this in fact is excellent articulated what I also strongly feel.
went from decades of oldschool crafting by hand and got good at it, then went into vibecoding and the experience of decades made me leverage agents significantly better than other peers, but it is also addicting and getting lazy, fast. to the point I type "git push" into claude regularly, not only because I am too lazy to switch my UI tab, but also because occasionally there is this bit of git friction with some local/remote states and claude/codex just resolve it instead of me getting distracted with it.
and now I essentially force my squirrel brain the opposite way again. everything new and exciting and MVP grade ideation happens with vibecoding in hours/days, shippable in isolation somehow so strictly limited blast radius, and then only the <10% of stuff that turn out valuable enough I go back to and rewrite/extract by hand, literally typing it out manually, _without_ autocomplete beyond what a lsp in zed gives me (no cursor magic, ...). AI is only used as a assistant to discuss solution options quickly like a coworker/mentor on the side, but even then I type the result manually.
And it feels like an alien now to do this but its getting better slowly and boy, even my handcrafted code gets better than ever, slow and deliberate, and also I switch to OCaml now for everything that matters, having learned it through LLMs for weeks. No more Node/Go/Rust for me, finally settled after like 20 years. So a quick slop of vibed experiments (react, node, go services, whatever fits best quickly), and the important stuff then slow careful crafted rock solid and lightning fast ocaml.
feels like this is the way. even if it hurts still, kinda like DOMS after going back to gym after long breaks.
just used mistral for a database/scraping creation tool and ended at <10k€ in token costs (via openrouter), beating gpt5.4-mini in output quality and speed and costs after actual testing A/B fairly. so its a super scoped task to be performed hundreds of thousands of time for some automation and mistral just did it better across all dimensions that gpt-5.4-mini. of course thats not a headline in terms of frontier model competitiveness, but for "the boring parts" it just was flat out better than anything else consistently. bonus points it handles mixed-language-content with nuances surprisingly well to turn web content in the wild into structured data really good and fast.
the very point is that theft means you no longer have something since someone else has. copying is you still have it and someone else now too. there is no harm done by copying, except you actually believe that exclusivity as a separate concept is important to you. (debatable, I don't).
cursor feels so 2025 to me guys. these days zed is just way better for my macbook battery and with acp can talk literally to my installed claude code and codex CLI tools, plus their own and custom providers ontop. I was kind of a decade of a vscode user and always just stayed through the evolutions until cursor, but at some point I just need a lean fast editor+lsp combo, git included and a chat pane next to it that uses my real subs underneath easily. (also: codex-cli can spawn and manage subagents and _resume_ them, acting like a real manager).
could be only me though, but longer interactions over days makes my codex gui app grind to a crawl and cursor was not only expensive with opus via api costs but also heating my room a lot. now I have a dozen zed instances open all crunching along with LLMs barely noticeable on system load (except the occasianal testuite runs but thats expected).
Okay so if this model is half a year behind, so let’s say January opus pre-nerf, this is it.
Inference is actually quite cheap for token costs, the frontier labs burn most of their money on training new models, priced into their token costs ontop of some margins and paying record salaries. So if this goes open, distills are tried out, independent providers around the world host it with actual price competition, the house of cards for anthropic collapses pre-ipo. The floor is opus (open models caught up), the current ceiling is Mythos (self inflicted ban due to the safety bullshit theater), and no way out.
It’s really comical I think it’s even the same guy that warned about gpt2 being too dangerous to release, well that mindset seems to now doing existential harm to anthropic, while the rest of the world essentially laughs and progresses anyway.
wholeheartedly agree. even as a techie with decades of experience, the meta business world is a complete clusterfuck its unbelievable bad. yes they bought and assimilated things like insta or whatsapp, but with their sheer amounts of engineers plus AI coding now even its not acceptable they cannot even get the basic UX of their money printing machine merged into one coherent thing? The god damn creators of react give us us this messed up slow garbage of a maze of config pages weakly connected with complete design breaks and menus I have to navigate with codex/claude to finally find the thing I need? I man accidential complexity can happen, but what are the thousands of engineers and trillions of tokens used for? its just sad
Between professional Elixir, Go Rust and Node over decades now I am arriving actually at OCaml now. Using LLMs to actually teach it to me.
Andd boy, a REAL type system is just something i won't ever again compromise upon. I mean yeah I did many years of Ruby/Rails and loved it back then, and Elixir in that regards at least on surface felt strictly better (sweet pattern matching, pipes, ...) but just SO MUCH CODE is written either at runtime or in loads of tests that essentially make up for the lack of a compiler guarantee about type errors i cannot unsee it anymore. Rust is way better here for example for sure, Trait system and all, but here the compile time tax is very real even after fiddling with optimal crate splits. Plus _sometimes_ a bit of simple mutable code just hits home in a few lines instead of often slower pure FP equivalents.
Happy to see that Elixir finally after years in the making is arriving somewhere, but I essentially left the ecosystem now since I really do either TDD (Type driven Development) now or quick solutions with node/go when quality isn't the concern... and now I discover OCaml (with Effects based multicore now) and yes the syntax is _a bit_ alien but damn it checks all boxes of all techstacks I ever wanted. I can write nearly Elixir style code, pattern match pipes and all, I can write (nobody does but I could) failry powerful OOP stuff, compile instantly, in a statically linked binary, with true parallelism, and a type system that is amazing (don't get me started about module functors). Beam is a impressive feat of engineering, but its also moving like molasses and deployment is nontrivial and quite cumbersome to operate (at least people need quite a lot of learning curves until theyre comfortable with this powerful beast). And then there is OCaml. And the tradeoff here is on the human side, nearly no one knows it, learning curve is high, so statistically no team would pick it in most businesses or has experience with it, and that specific situation is personally for me irrelevant now as a solo builder in an LLM age.
Lets see how good this becomes at some point, I am watching and would have loved to have this at least gradual typing available years ago!
with web framewworks like actix or rocket itsd actually not much different to python flask or nodejs express or .... . but i am a cheap person and prefer to cram as much stuff onto my $4 DO droplets as possible, and rust brings you very far here. might be not a concern for funded startups or big enterprises, but deploying small/mid projects to essentially minimal hardware really makes my wallet happy, plus I essentially never ever had to fix runtime issues, at all, for years. and this was before LLMs were a thing, so handwritten Rust in a fullstack way, at times even WASM frontend SPAs that still kinda just work.
I use both languages and in spirit you are right, btu in my experience its more nuanced. first of all the way inline unittests in the same file is a way to have very fast cycle times in TDD too in many situations, still slower than go full compile yet much less painful than full recompile in rust. second, you typically need way less debugging cycles in rust to begin with. so its more like slower but fewer cycles.
true fullstack (not only tech parts - from marketing to product to boring IT and financials) makers now have insane speedups, IF they previously did everything by hand properly and found ways to be good at it solo. Often the only scarce resource was your own time (or more correct: attention/focus hours per day) for execution and research. Now offloading stuff to agents, especially since you know the domain and the shortcomings by heart and when to (not) not trust AI is supercharged.
People fail to reap benefits for organizational scale and mostly fail completely, they try to make AI fit a human process spaghetti theater somehow, but a lone wolf can just change himself entirely in days and be adapted and leverage AI-first if he wants to. just like that. no coordination needed beyond rewriting personal scripts, which is fast with LLMs too.
in fact engineering job openings are skyrocketing, just far beyond the big megacorps. AI was a neat excuse to correct the covid overgiring phases, now axes happened while lots of people built lots of new things they couldn't before and at some point it "swamps" under the complexities of reality at scaling up, thus engineers are needed. jibs are shuffling around yes, but demand will be surging even more soon.
well the worse it gets the closer the reset button comes. guillotine can and will end capitalism if people are affected too much, and they don't care about any logic when suffering is high enough. happend in history every single time, will happen again, and no one saw it coming or believed it would happen again. still, it did. and will. like always.
Maybe I am in a minority position here, but despite me vibecoding for many months now (havent written a single line by hand and forced me todo so in the beginning), I still have my IDE open right next to Codex/CC and while the LLM is crucnching along and doing TDD loops I actually read whats created/changed and just sit with it judding if its only right on surface or semantically stupid underneath, essentially realtime-architecting and steering the code agents sometimes even midflight. so I do end up with these 200k+ LoC projects now since typing is lightning fast and 2/3 of my codebase is tests (I insist on regression tests after every steer) but in fact I perfectly know what each piece is doing and where it is, as well as the still not optimal parts and have a mental list for refactoring it later when I have time or a spare parallel agent can do it when feature work isn't crossing the same areas.
so I COULD take over by hand again like I did the decades before just fine, but I refuse to and instead play a codebase like a RTS - lots of stuff happening in parallel but at all times a understanding where is which thing going on and have the next steps in mind (sometimes directly queued as follow up instructions). For me vibecoding is a strict speedboost and literally gamified projects I work on, and the guardrails not only in textfiles but much more in executable code (linters, tests, dependency checks, playwright, ...) as feedback loops agents can spin on on their own made it all click together to the point my main bottleneck is stuff like the Codex app itself using high CPU and memory on my local mac.
Maybe I am the only one thinking this, but there is a growing backlash against AI mindset plus a lot of people using the free chatgpt app for "googling" nowadays in parallel too, while realizing slowly that they cannot really trust it to be correct. you know what would provide an actual outstanding value to users that is even at times increasingly searched for? the original 10 blue links with zero distraction or enshittification and a hard AI content filter applied bottom up already at ingestion level, like, what the quality team of google search did for decades (filtering out low quality spammy noise). Just the thing that made the original google great essentially, be an actual useful tool, not another chat people increasingly hate. But I guess I might stand alone here with that reasoning.
to be blunt actual porn sites are hyper optimized and mostly have a polished user experience, even frontier level tech stuff going on at times. you know how onlie credit card payment came really alive? not judging, but I'd not throw corn into the edge of shitty tech by default
Still anecdotal but the exact same coding task on the exact same repo (I clone from GitHub templates for projects) worked amazingly well in December with CC/Opus, couldn’t accomplish the goal anymore end of march, with essentially identical prompts, and 4.7 was just comically useless. But even these days I tried repeatedly and 4.6 still can’t do the thing it could in December.
I get what youre trying to say but this is actually a bad picture to defend. product and engineering should go hand in hand, with one side informing the other. Engineers sctually giving a shit about a product will tell product possibilities they havent even considered, product people caring about engineering will not propose utterly stupid things. and I for one can spot when a product is well designed but poorly made, as well as when a product is perfectly crafted yet useless. the sweetspot is both. and even with the speed multiplier of AI, having a proud in the craft and being actually good in it as an engineer makes a night and day difference for the final result.
Just switched into a CPTO role in a industry I have limited domain knowledge about (gastro) and first week I build up some automated world modelling for my small company to get a handle on things:
- crawl all relevant news sites and subreddits hourly, rss fetch style, and for every new thing let a small LLM think about what X means in essence and derive insight nuggets or thoughts about "what does this mean for customer group X", super fast and cheap oneliners essentially but at times surprisingly good snippets.
- add in designated slack channels and allow direct tagging of a bot so that internal discussions inputs also go into a database with a similar LLM treatment of extracting the interesting bits and pieces
- have a hourly cronjob look at the new stuff in those databases and create proper hypothesis/assumptions for customer behavior and needs, lean startup style, or link new stuff to existing stuff and refine it.
- have another hourly cronjob go through the new/changed hypothesis + some knowledge about our product/platform and formulate feature candidates worth building that are distinct and linked by evidence
- so far its all just code + LLM with some statistics for relevance, essentially a bayesian board of ranked experiments to try next, running fully on autopilot, and quality 80/20 is fully acceptable.
- I added a straightforward kano voting UI for internal folks on the most promising candidates to get explicit human judgements factored in, again some math formulas of weighing stuff changing the scoring significantly depending on vote results, and everyone can give their input/judgement in a asynchronous way, human touch decentralized, future planned as in-product little feedback ask widgets to get thecustomer voices in it, too
- I am the only one who can pick "the next thing to build" from the prepared and ranked list without much emotion but with my own judgement so I can also pick not just the highest ranking item but for good reason a different thing slightly lower. onclick github issue is created and framed with a user sotry and the hypothesis + all the context accumulated bottom up nicely formatted as markdown
- my codex/claude code on the platform repo knows how to work with GH issues and what the processis, so I pick up the next formulated ticket and brainstorm together HOW to build the hypothesis in a semi-supervised way (vibecoding, but with precise context and me driving it with my own judgement/ideas for how to make it happen in the best/leanest way that is faithful to the hypothesis)
- shipping it closes the issue, goes back to the feature candidate, moves it into a "measure" state, and then knows which metrics to observe, aka did the feature move the right needle and how much, and after a significant enough amount of time its clear it was good/bad/neutral and I can either do a cleanup or leave it there or report success, which then updates the underlying initial assumptions with strong real world evidence up or down, which naturally reorders the feature candidates list linked to the assumptions list, changing where to go next instantly
... essentially I built my own hamster wheel, reducing the "what to build next" into something that feels scientifically enough for people to trust it, with manual intervention through voting, and a stupid simple place to just drop any observations/ideas with slack without ANY special formatting or writing requirements and have it bubble up into concrete things to build autolinked with supporting/contradicting evidence from world news etc. The actual build/execution I do as you all then with guided vibecoding, reducing the cost of build to minutes/hours instead of days/weeks, with zero meetings needed, realtime and asynchronous.
... is this what you looked for? I streamlined product to become like 2 clicks from my end and zero product people needed beyond myself, and streamlined tech to vibecoding + grounded context, going into a kind of guided shipping loop with learning loops closing automatically along the way. and when AI takes it 30 minutes to flesh out a new thing its all the time freed up so I can talk ideas or feedback stuff with colleagues while the machinery is humming. Never been happier to be a devops/fullstack guy.
That’s statistically just luck then - plenty of outages this year already in Berlin time during work hours - I do remember the forced breaks with colleagues for sure.