Professional diligence, perhaps? A desire to not be blindly led into the kind of narrow, often first-party stack which is so often proposed by Claude Code?
With all due respect, not everyone is afflicted with the lack of care sufficient to allow them to launch vibe coded apps as low quality as https://podnami.com. Considered technology choices are one such aspect of the practice of caring about what you're building.
Stronger foundation (as to mean better tested), less accidental complexity of reinventing everything, transferability of knowledge, easier onboarding and review of changes.
Your argument is a mirror of the snark question "why don't LLMs write in assembly?" for those not looking at the output at all.
Statically typed languages with compiler hints are the absolute best languages for LLM's to work with. Successful compilation is incredible feedback, and it basically just means that there is a higher chance that the feature is in a complete and working state at the end of each agent iteration.
If anything, LLMs make it easier to choose from a broad set of options. The tradeoffs are the same as pre-LLM days, but the learning curve is more favorable.
I’ve spent time doing software at VW and a few of its subsidiaries, and this matches my experience.
Compliance is everything, and SAFe (Scaled Agile) is deployed as a blunt instrument.
Management treats software exactly like hardware production lines—everything is just an "engineering process" that can be optimized on a spreadsheet.
The underlying assumption is that individual engineering talent is just an interchangeable commodity. Once you view developers as replaceable cogs, outsourcing the entire infrastructure to the lowest bidder in India becomes the logical conclusion.
It’s a textbook case of process-over-people driving institutional tech debt.
> Management treats software exactly like hardware production lines
That's exactly my observation as well. Classic hardware-producing companies have an immense respect on the step of entering mass-production, as whatever issue that slipped through will be multiplied and physically spread across the world.
So they come from the mindset that the dominant mindset is to minimize the SURFACE-area of potential risk. This makes it really hard for them to compete in software-space, because in software the dominant mindset is to just estimate risk.
Neither is wrong, but applied vice-versa is.
- If you treat software like hardware, you end up cutting out everything that could make your product fit more than your decided main use-case.
- If you treat hardware like software, you're placing a bet on behalf of your customer that the product "will be fine", and a (very expensive) bet that this product won't create an aftermath which may destroy your entire company.
Companies which can't manage the distinction here end up putting hardware in the hands of customers they should have built differently and then spend all their resources on software updates just to somehow keep the core function working.
Why shouldn’t software be treated with the same rigor at Volkswagen scale?
Pretty much all software products typically talked about on HN are laughable at that scale, they have crashes or weird bugs way more often than the six sigma norm of 99.9999% reliability.
For example, I don’t think it’s even possible nowadays to buy a new iPad and use it with default apps and settings for any significant duration continuously. It’s well under 1 million minutes of uptime before failure and a hard restart is needed.
So anything more complex than the simplest possible use case of an iPad is even more of a joke under hardware norms.
> Why shouldn’t software be treated with the same rigor at Volkswagen scale?
No one said that it shouldn't.
What I wrote is, that the approach of minimizing any SURFACE of risk in software creates the (subjectively good and solid) software of previous car-generations (in Volkswagen terms: MIB2 ~ a bit downhill already in MIB3): A solid, predictable and closed product fulfilling its core use-case.
But it DOESN'T create a user experience with those "fun" niche features, competitive remote-access Smartphone features, exposed API's, sudden new features during lifecycle, funny "ludicrous modes" etc.
And today's customers are demanding those features, it's now a hygiene factor for a premium experience on Smartphones as well as on cars.
A Tesla is not considered a "Premium" car because of its premium hardware or manufacturing quality. They disrupted the car-industry by being the first to apply a software-dev mindset to it, and the consumer perceives this as premium.
Apparently I need to clarify, as it's not obvious from my previous comments: _I_ don't consider Tesla premium at all, it is NOT a premium hardware company.
It's bad quality hardware hiding behind an cleaner engine and some software features.
It is EXACTLY the product of a hardware company which keeps treating hardware-production like it's software, as described above.
Coding on the go is definitely a use case. Especially if you want something answered or make a big decision on an architectural change but need both docs and get informed about trade offs. The biggest productivity boost we made was auto deployment of new PRs so that the agent can make live changes and we can review stuff on the to. Suddenly you can do QA in a park in the sunshine without being in front of a desktop
I know of a publicly traded company which in its early years was built on beer. Literally. 3 guys in a co-working space in Cambridge, MA. Beer fueled their progress. 15 years later the software is still the backbone of the org.
Anecdotally OpenAI is trying to get into our enterprise tooth and nail, and have offered unlimited tokens until summer.
Gave GPT5.4 a try because of this and honestly I don’t know if we are getting some extra treatment, but running it at extra high effort the last 30 days I’ve barely see it make any mistakes.
At some points even the reasoning traces brought a smile to my face as it preemptively followed things that I had forgotten to instruct it about but were critical to get a specific part of our data integrity 100% correct.
Same here. I feel like all of these shenanigans could be because Anthropic are compute constrained, forcing then to take reckless risks around reducing it.
Same here. I was a fervent Claude code user at $200/mo until Opus4.7.
Freezing your IDE version is now a thing of the past, the new reality is that we can't expect agentic dev workflows to be consistent and I see too many people (including myself) getting burned by going the single-provider route.
On one hand I’m glad to finally see anthropic communicate on this but at this point all I have to say is… time to diversify?
Opus 4.7 via code has been inconsistent for me. Sometimes, it feels like working with a brilliant collaborator and is as good as 4.5 and 4.6 were. Other times, it takes dumb and lazy short cuts. It can be quite frustrating. Its response when I tell it it did something wrong is often to write a memory... which is then does not always read. The inconsistency isn't due to session length or age either. These are all new sessions. I feel like sometimes, I get routed do a dumber model or some other hidden setting is applied.
My experience as well. This is even worse than just having a mediocre model, because I can work around that. The inconsistency means it produces different outputs for the same prompt, and I can't rely on that as a business tool.
They lost me a little before then - Claude Code's regressions were so very obvious and there's no sign they've learned their lesson in this article or in the comments of those who work on Claude Code on HN. They'll continue to tweak and generally mess around with a product people are using, altering the behaviour without notice in ways that can severely impact use, for months! GPT5.4 has been remarkably consistent and capable, as a replacement. I've cancelled my max plan.
I started using Claude heavily on the 20th after having not used it for a year. Largely Sonnet 4.6, web, cowork and code. Can confidently say it is significantly worse than this time a year ago and regret that my new employer requires we use it, and only it.
I’ve never been one to complain about new models, and also didn’t experience most of the issues folks were citing about Claude Code over the last couple months. I’ve been using it since release, happy with almost each new update.
Until Opus 4.7 - this is the first time I rolled back to a previous model.
Personality-wise it’s the worst of AI, “it’s not x, it’s y”, strong short sentences, in general a bulshitty vibe, also gaslighting me that it fixed something even though it didn’t actually check.
I’m not sure what’s up, maybe it’s tuned for harnesses like Claude Design (which is great btw) where there’s an independent judge to check it, but for now, Opus 4.6 it is.
I noticed the difference, but coming from Gemini and xAI models it wasn’t that glaring. I still find that Opus makes much better plans than anything else I’ve tried, and it’s been very good at catching my mistakes in using public-key cryptography, also finding out why my crsqlite queries were failing despite no official documentation on the topic.
I’d never use such an expensive model for coding, so that might explain why I have little to complain about.
I find that it is better at thinking broadly and at a high level, on tasks that are tangential to coding like UX flows, product management and planning of complex implementations. I have yet to see it perform better than either Opus 4.6 or 4.7 though.
extra high burns tokens i find. ( run 5.4 on medium for 90% of the tasks and high if i see medium struggling and its very focused and make minimum changes.
Yeah but it also then strikes the perfect balance between being meticulous and pragmatic. Also it pushes back much more often than other models in that mode.
What's your workflow like? I'd be curious to test OpenAI out again but Claude Code is how I use the models. Does it require relearning another workflow?
Claude code can be configured with custom /slash commands and other details that don't necessarily transfer over to codex. /remote-control in cc is really great for walking away from my computer and continuing from my phone, for instance.
I guess I'm asking if their CLI tool is the same or if it functions different. I've never used anything besides CC so I wouldn't know if it's basically the same thing
Do you have to know Assembler to be able to write code in Java? With the point being that you rarely know the underlying mechanics - and the same if true for vibe coding.
Nah, but you have to actually put the work in to get the credit. Lazily vibe coding slop and then passing it off as your work is like claiming you cooked a microwave meal.
In microgpt, there's no alignment. It's all pretraining (learning to predict the next token). But for production systems, models go through post-training, often with some sort of reinforcement learning which modifies the model so that it produces a different probability distribution over output tokens.
But the model "shape" and computation graph itself doesn't change as a result of post-training. All that changes is the weights in the matrices.
Do they not? Many phone functions are already available through voice assistants, and have been for a very long time, and yet the vast majority of people still prefer to use them with the UI. Clicking on the weather icon is much easier than asking a chatbot "what's the weather like?"
My elderly mother has an essential tremor (though only in one hand now due to successful ultrasound treatment!) and she would still rather suffer through all her errors with a touch interface than use voice commands.
Some people seem to think that Deckard’s speech-controlled CSI software in Blade Runner is actually something to strive for, UX-wise. As if it makes any sense to use strictly nonvisual, non-two-dimensional affordances to work with visual data.
The sad part is that while everyone is chasing new interface modalities, the traditional 2D UI is slowly getting worse thanks to questionable design trends and a lack of interest.
