lol yall spending unlimited tokens out here. No worries that’ll just be another datacenter in my backyard it’s cool
4 Likes
It won’t be long until the open source models will be good enough to do everything locally instead of in the cloud with massive centers. The big companies are gonna want that anyway for security reasons.
You are speaking gibberish. How would it ‘measure everything against (my) companies specifics’? How should I tell it to do that? How would I have it ‘create a plan for any needed changes’? That is the point of the whole exercise isn’t it? It certainly sounds simple; but the actual results seem to elude me. Perhaps you are just a better developer than I am. So much for 40 years experience…
You keep using vague instructions and expect specific results. My experience with my AI is that AI is not conducive to this.
I would give a build agent several different .md files and created skills with all the specifics you are wanting in the build. I would give Agent 2 a library of screenshots of what you want it to look like. Then I’d tell agent 2 to take a series of screenshots of what agent 1 builds and compare them to the library of reference images and have agent 2 look for any discrepancies and create a written plan to fix them. Then pass that pack to agent 1. Repeat until it’s right.
I would give a build agent several different .md files and created skills with all the specifics you are wanting in the build. I would give Agent 2 a library of screenshots of what you want it to look like along with exact specifics in more .md files. Then I’d tell agent 2 to take a series of screenshots of what agent 1 builds and compare them to the library of reference images and specifics in the .md files and have agent 2 look for any discrepancies and create a written plan to fix them. Then pass that back to agent 1. Repeat until it’s right.
Will it apply that same solution to all the other 100+ views? You are looking for a solution to one problem. But these AI solutions are usually custom. If it is told to apply a style, it does so the way it has seen before. If I tell it to read our project and then create new components using the existing CSS/SCSS styling, I get results that simply are not correct. MOST importantly, the resulting code is different than all the rest of the code - littered with custom CSS in response to me asking it to do something like ‘align all margins to 17px; gutters 5px’. This means maintaining and debugging this code becomes different for different parts of the code. Nothing more fun than a project where each page has its own way of implementing the view and custom CSS per page.
Are you using Microsoft copilot?
I believe he has said before he is stuck on copilot. I will say that these corporate guidelines can still give Claude Opus a run for its money though. I went round and round all last week where I would hand it an existing pipeline and say I want you to use this as a baseline. Refer to these packages. Copy the formatting here. Only use these libraries. It would still sometimes go completely off book and do whatever it wanted.
Rather than multiple agents one checking the other though I’m using spec based development of it. So a document will be created saying do these tasks. Follow these examples. Create these tests that validate that you did the thing. That document becomes a task list. Each individual task is carried out in individual sessions. I validate the output.
Honestly it works. It’s a way to make software while in meetings, answering emails, etc and half paying attention and it can do a pretty decent job. Can it do as good as a dedicated dev focusing with no meetings and distractions? No. But who lives that life?
If he can integrate Playwright into his development environment with AI he can probably have the AI “view” renders of the pages to check against style guidelines.
1 Like
This is exactly what I was trying to say. If he is stuck using Copilot then he doesn’t stand a chance though.
The inefficient garbage code this will add to whatever it’s creating is gonna be so epic to untangle one day lol.
No worse than the garbage code interns and junior devs have created in the past but now at scale lol.
2 Likes
I’d bet if you got the output your looking for then send it through another agent get rid of any “garbage” without changing the view or functionality it’d be decent.
I doubt it. At least at this stage of ai models the agents are looking for outcome not efficiency. So your scripts and such will come out working but the way it “works” isn’t going to always be optimized for the platform . This admittedly is learned by the model from bad development from other developers code it’s seen that ignores the hardware and platform and builds apps basically just to run. This can be fine and my experience it’s just what developers especially junior developers do they make things that “work” but do they work well.. debatable. Errors, load times and memory leaks or odd artifacts are going to be the norm at this phase of AI development especially if vibe coders and folks that can talk the talk but don’t know how to review and fix issues get to push code without reviewing it proper.
Good for me though cause then when the next big AI caused security event takes place my job will be secure in the federal space for a decade based on how reactive stuff is. 
2 Likes
Going off this tangent as it’s solemnly been discussed; During my company’s acquisition from small to multi-national, another problem with AI for corporations, much like pretty much any other software license for big companies is, sometimes, you’re stuck with whatever the higher ups decided to go with, whatever the reason may be (cheap, best bang per buck, most useful, etc).
Workers can’t just choose to use another AI agent, we can’t just say, well, this agent is better at this and this other agent would be better for something else. We’re stuck with whatever we’re given, and if we’re given a crappy agent, then that will affect our work. Like, I’m glad that you, being a free agent 
can freely chose an AI at your leisure (provided you have the currency), but we simply cannot just be like, “Well, I wanna try .”
Because if it works and makes it to production, my company now has to decide if they want to keep the code, and any derivatives thereof, they’ll have to pay for the licensing, otherwise, they’ll have to completely re-write everything, which could just be a few lines, to whole files (depending on how long this flies under the radar). I could potentially be in hot water for instigating a multi-thousand/mulit-million dollar expenditure; Especially if we’re supposed to be saving money that quarter.
@TheThrowingGnome You are looking at this from the lens of a free agent. You have the freedom to choose which agent to use, you don’t have to worry about security, you don’t have to worry about it conforming to company standards (which may be it’s own issue, though out of scope here). Put yourself in our shoes. You’re collaborating with dozens/hundreds/thousands of people, all with their own idiosyncrasies, you are governed by a slew of rules you cannot change, your code needs to go through many hands (for example, as previously discussed, security). You only get one (or if you’re lucky, a small handful of) agent(s). You’ll also likely have third party oversight/audits (especially if you have gov’t contracts).
As it stands, right now, how comfortable would you feel putting your AI-derived creations on a server hosted at your house that’s open to the world? If you don’t know how to do that, you can use AI to teach you (In for a penny, in for a pound; as the saying goes). Granted, this might not resonate with you quite as much, because iirc, most of your apps are client side, but imagine if you had something with a database. Further, imagine that app required that the database retain sensitive information. Would you feel comfortable creating this app with AI, no manual input, and putting it straight on the WWW for everyone to use? What if you got feature requests? Bugs? What if you get hacked? Would you feel that the agent made the hacked data worthless or hard enough?
Bonus: Imagine you’re part of a large company (large enough that you can’t just chose whatever agent you like), imagine you’ve been given the directive to fix a bug, and it’s been suggested that it stems from someone else’s AI code. You ask the person who worked on it last for details, docs, or for a good place to start. They shrug and say, “I was just told to put this prompt in.” Now, all you have to go off is the code itself and whatever comments the AI made. Problem is, your job is the same as theirs, just put the prompt in. You have very little knowledge of this language, possibly any language for that matter. Heck, you don’t even know what this block is supposed to do and all the variables are effectively var1 , string1, array1(I know agents tend to be a bit more descriptive, but go with me). You provide the block for context and ask the AI to fix the bug, which you’d naturally describe. It comes back with something, you test it and you’re still able to reproduce. You try again, and again, and again, but still the bug persists.
In summary: AI use as a free agent is incredibly different than AI use in corporations, and hopefully, I’ve given you enough insight to walk a mile in our shoes
3 Likes
To be fair this was folks sentiment during the internet boom. Would you feel comfortable with all your records being digital and easy to lose instead of a physical copy you can touch?
Now paper copies are silly in most cases. With that said the inbetween time there was lots of abuse of the internet boom did to lack of regulation. Think of ponzi schemes like madoff where he used the magic of computers to convince people he had a great portfolio that magically printed money. (Spoiler it did not)
I imagine there will be lots of bad inbetween the final good state where AI is ubiquitous but I really do agree today right now is the worst it’s going to be and it will get better eventually. In the mean time there will be loads of pain and trash bad things produced as an outcome of AI and the people using it not knowing better.
2 Likes
This is fair. We don’t know what will happen in the next few years or decades. Personally, I would argue that the internet is worse off, antiquated client security measures or just a general lack of any security at all, en-
-ification, and sameness to name a few. However, there could very well be a future where AI is advanced enough to make the shareholders happy by employing lower or even minimum wage workers (Can’t wait to see how a certain party manages to convince people that AI is good despite the fact that it wrecks the economy while simultaneously lowering employment, but I digress).
All in all, if we’re arguing on the sole base that AI is a good technological advancement, without any external morals, ethics, or philosophy, heck yeah! AI would give anyone the opportunity to do software development. Not only that, but AI could be an useful, even essential, tool, emphasis on “tool”, that devs could add to their toolbox, much like a computer, an IDE, unit testing, or whathaveyou. However in the world we live in, as it stands right now, the effects of which you can witness in front of your vary eyes (via Andy), devs are training their replacements; people are being “given a chance to find a new place in the company” (read: laid off) by effectively using AI well, or refusing to use AI.
I wish I could believe in a world where big companies allowed devs to keep AI as a tool, and nothing more. But big companies or large corporations, especially those that are public, only care about one thing, showing a green number to their shareholders in the next three months. I genuinely wish I had Gnome’s faith that my company isn’t putting me and my fellow dev’s (devs’?) head on chopping blocks, but that simply isn’t the case
Tbh, I think if AI continues to proliferate the tech industries for long enough, and if it has a boom or a significant advancement, I think we’ll be looking at, at minimum, another great recession, both economically and job wise. Bottom line, companies do not like paying people, we cost a lot, we complain, we are lawsuit machines, we require downtime and maintenance. We are neither efficient or consistent. AI though… AI just does what you ask it to do, and that’s the dream of corporations. A mindless drone that costs very little money
1 Like
I totally get what you’re saying. I guess it’s quite the luxury to be able to use a work flow like: give an idea to Perplexity to generate market research, then give that research to Chat gpt in thinking mode to review and generate a plan to build an app, give the plan to Claude via Lovable to build the MVP. Go back and forth with Claude to set up the backend in Supabase and ensure everything is working, use Gemini to generate necessary images, send it to git hub and then pull that into Codex 5.3 to find security issues and fix bugs.
What I just described is lightyears ahead of simply fumbling around with CoPilot. I would point out that if you’re a professional developer who is hamstrung by using copilot, then your opinion that AI can’t do this or can’t do that should be taken with a grain of salt. All the different models are best for different things and to try to use only the worst one for everything is like asking a carpenter to build a house with nothing but a hammer
As far as security goes, it came out last week that Anthropic had Claude go through their whole company looking for security issues and it found over 500 security risks that were missed by human engineers.
2 Likes
I’m going to be honest, and I only speak for myself as a dev, but if that was my workflow just to get a short block of code out the door, I think I’d just find a different career. Even to just create an app for personal use, depending on how important it was to me, that still feels like a lot. Keep in mind, professional devs don’t create whole ■■■■■■■■ with AI, we’re typically working with someone else’s code (in some cases a lot of many other people’s code). I can’t imagine even having to do half of that, just to never actually touch the code. And honestly, I may not be the best dev, but I didn’t sign up just to babysit another coder (training is different, I’ll happily train a new dev).
Even further, I’m still not sure I’d be comfortable with putting putting AI generated code, hot off the press, straight onto my homelab and open it for the world to use and abuse. I’d still want to review it, which requires someone who can not only read the language, but also understand how people can abuse the system.
I’m not talking about what AI can/can’t do, in fact, I’ve even stated several times that if I was told (and shown) that AI was simply another tool in the toolbox (like it is for you, and other “freelance devs”, if you will), I would happily use it, I’m actually pretty happy with the foundation that even copilot provides as a starting place, the sourdough starter, if you will. As a tool, I’m more than happy to exploit AI for all it’s worth. But if my future is anything like Andy’s, wherein no human hands shall touch the AI generated code is a mandate, that’s a bleak future, and says a lot more about management/higher-up’s end goal is than anything else.
I spend 95% of my time on any one story going “wtf was this person thinking?” or “how tf did this even work in the first place?” than I do coding. As much as I dislike it, I would much rather do that than spend the same amount of time generating prompt after prompt just to get something what I want without ever having touched it.
To summarize and possibly clarify the last two paragraphs:
- AI as a tool (the good path/future), fantastic! I’m all down for that. Never heard of a tradesmen complaining about getting a new power tool after having used manual tools for so long (I’m sure @SlowThrow can speak to this).
- AI as a replacement, not great. How would you feel if your job was replaced by machines? How would you feel if your job was to train that machine how to do your job? How would you feel about all your time learning your specialization and all of the sudden, you’re not doing any of that. Instead, your job now is to just type prompts
In fact, let me go even further, allow me to put my shoes on you. Let’s assume that there was an engineering AI that’s as good as dev AI (at least, as good as you claim dev AI is; idk if this exists yet, but if this bubble goes on long enough, you can bet your bottom dollar it will eventually exist) As an engineer, you have a very tight margin of error (correct me if I’m wrong). What if instead of all the time you spent designing a construct or a system, you are just entering prompt after prompt, with little to no input after the fact. For now, you can validate that it won’t collapse, but even then, management might decide that the AI is trustworthy enough that you don’t need to validate it either. I suppose what I’m asking is, provided there was an engineering AI that’s as good as the dev AI, how would you feel doing our workflow for your job, knowing that your job is to train your replacement?
Tbc, I’m not talking about security issues in the AI agent, I mean the apps in which it creates
2 Likes
Idk man just this morning my security team was talking about building threat modeling apps with AI agents and my first response was “really? We cool with this?” Followed by a this is corporate mandate we don’t have a Choice. It’ll be interesting because I’m positive we will be creating unicorns every time we make a new agentic AI built app and that’s gonna get messy and non standard fast…
1 Like
I do have the ability to choose multiple models for use in Copilot. To date; I have not seen much difference in results. That is probably because I am using AI for specific tasks in existing code; not for building new apps from ground up.
1 Like
I honestly don’t trust an AI company reporting its own metrics like that any more than i trust the anti virus companies who similar things in threats found…
With AI it’s even worse because AI sometimes just makes stuff up lol how many of those threats where legitimate
3 Likes