here’s the uncomfortable truth: AI writes decent code now, and it’ll write better code next year than this year and so on. so coding by hand doesn’t hold up long-term unfortunately (it’s sad to my years of nvim muscle memory). what will potentially hold up is shifting what we practice from producing code to judging code, because i think judgment is the part that stays scarce. there’s a skill gap between people who paste errors into a chat and people who scaffold proper context, write good specs, decompose problems, and verify outputs rigorously
webofbits
achempion
I can’t relate to this experience at all. It’s true that benchmark scores are improving as they should with a better training data.
It doesn’t translate directly to a mergeable code and maintainable software without defects. Unfortunately, it’s still impossible to shortcut expertise in related domains (3d designer will still produce much better result and way faster than me who never saw Blender interface given equal access to LLMs, I’ll produce much better result with LLM than LLM on its own). The code quality isn’t improving with new models either. They still produce bloated and overengineered solutions.
I maintain Waffle library for file uploads and recently I needed to patch the hackney dependency as old major version had CVEs and new major version had missing features. I wrote related PR mostly by hand and used LLMs as a second pair of eyes for a code review. It helped me with catching some potential issues, but I don’t see how I can get the LLM full drive on this feature. It would affect hundreds/thousands of production apps relying on Waffle and bad API/naming decisions would slow down future releases. I also saw other implementations that were heavily LLM assisted and didn’t see them matching quality bar to be mergable.
Also, I’m sure nobody prompts the same way or defines requirements so precisely as in benchmarks (see examples from swe bench, scroll to right for problem_statement and hints_text columns to see the prompt), especially when someone with a little domain experience would try to write a prompt.
11
jam
This is a great question. For anyone reading: What existing product has gotten better? What new product has been created that you think is excellent?
You might like zed’s delta https://delta.dev/. I’ve played around with it and I like the approach they are taking with annotation / diffs. There’s also plannotator if you prefer TUI.
Maybe. I think things are trending towards managing a black box for any system that doesn’t threaten lives if it has bugs, unless token costs jump and make it unsustainable. If programmers are mainly doing code review, that’s still a huge bottleneck. If that’s the case then they might as well manually write (or heavily steer) the important pieces themselves to try to cut down review time. Otherwise it becomes too easy to rubber stamp when you’re overwhelmed by all the generated code.
So I think software creation will become a barbell — on one side, black box management and on the other, a significant amount of manual coding / review with agents assisting in other ways. Though I suppose with black box management, it is possible that the system gets so large that the agents can’t be effective and it collapses.
True. Agents will happily crap out something that works but would not be considered maintainable by existing standards. The question is if it matters. My hunch is it’ll eventually bite but it might take a while.
I like this idea. I think you need to target the right niche and fend off any vibesloppers. Seems like there will be a lot more competition. I hope that quality software can be a differentiator and ultimately prevails.
2
cmo
I changed the emphasis 
When I say go through every line, I am not talking about just reading it.
I will have a deeper and longer lasting understanding if I have written something myself versus reading what someone else has written. Same reason I write notes by hand that I will likely never read. So that the information sticks. This is all pretty well understood, hence courses and books telling you to write the examples yourself, professors making you do the problems rather than just attending a lecture.
It’s also frustrating having to fix all the nonsense it adds in when I ask it to do something of size.
6
martosaur
Will it though? I don’t think I’ve noticed a noticeable improvement since… gpt-5.6 I guess. Models seem to score higher on benchmarks, yes, but code-wise they still fall into all the same traps. Almost as if they developed their own style and hit a ceiling with it.
6
jeremyjh
I felt all these things beginning early 2024. Essentially, the end of my craft was pretty clearly in sight. But, it wasn’t working for me yet.
Agentic coding wasn’t useful to me until February this year, or roughly Opus 4.6. It probably would have been usable at Opus 4.5 but I hadn’t been doing enough coding to notice until I had some vacation in February. Once it actually was working, I became very excited about all the things I could build. And I did build a lot of really cool tools for myself. Now whenever a library or tool I use doesn’t have the features I want, I just have the agent build them.
I never had the patience to learn Nix before but with AI generating the code, all my environments are reproducible - I have complete homelab setup with Nix, Incus and OpenTofu that gives me a ton of pleasure and real utility and I never would have built it without these tools.
This is still the question though. Everyone I know having fun with these tools is doing **** like I described. At work - the AI slop is nauseating. I am so tired of reviewing claudevomit without being able to directly talk to Claude. Not that I ever want to speak to Claude again anywhere, that insufferable smug prick.
Using my own preferred tools and methods, and working on my own, I’m having a great time. But I haven’t heard of anyone having a great time at work. And most people who are trying to build products alone are either just producing more navel-gazing agentic paraphernalia or are grifting like the cryptobros they were last month.
Still, at work I’ve fixed bugs that I’d lost all hope of ever tracking down. I have built improved tooling for standing up ephemeral environments with production snapshots, making production support much easier. We’ve built lots of little integrations between the tools we use that we just never would have spent time on before. But working on the main platform … together with other humans … has become hell.
Big picture - I don’t need a coding job. I mostly do work best understood as product design; I also do solution design with clients, help the business plan and prioritize all our tech spend and all kinds of random stuff that falls in the lap of a founder or early employee. Right now, human minds are required for that, and so I don’t feel any existential pressure from all of this. I’m old enough and close enough to our next exit that I don’t really have to worry about the future in a personal way. But, my son is 17. I don’t know what to suggest for him to do. He wanted to go into engineering or software and I’m just not sure anyone should be doing that. Not because AI can do it all, but I’m concerned it is going to be almost impossible for juniors to actually learn this trade.
9
mudasobwa
Creator of Cure
jeremyjh
I haven’t used it much - because Sol 5.6 is basically good enough for everything I’m using coding agents for - but Astra has done a very good job for me of humanizing claudespeak in presentations and documentation. You might consider giving it a shot on cure-lang.org .
1
Desty
Yes, I’m 100% with you on that. The past 2-3 years has been extremely depressing from a social perspective: I entered this industry because I loved figuring things out and writing programs since getting a C64 and then Atari ST as a teen in the 1990s.
Because of my personal experience, I assumed most other programmers got into it for similar reasons. But that assumption was completely invalidated when ChatGPT and then Copilot and Claude arrived on the scene, and it appeared a large majority of programmers went “oh thank God, I never have to write another line of code… how I have hated this boring, pointless chore!”.
So that was a huge culture shock for me: the people I had thought were my peers apparently hated this craft all along (despite being the same people who used to post excited blog articles about complex Kafka setups and best practices… perhaps that was all just corporate ladder-climbing behaviour).
Having done my doctoral research on an AI-related topic, I’d agree with your assessment that these very-good-plausible-text-predictors are fundamentally not capable of AGI since they have no underlying models of anything other than “what token is likely to appear next”. There is clearly some emergent effect that enables it to produce convincing, even sometimes useful but mostly mediocre results for people, but it doesn’t manipulate any explicit models or learn anything. The fact that these things work at all is a testament to the ludicrous amounts of energy, compute and data pumped through them.
At first I tried ChatGPT for programming help when it came out, and realised I utterly detested it and could feel my problem-solving skills weakening within a week or two. It sucked out every ounce of joy I got from programming, and I quickly resolved not to use it ever again in that capacity. After that, I used it as a research tool to point me to published papers, but even then, its summaries and claims about papers were just way off base.
Now I’m sick of LLMs, but much more importantly, sick of the amount of airtime they consume - not this thread, which I find cathartic and humanising… but all the programming newsletters I used to read (Hackernoon, Medium etc) are just drowned with either slop articles, or articles about how to use Codex etc more effectively.
When I open newsletter emails now, I scroll down and delete the email as soon as genAI topics are mentioned.
When a recommendation pops up for a new game (even an open-source one), I click into it, already looking for signs of slop. I’ll check the Github repo for AGENTS.md (or the “banned user” banner for Claude).
And, getting away from the social and mental areas… I just can’t trust any codebase that has been exposed to LLM-generated code. In particular I’ve stopped using Elixir which has been utterly infested with LLM commits for a while. IMO it’s too much of a risk to expose customers to something that gets a score of 0 (F) from slopscan, and more importantly I just don’t want that in my life. It’s such an ergonomic language, but there are others which haven’t been tainted by those plagiarism engines that flood codebases with gigantic PRs and make understanding impossible. Getting away from LLM-generated code is almost impossible (it’s even in Linux now, so I’d have to move to netbsd or even Haiku), but I’ll minimise it as much as reasonably possible.
17
jeremyjh
This is only a meaningful description of a base model that has only been pre-trained. After post-training the model is not predicting the most likely token to appear next in the internet corpus - it is predicting the token that is most likely to lead to a solution that satisfies the grader in an RLVR. Most of the model’s training has taught it how to design and develop solutions, and use tools to research and test their solutions before submitting them for grading. Transformer-based token predictors are just the bottom level substrate in which those learnings are implemented.
2