How widespread is AI usage in actual modern commercial frontend-development and in what way? I'm a team lead and I can't find a place for it in my team.I don't mean autosuggest in Webstorm, Google search overviews, or something similar that everybody uses. Something more akin to "create proper context containing the way you code, feed Figma project to an AI and ask it to make a React app, tweak it a little and send it to QA". Bosses push our team to use AI to that degree, but honestly I have no idea how. Feels like they want us to shit out "business optimisation" out of nowhere because they read a retarded article by some easily impressed journo.Creating entire commersial project feels dangerous, like some shit can slip on reviews or we eventually lose our guard, become lazy and amount of AI mistakes piles up. And I have no desire to spend time making AI understand what I want. Besides, in our company it is forbidden for us to use non-internal models, and ours aren't good and it takes a lot of time for them to generate something.Or am I just backwards? How do processes work in your team? The AI thread seems to be full of vibecoding normies, so I desided to make a separate thread.
>>109775713here's a bejitabro that broke containmentmods wtf?
bro typed a whole ass essay
>>109775713I find it strange that every time I hear about AI actually being used by employedfags there always seems to be a mandate by bosses/the company.
>>109775713I'll set up your entire stack + training for 8k via remote, 12k in person.White male with average dick here, btw.
>>109775779>>109775713Because cheap brown workforce + Ai works 8 out of 10 times short term... And that's enough for an average SOHO Ceo
>>109775713Use of AI is inversely proportional to the maturity, experience, skill and knowledge of the staff.A lot of new staff in our company are essentially illiterate and can’t spell. They are heavy users of grammarly, AI, and whatnot to generate verbiage.The same is true for developers. If you can’t even type, AI is your ticket to generate reams of shit that looks superficially acceptable to managers and other retard-level positions.
>some shit can slip on reviewsJust like with humans. You should treat is as if it was really good fuzzy autocomplete. Like autocomplete AI gets shit wrong sometimes. If your autocomplete makes you write shit, stop using it or try to tweak the settings/prompt. If autocomplete makes you lazy who gives a fuck.>in our company it is forbidden for us to use non-internal modelsWith bad models you are wasting your time. Again it's like using autocomplete that is slow and shit. Nobody would use it.
The real answer depends on your team's mandate.>Better to be first than rightThey're just going to push and push until it works. It can be duct tape and popsicle sticks and until something fucks up, it'll be good enough. But God help you the longer you go gluing more and more popsicle sticks without a problem or a refactor.> More mature teamsIt will be some form of AI-DLC.The team will go over the requirements, whiteboard that shit, and type up a preliminary plan file in markdown, and feed that into a prompt asking for a plan. Then you take that plan back and have the team try and shoot holes in it. Is it breaking functions into units of work? What do the unit tests look like, because LLMs can be really stupid and pedantic about coverage. What about integration testing?Once there, you run the prompt, requiring all tests pass for the LLM to be done. If things don't work out, stop, go back and modify the plan, start fresh.The goal here is to really automate keyboard time, not whiteboard time.This all gets SIGNIFICANTLY more expensive and time consuming on brownfield projects, and your productivity gains over doing it by hand won't be nearly as attractive. Also, some places will mandate 1 day a week no LLMs allowed, so that their d evs don't forget stdlibs and shit, but this is increasingly rare, because Economic Pressures™
>>109775713It's grim on my team.We were pushed to find a use case for it" so people use it for everything.People make mock designs using AI, then tell the AI to program it, then tell the AI to check the work, then tell the AI to test and review it. Messages on slack are written using AI. We have session where people "share knowledge", but it's just Indians reading off of slides created by AI. I asked someone about a problem once, and as I was talking, they were inputting my question into AI.They barely think about stuff anymore.The product quality is shit now btw. Defect rates are 2-5x what they were a year ago (around 50 defects per month to around 160-290).
>>109775713You'll want to cut your lines. I'm approaching normal. I'm at terminal....
>>109775713The killer app so far has been codeium/windsurf single line autocomplete in vim. Anything greater and ai spams you out of saliency of the system. Basically a new code base every week.Given the "3 90%s" of software dev, ergo if it is estimated to take a month to coomplete it takes a month to boilerplate, a month catch edge cases and a month to nit the complex emergent bugs (non deterministic, nat hole punchung, memory leaks, etc) AI turns it to 1 day to make a boilerplate, 1 day to fix edge cases but than 3 months to understand the system enough to iron it out
>>109775713>Besides, in our company it is forbidden for us to use non-internal models, and ours aren't good and it takes a lot of time for them to generate something.This is probably most of your problem. The state of the art models are really really good at frontend.
>>109776031>>109776105>>109776526>>109776560OK, then how does your workday look like? Lets assume the comany ordered you to create a new microfrontend app (a new page) within a bigger app. Your actions? When would you use AI and when code yourself?
