Anecdata, but I'm still finding CC to be absolutely outstanding at writing code.
It's regularly writing systems-level code that would take me months to write by hand in hours, with minimal babysitting, basically no "specs" - just giving it coherent sane direction: like to make sure it tests things in several different ways, for several different cases, including performance, comparing directly to similar implementations (and constantly triple-checking that it actually did what you asked after it said "done").
For $200/mo, I can still run 2-3 clients almost 24/7 pumping out features. I rarely clear my session. I haven't noticed quality declines.
Though, I will say, one random day - I'm not sure if it was dumb luck - or if I was in a test group, CC was literally doing 10x the amount of work / speed that it typically does. I guess strange things are bound to happen if you use it enough?
Related anecdata: IME, there has been a MASSIVE decline in the quality of claude.ai (the chatbot interface). It is so different recently. It feels like a wanna-be crapier version of ChatGPT, instead of what it used to be, which was something that tried to be factual and useful rather than conversational and addictive and sycophantic.
A small app, or a task that touches one clear smaller subsection of a larger codebase, or a refactor that applies the same pattern independently to many different spots in a large codebase - the coding agents do extremely well, better than the median engineer I think.
Basically "do something really hard on this one section of code, whose contract of how it intereacts with other code is clear, documented, and respected" is an ideal case for these tools.
As soon as the codebase is large and there are gotchas, edge cases where one area of the code affects the other, or old requirements - things get treacherous. It will forget something was implemented somewhere else and write a duplicate version, it will hallucinate what the API shapes are, it will assume how a data field is used downstream based on its name and write something incorrect.
IMO you can still work around this and move net-faster, especially with good test coverage, but you certainly have to pay attention. Larger codebases also work better when you started them with CC from the beginning, because it's older code is more likely to actually work how it exepects/hallucinates.
Agreed, but I'm working on something >100k lines of code total (a new language and a runtime).
It helps when you can implement new things as if they're green-field-ish AND THEN implement and plumb them later.
I have my own anecdata but my comment is more about the dissonance here.
I've had interesting chats where it explained that it's choice of tailwind for example was because it had a ton of training knowledge on it.
I've also had it try to build more in one chat than it should many times.
For some reason openai codex handles building too much without failing better - but that is total anecdata from my particular projects and ymmv.
I've had these things try to build big when a little nudge gets them to change direction and not build so much. Explain which libraries and such and asking it to change the tech stack and the steps to build at once seem to make things much better for my use cases.
Also running extra checks and cleanup later is a thing, that sure a human might have seen an obvious thing at time of build, but we have bigger memory context comparatively imho.
A counterpoint is Google saying the vast majority of their code is written by AI. The developers at Google are not inexperienced. They build complex critical systems.
But it still feels odd to me, this contradiction. Yes there’s some skill to using AI but that doesn’t feel enough to explain the gap in perception. Your point would really explain it wonderfully well, but it’s contradicted by pronouncements by major companies.
One thing I would add is that code quality is absolutely tanking. PG mentioned YC companies adopted AI generated code at Google levels years ago. Yesterday I was using the software of one such company and it has “Claude code” levels of bugginess. I see it in a bunch of startups. One of the tells is they seem to experience regressions, which is bizarre. I guess that indicates bugs with their AI generated tests.
Alternatively, it could be there’s a large swath of people out there so stupid they are proud of code your mom can somehow review and suggest improvements in despite being nontechnical.
For example I’m working on a huge data migration right now. The data has to be migrated correctly. If there are any issues I want to fail fast and loud.
Claude hates that philosophy. No matter how many different ways I add my reasons and instructions to stop it to the context, it will constantly push me towards removing crashes and replacing them with “graceful error handling”.
If I didn’t have a strong idea about what I wanted, I would have let it talk me into building the wrong thing.
Claude has no taste and its opinions are mostly those of the most prolific bloggers. Treating Claude like a peer is a terrible idea unless you are very inexperienced. And even then I don’t know if that’s a good idea.
From my observations, generally AI-generated code is average quality.
Even with average quality it can save you a lot of time on some narrowly specialized tasks that would otherwise take you a lot of research and understanding. For example, you can code some deep DSP thingie (say audio) without understanding much what it does and how.
For simpler things like backend or frontend code that doesn't require any special knowledge other than basic backend or frontend - this is where the bars of quality come into play. Some people will be more than happy with AI generated code, others won't be, depending on their experience, also requirements (speed of shipping vs. quality, which almost always resolves to speed) etc.
This is one variable I almost always see in this discussion: the more strict the rules that you give the LLM, the more likely it is to deeply disappoint you
The earlier in the process you use it (ie: scaffolding) the more mileage you will get out of it
It's about accepting fallability and working with it, rather than trying to polish it away with care
And sure, AI could “scaffold” further into controllers and views and maybe even some models, and they probably work ok. It’s then when they don’t, or when I need something tweaked, that the worry becomes “do I really understand what’s going on under the hood? Is the time to understand that worth it? Am I going to run across a small thread that I end up pulling until my 80% done sweater is 95% loose yarn?”
To me the trade-off hasn’t proven worth it yet. Maybe for a personal pet project, and even then I don’t like the idea of letting something else undeterministically touch my system. “But use a VM!” they say, but that’s more overhead than I care for. Just researching the safest way to bootstrap this feels like more effort than value to me.
Lastly, I think that a big part of why I like programming is that I like the act of writing code, understanding how it works, and building something I _know_.
Doing nonsensical things with a library feed it the documentation still busted make it read the source
If you do spot checks, that is woefully inadequate. I have lost count of the number of times when, poring over code a SOTA LLM has produced, I notice a lot of subtle but major issues (and many glaring ones as well), issues a cursory look is unlikely to pick up on. And if you are spending more time going over the code, how is that a massive speed improvement like you make it seem?
And, what do you even mean by 10x the amount of work? I keep saying anybody that starts to spout these sort of anecdotes absolutely does NOT understand real world production level serious software engineering.
Is the model doing 10x the amount of simplification, refactoring, and code pruning an effective senior level software engineer and architect would do? Is it doing 10x the detailed and agonizing architectural (re)work that a strong developer with honed architectural instincts would do?
And if you tell me it's all about accepting the LLM being in the driver's seat and embracing vibe coding, it absolutely does NOT work for anything exceeding a moderate level of complexity. I used to try that several times. Up to now no model is able to write a simple markdown viewer with certain specific features I have wanted for a long time. I really doubt the stories people tell about creating whole compilers with vide coding.
If all you see is and appreciate that it is pumping out 10x features, 10x more code, you are missing the whole point. In my experience you are actually producing a ton of sh*t, sorry.
Spend a few hours writing context files. Spend the rest of the week sipping bourbon.
10x means you could have built something that would have taken 4 or 5 years in the time you've had since Opus 4.5 came out.
Where's your operating system, game engine, new programming language, or complex SaaS app?
Honestly, this more of a question about scope of the application and the potential threat vectors.
If the GP is creating software that will never leave their machine(s) and is for personal usage only, I'd argue the code quality likely doesn't matter. If it's some enterprise production software that hundreds to millions of users depend on, software that manages sensitive data, etc., then I would argue code quality should asymptotically approach perfection.
However, I have many moons of programming under my belt. I would honestly say that I am not sure what good code even is. Good to who? Good for what? Good how?
I truly believe that most competent developers (however one defines competent) would be utterly appalled at the quality of the human-written code on some of the services they frequently use.
I apply the Herbie Hancock philosophy when defining good code. When once asked what is Jazz music, Herbie responded with, "I can't describe it in words, but I know it when I hear it."
That’s the problem. If we had an objective measure of good code, we could just use that instead of code reviews, style guides, and all the other things we do to maintain code quality.
> I truly believe that most competent developers (however one defines competent) would be utterly appalled at the quality of the human-written code on some of the services they frequently use.
Not if you have more than a few years of experience.
But what your point is missing is the reason that software keeps working in the fist, or stays in a good enough state that development doesn’t grind to a halt.
There are people working on those code bases who are constantly at war with the crappy code. At every place I’ve worked over my career, there have been people quietly and not so quietly chipping away at the horrors. My concern is that with AI those people will be overwhelmed.
They can use AI too, but in my experience, the tactical tornadoes get more of a speed boost than the people who care about maintainability.
> the tactical tornadoes get more of a speed boost than the people who care about maintainability.
Why are these not the same people? In my job, I am handed a shovel. Whatever grave I dig, I must lay in. Is that not common? Seriously, I am not being factious. I've had the same job for almost a decade.
Honest question. How does one do that? My workflow is to create one git worktree per feature and start one session per worktree. And then I spent two hours in a worktree talking to Opus and reviewing what it is doing.
Has your output kept pace with the code? Because months in hours means, even pushing those ratios quite far, to be years in days.
Has your roadmap accelerated multiple years in the last few months in terms of verifiable results?
