> Software fundamentals - though useful - are table stakes skills at this point.
I'm having a difficult time even seeing what we're talking about here. I see "seniors" in our industry that don't know what I would call the fundamentals of programming or software development; apparently not all fundamentals are created equal.
Not even SotA models are good enough to generate code (beyond functions or small, very simple modules) that I'd be happy shipping, so I've decided to just not have them do that. And given this, it has basically turned out that what's left is information gathering + analysis + design overview stuff.
I've just recently started trying out DeepSeek 4 Flash and I was very skeptical at first because I've had some really good experiences with GPT-5.{4,5}, and couldn't possibly believe that this model they charge nothing for could give me similar results, but it absolutely shreds through things and ends up giving me very good answers in almost no time. I also like that it doesn't really seem to have much personality, it's given me mostly just facts and data so far without any additions to the prompt by me.
In my own agent I also specifically prompt to remove flowery language, snark, etc., but I haven't tried it with models like GPT-5.x which I've found has too much personality and tries to make it seem like I'm talking to a human too much.
My personal experience with trying to front-load tons of planning and speccing out with LLMs is that at best it's a small improvement on code quality but with considerably more time spent.
As a result I've abandoned the idea of having LLMs generate code except for very small, localized and tightly scoped things. They really can't produce much more than a function or a small module without shitting the bed (last time I vibecoded was with Opus 4.6, Composer 2 and GPT-5.4). I use it almost entirely as another signal in analysis, which naturally makes it fit in better because all the other signals (reading the code, stepping through the code, writing the code myself) are already there so when the LLM points things out the information it actually renders can be taken in much more easily (and seen through more easily when it's false or irrelevant).
I think it's neat that people find fun ways to develop, but I think dressing up vibecoding in a fancy dress and layering SpecLang, sometimes in multiple steps, on top of it, is an exercise in trying to use the tool more instead of trying to use it in its most useful capacity.
I expect you'll be told to try Opus 4.7, and in short, JuSt WaiT FoR ThE NexT MoDel, BRo.
This has been my experience every time I've suggested that there are any sort of inherent ontological/conceptual or computational limits to the sophistication of LLM mimicry.
Anyone who frivolously suggests throwing away possible independence in favor of dependence on a Silicon Valley company is either incredibly naïve or acting in bad faith.
Not necessarily so. I can see how a bid to predict how thing will be in 1 year in AI-based coding is likely a losing one. So the idea is to extract the maximum value now, and turn it into profits that would buy you whatever is adequate for the next steps. For comparison, the AI-based coding landscape a year ago, in May 2025, wasn't even close to what we have now, and half the key tools did not exist.
OTOH, as we see, the larger models demonstrate diminishing returns, smaller models demonstrate improvements, and hardware does not show any signs of becoming cheaper, so holding on existing decent GPUs may, too, be a winning strategy in longer term.
I'll choose not to respond to your personal attack.
But in term of actually running a dev team - you are free to use QWEN or another quantized local model that can run on an RTX 5090 for coding if it makes you feel more independence. However you would struggle and spend many many more hours achieving the same thing, with a lot more debugging time, long delays before it's done, and many more prompts.
It's just not the right approach. I use QWEN and other local models all the time, but for more clearly defined monitoring and classification tasks.
I lost about 50% functionality in my hands in 2018-2019 and couldn't type more than an hour or so per day, what really saved me was dictation via Dragon NaturallySpeaking, and for coding I used dragonfly to create programming grammars. I'm happy for this guy for finding a solution but LLMs (in this shape) were late to the party.
Were you as fast dictating code (not just outputting it, but navigating, editing, refactoring etc.) as you were before needing it? If it was slower, then LLM assistance is significant in making such disability no longer an impairment for the job.
Once you get into it it's not any slower than typing; if anything I think you can actually be faster with a good enough grammar, and it's precise too.
The problem I had eventually was that dictating all day is itself something that wears you out, so you get hoarse and honestly a bit more tired from it when you do it too much.
But no, overall a good Dragonfly grammar should be something that keeps you at least up to par with your typing speed.
While I agree that Mercurial probably lost a ton of users for not clearly addressing performance issues for quite a while, I've found it's never too late to switch from using git, to be honest. I am personally using Fossil for all of my projects and it's been a great experience overall. I didn't have particularly challenging needs when it came to git so I wouldn't say that it's been a major headache overall for me, but I also think Fossil just is better as a default than git for everything that I do (and everything I've ever worked on):
- Assume we want to sync to main repo as a default when issuing operations, keeping us in sync more often and easily
- `commit` just commits all indexed files with changes, no need for staging
- Worktrees by default (admittedly this is more of a convention and you can certainly do the same with `git worktree`, but it's very prominent in how they show you how to use Fossil).
- `fossil ui` for having your own mini-GitHub is great, and having changes you make there sync with your remote is incredibly convenient.
- `fossil serve` on your remote is a great way to make your mini-GitHub an actual persistent service. I've used SSH remotes for `git` but as far as setup goes this actually is a close second in terms of convenience. Nevermind setting up a more involved forge; `fossil serve` is overperforming for what you get by a lot here.
All in all switching to Fossil after a ton of time using git has been a great experience overall.
Wouldn’t previous human-coded implementations already be in LLMs’ training sets? OP article even mentions a number of previous implementations. And even recent LLM “clean room” designs seem to gloss over how the final LLM implementer still has access to previous implementations’ code.
Still, OP claims to have done the best job to date at creating (via AI) specs, and the non-optimal Rust implementation, so a net gain?
I don’t believe any of the extant open source rar implementations cover the range of features and versions OP’s does. I think that’s the point - OP’s isn’t the cleanest or fastest implementation, but it is the most broad open source version available.
It probably wouldn't take a single person who knew what they were doing more than a year to re-implement Bun in basically anything, by hand and from scratch, i.e. not even looking at source. Writing the code for something you already understand and have built before is incredibly fast.
I'm sure they'll market what you said, but it's so ridiculous that I would hope people would see through this stuff.
I'm having a difficult time even seeing what we're talking about here. I see "seniors" in our industry that don't know what I would call the fundamentals of programming or software development; apparently not all fundamentals are created equal.
reply