If you’re honestly asking, LLMs are much better at coding than any other skill right now. On one hand there’s a ton of high quality open source training data that appropriated, on the other code is structured language so is very well suited for what models “are”. Plus, code is mechanically verifiable. If you have a bunch of tests, or have the model write tests, it can check its work as it goes.

Practically, the new high end models, GPT 5.4 or Claude Opus 4.6, can write better code faster than most people can type. It’s not like 2 years ago when the code mostly wouldn’t build, rather they can write hundreds or thousands of lines of code that works first try. I’m no blind supporter of AI, and it’s very emotionally complicated watching it after years honing the craft, but for most tasks it’s simple reality that you can do more with AI than without it. Whether it’s higher quality, higher volume, or integrating knowledge you don’t have.

Professionally I don’t feel like I have a choice, if I want to stay employed in the field at least.

A later commenter mentioned an AI version of TDD, and I lean heavy into that. I structure the process so it’s explicit what observable outcomes need to work before it returns, and it needs to actually test to validate they work. Cause otherwise yeah I’ve had them fail so hard they report total success when the program can’t even compile.

The setup I use that’s helped a lot of shortcomings is thorough design, development, and technical docs, Claude Code with Claude 4.5 Sonnet them Opus, with search and other web tools. Brownfield designs and off the shelf components help a lot, keeping in mind quality is dependent on tasks being in distribution.

The vast majority of my experience was Claude Code with Sonnet 4.5 now Opus 4.5. I usually have detailed design documents going in, have it follow TDD, and use very brownfield designs and/or off the shelf components. Some of em I call glue apps since they mostly connect very well covered patterns. Giving them access to search engines, webpage to markdown, in general the ability to do everything within their docker sandbox is also critical, especially with newer libraries.

So on further reflection, I’ve tuned the process to avoid what they’re bad at and lean into what they’re good at.

So this article is basically a puff piece for Code Rabbit, a company that sells AI code review tooling/services. They studied 470 merge/pull requests, 320 AI and 150 human control. They don’t specify what projects, which model, or when, at least without signing up to get their full “white paper”. For all that’s said this could be GPT 4 from 2024.

I’m a professional developer, and currently by volume I’m confident latest models, Claude 4.5 Opus, GPT 5.2, Gemini 3 Pro, are able to write better, cleaner code than me. They still need high level and architectural guidance, and sometimes overt intervention, but on average they can do it better, faster, and cheaper than me.

A lot of articles and forums posts like this feel like cope. I’m not happy about it, but pretending it’s not happening isn’t gonna keep me employed.

Source of the article: https://www.coderabbit.ai/blog/state-of-ai-vs-human-code-generation-report

The last 6 to 12 months of open models has pretty clearly shown you can substantially better results with the same model size or the same results with smaller model size. Eg Llama 3. 1 405B being basically equal to Llama 3.3 70B or R1-0528 being substantially better than R1. The little information available about GPT 5 suggests it uses mixture of experts and dynamic routing to different models, both of which can reduce computation cost dramatically. Additionally, simplifying the model catalogue from 9ish(?) to 3, when combined with their enormous traffic, will mean higher utilization of batch runs. Fuller batches run more efficiently on a per query basis.

Basically they can’t know for sure.

LLMs have made it really clear when previous concepts actually grouped things that were distinct. Not so long ago, Chess was thought to be uniquely human, until it wasn’t, and language was thought to imply intelligence behind it, until it wasn’t.

So let’s separate out some concerns and ask what exactly we mean by engineering. To me, engineering means solving a problem. For someone, for myself, for theory, whatever. Why do we want to solve the problem, what we want to do to solve the problem, and how we do that often blurred together. Now, AI can supply the how in abundance. Too much abundance, even. So humans should move up the stack, focus on what problem to solve and why we want to solve it. Then, go into detail to describe what that solution looks like. So for example, making a UI in Figma or writing a few sentences on how a user would actually do the thing. Then, hand that off to the AI once you think it’s sufficiently defined.

The author misses a step in the engineering loop that’s important though. Plans almost always involve hidden assumptions and undefined or underdefined behavior that implementation will uncover. Even more so with AI, you can’t just throw a plan and expect good results, the humans need to come back, figure out what was underdefined or not actually what they wanted, and update the plan. People can ‘imagine’ rotating an apple in their head, but most of them will fail utterly if asked to draw it; they’re holding the idea of rotating an apple, not actually rotating the apple, and application forces realization of the difference.

I’ve used it most extensively doing Ruby on Rails greenfield apps, and also some JS front ends, some Python mid sized apps, and some Rust and Nix utilities. You’re absolutely right about it struggling with code base scale, I had to rework the design process around this. Essentially, design documentation telling the story, workflow documentation describing in detail every possible functionality, and an iteration schedule. So the why, what, and how formalized and in detail, in that order. It can generate the bulk of those documents given high level explanations, but require humans to edit them before making them the ‘golden’ references. Test driven development is beyond critical, telling it everywhere to use it extensively with writing failing tests first seems to work best.

So to actually have it do a thing I load those documents into context, give it a set unit of work from the iteration schedule, and work on something else.

It does go down some seriously wrong paths sometimes, like writing hacky work arounds if it incorrectly diagnosing some obscure problem. I’ve had a few near misses where it tried to sneak in stuff that would bury future work in technical debt. Most problematic is it’s just subtle enough that a junior dev might miss it; they’d probably get sent down a rabbit hole with several layers of spaghetti obscuring the problem.

I’m a professional developer and have tested AI tools extensively over the last few years as they develop. The economic implications of the advancements made over the last few months are simply impossible to ignore. The tools aren’t perfect, and you certainly need to structure their use around their strengths and weaknesses, but assigned to the right tasks they can be 10% or less of the cost with better results. I’ve yet to have a project where I’ve used them and they didn’t need an experienced engineer to jump in and research an obscure or complex bug, have a dumb architectural choice rejected, or verify if stuff actually works (they like reporting success when they shouldn’t), but again the economics; the dev can be doing other stuff 90% of the time.

Don’t get me wrong, on the current trajectory this tech would probably lead to deeply terrible socioeconomic outcomes, probably techno neofeudalism, but for an individual developer putting food on the table I don’t see it as much of a choice. It’s like the industrial revolution again, but for cognitive work.

Most if not all leading models use synthetic data extensively to do exactly this. However, the synthetic data needs to be well defined and essentially programmed by the data scientists. If you don’t define the data very carefully, ideally math or programs you can verify as correct automatically, it’s worse than useless. The scope is usually very narrow, no hitchhikers guide to the galaxy rewrite.

But in any case he’s probably just parroting whatever his engineers pitched him to look smart and in charge.

I had some similar and obscure corruption issues that wound up being a symptom of failing ram in a main server node. After that, only issues have been conflicts. So I’d suggest checking hardware health in addition to the ideas about backups vs sync.

Stories like this are sometimes more complicated than they appear. The infamous examples of $500 hammers, for example, were anti sparking hammers for working around flammables or munitions, hence requiring special materials, certification, and low production runs.

For this case, we have liquid hand soap dispensed by a pump. Pumps require a sealed vessel. Unlike commercial planes, military planes are required to anticipate prolonged operation with an unpressurized cabin. At max altitude of a C17, atmospheric pressure is only 20% of sea level. Off the shelf dispensers are unlikely to be designed to withstand that pressure difference, let alone function normally. In a high demand environment like aerospace, even apparently minor failures like an exploding soap container needs to be taken seriously due to the possibility of unexpected cascading failures. Why not use bar soap, then? Unfortunately this too has complications, like not being able to be securely mounted, liquid soaps having superior hygiene and cross contamination characteristics, and necessity for military standardized soap, sometimes designed for heavy metal, eg lead, which is likely if the cargo were munitions.

This unusual set of requirements unlikely to be seen outside the military context, so whether designed by Boeing or off the shelf the unit would likely have low quantity manufacturing runs, significantly increasing per unit costs. Combine that with the necessary certifications and the per unit costs balloon even further.

While a soap dispenser having an 80x markup seems absurd, it might be more reasonable than it seems at first glance. To be clear, there absolutely is military contractor graft. I just don’t expect even a $10,000 soap dispenser would be a substantial proportion if it even within the C17.

I recently removed in editor AI cause I noticed I was acquiring muscle memory for my brain, not thinking through the rest past the start of a snippet that would get an LLM to auto complete. I’m still using LLMs, particularly for languages and libraries I’m not familiar with, but using the artifacts editors in ChatGPT and Claude.

The comments from that article are some of the most vitriolic I’ve ever seen on a technical issue. Goes to prove the maintainer’s point though.

Some are good for a laugh though, like assertions that Rust in the kernel is a Microsoft sabotage op or LLVM is for grifters and thieves.

FOSS in general needs better means of financial support. While the software is free and libre, developer time is not, and ultimately they gotta eat and pay bills. I hope they get positive results and don’t catch much unnecessary flak.

People haven’t really changed. As always, power corrupts. When the rewards are great enough, it seems people are often enough willing to compromise their integrity.

My first programming experience, an online class, was in a Linux VM. Linux made programming easy and delightful, Windows always made it a huge pain. As time went on, more of what I did was easier on Linux, and now everything is.

Title worried me for a moment that they were dropping Steam Input; happy to see they seem intent on the opposite.

The moment that shocked me was when printers, network cards, and even motherboard integrated Ethernet didn’t work on Windows without driver downloads but Linux was plug and play. Full reversal of the situation.

Have you tried recent models? They’re not perfect no, but they can usually get you most of the way there if not all the way. If you know how to structure the problem and prompt, granted.

Them using Google indexes anonymously isn’t intending to solve the problem you think it is. It’s more about incentive structures. Google’s “free” search optimizes for ad revenue now. The API access doesn’t as much, and Kagi certainly doesn’t have an ad incentive. So privacy is a nice bonus, but the real benefit is a customer serving incentive structure.