your narrow framing here and your specific complaints aren't universal to the subject.
Yeah? You're surrounded by people casually paying for AI and mandating you use it in your personal life? Idk about you, but the only place I've heard "velocity" used is in some agile routine at a corporation.
I don't know how old you are, but in my experience, as you progress in your software career, you end up helping people get better, sometimes very mediocre people, review their code, sometimes very bad code, plan what to do next, essentially do what you would do when using something like Claude Code.
I don't think LLMs are going to go away, but you can learn to use them, without relying on them. For example, I find I extremely valuable to have AI review my code. As I get older, my attention span os shorter and I'm making more mistakes. Having a mechanical pair of eyes available is helpful. I never got such good reviews from my co-workers.
And obviously, if you like coding, do the coding yourself. Don't let machines take it away from you.
I prefer computer science theory and software architecture, so I'm kind of glad that I can delegate the coding part, but still not completely ignore it. Just old school XP pair programming.
you end up helping people get better, sometimes very mediocre people, review their code, sometimes very bad code, plan what to do next, essentially do what you would do when using something like Claude Code
I think these are only superficially similar. You're doing some similar things like reviewing code you didn't write. But you're not actually working with and helping another person and watching them grow, and that's a big part of what makes it rewarding to me. Not everyone is going to learn and grow and become a superstar, but some number of people will, and you'll probably value those connections for your entire career.
Yes, you are right, but I've mentored and guided more people in my life, than people whom I had connection with. While working in a company, my primary goal was to ship good software, and in my case that meant more planning and mentoring, less coding. There obviously were a few people, who I wanted to see grow. I'm not saying that AI should replace coworkers. But the skills needed and frustrations experienced are similar.
This matches my experience working with agentic coding (which is just a subset of using AI for coding). I treat the agent like I would any subordinate developer and I use the skills I honed in my career as a manager of software developers. The code review cycle feels very familiar to me and honestly I'm getting better results from the frontier models these days than many (most?) of the software developers I managed when I was doing this professionally.
I don't think that's right at all. If you tell the AI that it made a mistake during a code review you're shouting into the void. Unless it will later pick up exactly what you said in this review for some reason none of the future Codexes or Claudes will remember having this conversation with you. You're not mentoring anyone. At best you are building a corporate knowledgebase but generally people are just giving advice to something with the lifespan of a fruit fly.
I hear what you're saying, but it doesn't really match the actual working reality of the current state of agentic coding which absolutely retains its own local memory (OOB, not in-repo) of steering and guidance received from the operator for use across sessions.
Beyond that, part of the technique of working with AI coding agents is to direct your steering and guidance back to in-repo documentation and AGENTS.md/CLAUDE.md style resources which serve as the exact memory you claim can't exist.
Just as in-repo documentation needs to be written to assist future human users and developers, the same documentation burden exists for guiding and steering future AI agents who will be interacting with the repo.
helping people get better, sometimes very mediocre people, review their code, sometimes very bad code, plan what to do next, essentially do what you would do when using something like Claude Code
The similarity is superficial, primarily because the person you are mentoring can learn and grow. Claude can not. Reviewing LLM generated code ultimately does nothing for the future, while reviewing a junior developer's code eventually leads to a new senior developer.
I disagree. We had a boom of software development. I've seen streams of junior developers who did not want to learn, they wanted to get the job done with smallest amount of effort and go home. I respect that attitude, but I have zero interest in teaching them. I'd much rather use AI to get stuff done, have 3x smaller teams that consist of people who actually want to learn how to be better software engineers and architects.
I would be curious to hear more about software architecture in these times if anybody has good links. It has always been a difficult subject with a lot of religion but a lot of the economics have changed with LLMs and also the LLMs like to ignore your architecture and rewrite random things or duplicate code all over.
Software architecture principles have not changed for decades and will not change with LLMs. The main mistake people make is assume they give it vague input, pay some coins and working code will come out. If you want valid output, you need to the same you would do with a junior dev, be specific, pair with them, watch their work in small increments, never let them submit a large chunk of work. You need to have the design in your head and ideally on paper before you start out. If some parts are not clear, you do experiments to clarify/evaluate. LLMs are good at this. But LLMs also have no taste whatsoever, so you never can allow them to make decisions. They can only be allowed to fill the blanks in your design. I guess the rule is, don't vibe code no matter how tempting it is.
[Using AI chat] made me lazy. It made me stop caring. It made me a worse programmer. It made me depressed. Because I stopped doing the hard work, I stopped learning, I stopped growing, I stopped being the one making the software. [...] [W]riting and coding without AI helped me realize how much I love coding and how using AI was making me a worse programmer and negatively impacting my mental state.
I appreciate OP's decision to write this all out so explicitly.
Upon reflection, I've come to believe that quality software is not about how many features something has or how quickly it is built but rather how cared for it is by the designers, developers, and product people who work on it.
Well, yeah.
Nobody says you should sacrifice the latter for the former (or sacrifice it more just by virtue of using some LLMs).
I'm not saying that it comes for free; just like with almost any powerful tool, one needs to learn to use it responsibly and with judgment, which takes effort and maybe even some discipline (for all the disgust I feel for this word, I guess this is the place to use it).
During the time when I started using the LLM coding tools, I also started to use the chat functionality. I'd ask it to recommend me books. I'd ask for career advice. I'd float business ideas by it. I talked to it like it was a friend. I'd come up with grand ideas, and it was always supportive. And I'd ask it for medical advice. I'm embarrassed to admit this, but one night when I was sick, I described all my symptoms and it told me to go to the Emergency Department at the hospital.
My sympathies.
AI is useful. Don't give it work you want to do yourself, don't trust it with things you need to understand, and use the shit out of it everywhere else. And plant a tree.
Read half of this this moring; came back to finish it, and found the sites been knocked off somehow?