Ask HN: Has anyone else found it harder to review code recently?

I’ve found that junior devs can be incredibly creative and produce seemingly high-level code that they don’t fully understand. Results are often bad because they didn’t consider tradeoffs but still, pretty interesting code. Even before LLMs were a thing, so I wouldn’t attribute it to ChatGPT.

When I read about people using AI to write code, it always seems to me that it would be a lot faster and less hassle if they would write the code themselves, without even considering the fact that it would give them more experience and make them better.<p>I have used some of those tools myself, and for the code that I could use help of an AI tool, I, again and again, receive junk: code that looks plausible but that does not compile, uses apis or libraries that do not exist and so on. In the end, it just made me waste time.

Does anyone else not understand the hype with AI generated code? I almost never use AI at $DAYJOB; only for &quot;search engine&quot; type tasks where I&#x27;m looking for an answer or perspective on something that&#x27;s hard to search for with google - not code generation.<p>Feels like more of a fight with the AI and less time thinking about the bigger picture of the changes being made (e.g. systems involved, business considerations, documentation, etc..)

This sounds like a jobs plan!<p>AI will create the world‘s largest code mess, while also running the entire world, and everything will be slightly broken like in the movie Brazil.

Warm take?<p>Just reject and ask for an explanation OR pair with the coder to have them explain the code.<p>If the reviewer can’t understand it without an explanation the rest of the team can’t understand it unless they git blame then ask the coder, assuming they are still at the company!<p>The “dump 1000 lines of previously unseen and undiscussed code as a code review at a reviewer” method is an antipattern.<p>Either the reviewer should be heavily involved or break it into smaller, well explained chunks with design documentation to point to.<p>This saves everyones time in the long run.

AI generated code is &quot;plausible&quot; seeming, but often has subtle bugs in my experience. Bugs and false assumptions if you don&#x27;t spend a long time telling the AI exactly what you are trying to do. And if you are a novice, you probably don&#x27;t know exactly what you are trying to do in the first place.

I have had this problem a LOT recently, although in the past it has like you&#x27;ve said, been easier to review code.<p>First, let me blame microservices. I heard your eyes roll, but just wait a minute.<p>Second, there&#x27;s all these new frameworks that kind of sit on top of a language just to make it practically a new language. Things like pydantic and sqlalchemy, and other things that parse data, check data, and move data around are their own kind of magic with anything more than a trivial data model. You have to learn how the whole modeling works. Then try to apply it to what is going on. In the end, I&#x27;m generally not sure the code is any better, and it&#x27;s certainly less readable than just a function that parses your data. The frameworks are more complex though, and have all sorts of features. Features you need to learn when someone new brings that framework into your code.<p>But AI&#x27;s can easily generate boilerplate code to handle things that even the author doesn&#x27;t understand. This is not good.<p>So now you&#x27;ve got a &quot;simple&quot; microservice that uses a bunch of libraries and things on top of a language. Such as for python, it&#x27;s pydantic and mypy and numpy, just all sorts of things that are very good, but hard to understand what is going on from a code review basis unless you are also familiar with these frameworks.<p>To go back to blaming microservices, now that everyone wants to have a bunch of little containers running around, even a simple ecosystem can use all sorts of languages, frameworks, build systems, etc. You have to learn all of them for every service.<p>To contrast this with how it used to be in my opinion, one product had one ecosystem. One product would be more monolithic, using fewer moving pieces. Once you learned the pieces, of which there were fewer, you can easily review code.<p>While it can be that many codebases are smaller by using more off the shelf libraries, that smallness hides problems. Many times people don&#x27;t understand the deep ways that the frameworks work, and where the problems may be, especially with performance and scaling.<p>But this small code doesn&#x27;t make nearly as much sense as even larger amounts of just language code where everything is in front of you to review.<p>In the end, you can over-engineer things in a very hidden way.

As a sysadmin I found a crescent level of crappiness and monstrosity that&#x27;s essentially unmanageable, undebuggable, incomprehensible so well... I dot not even try sometimes to review certain yaml&#x2F;js monsters...<p>Most young devs have completely NO knowledge of how a system works, no idea about how to deploy and so on, they are &quot;born on someone else APIs&quot;, meaning some cloud vendor and they do not even want to know. The results are fragile big monsters so buggy that no one looks at logs, at least not if something do not break.

I&#x27;ve gotten to where I sometimes have questioned whether code has come from an LLM or was actually written by a human on some OSS stuff.<p>I&#x27;m not really -good- at it but the explanation for false positives has typically been &#x27;someone not up to a task&#x27; where it gets to the point I call it as an LLM.

Yes, it&#x27;s been happening more and more over the past decade. 10 years ago, code was on average relatively clean and easy to understand. Concepts like &#x27;high cohesion, loose coupling&#x27; were drilled into us. Nowadays, developers barely know what these terms mean. Most developers don&#x27;t seem to take pride in their work anymore. Seemingly, they&#x27;re trying to maximize their lock-in factor via complexity. Probably most do it subconsciously, but the effect is so strong and it has gotten so much worse over time that it feels intentional. Like people have been conditioned to add as much complexity as possible. Coding has become mostly unpleasant nowadays because of this.<p>Most experienced developers know that unnecessary complexity is the absolute worst enemy. You literally cannot overstate the harms of unnecessary complexity... It is absolutely and inherently harmful. But to the junior or mid-level developer, complexity is a sign of intelligence; that, along with the ability to churn out thousands of lines of code per day.<p>On my own projects, I never allow this complexity, but when you&#x27;re working for a company, they don&#x27;t like it if you point out that there is a complexity issue. They&#x27;ll think that maybe you&#x27;re just not smart enough and are jealous of or trying to demoralize the &#x27;genius junior dev&#x27; who is churning out 2k lines per day! Truly Kafkaesque situation.<p>I honestly didn&#x27;t know what to do in my last job. I was doing a lot of PR reviews but I just let the &#x27;most productive&#x27; junior dev continue adding complexity because that was what the founder wanted me to do. Every time I tried to talk about reducing complexity, I would get brushed off, so I just stopped trying.<p>It&#x27;s quite a ridiculous situation actually. Because all the code I write is high quality; highly maintainable, everyone is able to easily add features and make changes to it, but when I work on other people&#x27;s ugly, over-engineered code, it&#x27;s a struggle.<p>So from the outside, it looks like I&#x27;m slow when working with other people&#x27;s code, and it looks like other developers are fast and adaptable since they can easily work with my code... So basically I look like I&#x27;m the one who is a low performer.<p>The winning strategy is clearly to write over-engineered code, then try to socially engineer the situation so that you only end up working on your own code or other people&#x27;s high quality maintainable code (if there is such a thing at your company because people who produce such code tend to get laid off)... While at the same time, you need to try to ensnare your colleagues to work on your complex code so that they end up looking unproductive relative to you... Because huge amount of code + visible features is how directors decide on promotions and layoffs... It&#x27;s always about picking low hanging fruits, sprinkling sugar on top and then personally delivering it to the boss on a silver platter; easy and visible.<p>Much of software engineering nowadays is social engineering; ensuring that you are only assigned to decent quality maintainable and highly visible projects, always hitchhiking on top of the work produced by good developers and dumping your own low-quality output on others to entrap them. Sigh... Then after some time, these big companies end up with ridiculously low productivity expectations... Which is great for social-scheming low performers who are used to this game of racing to the bottom.<p>Also, people like me who can see what&#x27;s going on are never promoted to positions where I can have the last say on such things. It feels like the entire tech economy is just a massive bullshit jobs factory at this stage. All about pretending to be highly productive while in fact being counter-productive.

I’ve been coding over 20 years. I’m guilty. But I’ve also learned that for most domains it doesn’t really matter if you understand the code, if it work, it works. Of course you have to watch for security and safety issues. Maybe ask an llm to explain the code under review?

Perhaps a case of over-engineering?

Honestly, I rarely write more than 10 lines of code without an LLM. I copy existing code files into LLM, ask it to add feature I want. Iterate a bunch of times on the solution. Manually fix the last bits. Check in the software.

On the non-coding side, I’m seeing lots of docs with sections blatantly copy pasted from ChatGPT.<p>Pro-tip: if your normal writing includes typos and sloppy formatting, then be careful pasting a block of three bullet points where each point is a title-cased bold string, followed by exactly two sentences, with perfect grammar and no typos.

Well I faced the same problems earlier today. I had to fix up my LLM generated code before sending it on to my line manager Steve for him to review it. I ran it through ChatGPT a few times to make it seem liked a seasoned developer has written it and to increase its complexity. Its not working out though. My plan now is to use a multi GPT system. I will feed my ChatGPT code into Google Gemini and go around in a circle at least 10 times. That should put Steve off and buy me some more time.