Type here.
7 Likes
Well quite.
We’ve all run into the thing where we correct the AI and then it says “actually yes, I was wrong. Thanks for correcting me. You’re instinct was good”.
It doesn’t exactly fill you with confidence, especially when people are starting to rely on AI for all kinds of day-to-day life stuff and even asking it for medical advice.
6 Likes
I’ve been using Cursor extensively to build small utilities for myself, primarily for work, and at that scale, it has surprised me by its effectiveness. There are now a number of automations I have going that range from eliminating drudge work to actually implementing some complex algorithms for design (I work in 3D modeling for product design…here I am not talking about fancy Pinterest image generation, but rather more geometry-focused applications). Pretty much everything I make is for myself, and the usefulness relies on my domain expertise and expertise in coding.
For my use case, it seems to have value…I am pretty ambivalent about it all for many reasons already mentioned - energy costs; the fact that users are being shielded from the true costs to get them addicted; IP theft - and so am waiting for the other shoe to drop. And I would absolutely struggle to try to use it for building something more durable for broad deployment.
2 Likes
They are pretty impressive. Well crafted prompts explicitly describing every requirement and sharing the context (CLI-tool) let me create big parts for the existing system in seconds. And it is almost always right on the first try. It can copy the organizational code style and does not import new libs or anything. It even reuses the components I have created before. Its refactoring and explaining skills are of better quality than most humans. I feel that at least Gemini can understand even big code bases’ logic pretty well at least when guided by prompt engineering.
Yes, I do work 2 hours a week now and can barely even bother reading the generated code any more. Never heard of any professional coder saying how their work efficiency’s 10x’d or even doubled, but if everyone were really just generating and validating output 8 hours a day, it would more likely 50x. People like to treat LLMs as mere code tools but they can do so much more with their insane generalizability. Planning, communication, specification, documentation, research etc. Hell, I even ask ChatGPT for chord progressions.
It is a weird direction we are heading. Problem-solving skill degradation is unavoidable and the programming job market is fucked hard. Security, ethical, privacy concerns. I do wish they would bump the prices, but surely organizations will have money to keep on using those as the gains are unbelievable, workflows are just being figured out and the progress during couple of years has been crazy. It makes my skin crawl just thinking about being an AI janitor for the rest of my career, generating away, at least 100k LoC a day boy! But that’s not even the endgame. They are building autonomous multi-agent (fine-tuned models for different tasks) platforms where human feedback required is minimal and the models discuss with each other. What then when no one even wants to validate the output any more and it’s just how it is.
Well… ofc it is not just all so bleak and I use the tools myself, but can’t help but wonder what the future will bring.
3 Likes
I prefer it’s more like having a college intern that has read every book and technical document ever written. … who is also tripping on lsd
4 Likes
I think vibe coding is fine if you’re learning and understanding from what you’re creating, but if you’re just prompting to create apps and throwing it out in the mix of everything else, I can’t see how that’s beneficial
I have been coding for a while now (early 80s for fun, ‘96 to present professionally), and have recently pushed my org through their AI enablement project, and … yeah a lot of sentiments in this thread are spot on.
It’s great for low risk apps but for safety or risk critical systems it’s pretty dicey. You have to force it to adopt testing and design best practices (TDD, CUPID) and then thoroughly scour the code.
Context is king: the more domain-centric information you can give it, the more grounded its output will be. But the more context you give it, the fuller its context window becomes, resulting in hallucination. Because of this, the transition to markdown as the lingua franca for knowledge is inevitable.
Juniors are getting screwed; the industry foolishly thinks that LLM agents will replace the need for new hires, when it fact it’s just setting us up for collapse on ten years when all the seniors are retiring and there are no apprenticed juniors around to pick up the work.
Token pricing bait and switch is inevitable. We just saw this with Copilot and we will probably see this with Anthropic after its IPO.
For music apps I love it. Shameless promotion for my open source, Claude powered studio knowledge management tool, Sonolex! Sonolex — Know Your Studio I use it everyday; scratching your own itch is a wonderful use for this sort of technology.
6 Likes
Type here.
10 Likes
I feel like a lot of people have a black and white relation to LLM’s. AI is a tool, and just like any other tool, you need to learn how to use it properly to make good stuff. For this generation of coding LLM’s, I believe it’s very much about managing and optimizing agent context window (”context engineering”), often outsourcing smaller tasks to individual agents, while having a stronger model keeping track of foundational coding principles, project principles etcetera (what many call a harness). This is what weeds out the Shopify vibe coders from the ones who can build working systems before they release anything, even without a coding background.
3 Likes
Good to see reasonable and nuanced discussion. I’m honestly split on AI ethically. I have ‘vibed coded’ one released app for a very niche purpose to address a specific need, and it’s being actively used by far more people than I expected. I can and do code stuff myself, ever since learning BASIC and then Pascal in school, but never stuff on this level.
I have also vibe-coded my own midi sequencer that has slowly morphed into a semi-DAW, mostly with specific things to make my productions and performances better, that other DAWs lack. Honestly I didn’t expect it to work, and then didn’t expect it to work well, but now, it’s definitely in contention to displace Ableton for my needs. Not sure I will bother to release it as it’s highly targeted at my workflow and that of my other bandmate.
I feel this is where AI is best targeted: at bespoke targeted tools that just would not get made previously. If this could also be harnessed in the war against enshittification that would also be great: use the tools of the techbros against them.
6 Likes
This is a perfectly reasonable argument about AI being another tool and having focus to use it properly, but the whole argument people are using is that you need experienced people to understand and verify the output.
The current “senior” engineers that are now the gatekeepers of output quality all learned how to make things the old-fashioned way through many years of study and real-world experience of seeing what worked and what didn’t, not just in terms of actual code but in terms of process, delivery, testing, QA etc.
If people don’t get that knowledge from years of hard-earned experience, who are going to be the future experts to verify the output? At some point, the code that trains the models in the future will be the code that was written by the models today.
Do we want future generations of doctors, engineers, coders etc to be using AI-generated “knowledge” to be the basis of their work? It’s a circular logic.
5 Likes
Why would you definitely need that though? This is not an issue if you know your limits, at the current generation of LLM’s I wouldn’t try to make my own DAW or something ridiculous like my own antivirus software.
But what if someone utilizes their deep knowledge in synthesis or electronics to make their own dope sounding synth engines wrapped in an existing framework? Like using Max, but instead packaging it into a VST using a tried and tested system with separations of concerns. Conversely, there are many great coders who are absolute shit at synthesis.
That’s neither black nor white to me. Coders have been using frameworks and modules to make romplers and stuff for years. The only difference is that it has become even more accessible know. This means that those who have mastered their craft can use that to their advantage in new ways. This is great news. And to make it, you have to earn your reputation by being humble and able to control what you build.
1 Like
I think you’re missing my point.
Who is going “master their craft” or have “deep knowledge” in the future if everybody takes a shortcut by using AI? At the moment, those with mastery in any given field got it the old-fashioned way. If we don’t have a pipeline of people learning things properly, we won’t have people with “mastery” in the future to actually train any AI or know how to do anything or verify the output, we’ll just have a generation that are reliant on AI and nobody with actual mastery of anything. What happens when the masters die out?!
The current AI is trained on previous years worth of real human knowledge. Where is that knowledge coming from if nobody is learning anything properly in the first place?!
I can’t believe people aren’t more wary of this, as if we aren’t already excessively reliant on technology and the tech bros that control it. We know from recent experience that we can’t trust these people not to just do whatever lines their own pockets or change the goalposts whenever it suits them.
12 Likes
And you’re missing mine, respectfully. I really disagree with what you said here. On the contrary, I think mastering a craft is going to be extremely important, both for coders and for musicians. Who’s going to buy a vibe coded synth from someone with no synthesis experience or ability to curate interesting sounds? On the other hand, if someone like Howard Scarr built his own synth using LLM’s and a tried and tested library of DSP modules to make sounds that make us feel things deeply, I would personally be all over it.
Instead of rebuilding rudimentary things, both musicians and coders can focus on doing entirely new things, creating novel experience instead of reinventing antialiasing for the millionth time.
1 Like
I apologise if I’ve missed your point, I’m not trying to be argumentative for the sake of it.
There are already loads of people rushing out new synths and music tools, they’re all over KVR. Many of those people, by their own admission, have absolutely burned through AI tokens by aggressively vibe-coding for a few weeks with no prior knowledge of software engineering or DSP. They aren’t hiding their lack of domain expertise at all.
And people are buying that stuff.
I’m not questioning what’s possible, I’m questioning exactly HOW people are going to become experts in a given subject in the future. You can’t “fast track” years worth of learning and hard experience by having access to the worlds largest (and by no means entirely accurate) encyclopaedia.
Being able to do something quickly doesn’t make you an expert in it.
I respect your position that you’d rather give your money to somebody who has experience and reputation (I feel the same way) but plenty of people couldn’t care less where something came from, who made it, or how.
5 Likes
Good discussion around the mastering the craft topic.
To be good at using AI in a hardcore sense (ie not music apps— trading systems, etc) you need to have good design sensibility. To have good design sensibility, you have to have intuition into software engineering (“solving problems created by solving problems with computers”). And to have intuition into software engineering, you have to have intuition onto computer science (“solving problems with computers”).
This is why the industry is pivoting hard towards senior devs to do AI work, because they have the intuition needed to guide the AI towards a good design (which it kind of sucks at, unless it’s a traditional 3 tier web site thing).
You can’t fast track that intuition, as @tallrobphilp says above. I once had the absolute pleasure of spending about 400 hours pair programming with Matt Godbolt (of Compiler Explorer fame), and I learned more on those 400 hours than I had on the previous 4,000. Seeing the blind alleys he went down, observing his sense of taste , learning his testing katas and UNIX she’ll command line tricks.
With agentic AI it seems this sort of medieval guild style apprenticeship is mostly dead, and I have no idea how the younger generation of programmers are going to pick it up, especially when the temptation will be to let the AI do it for you. Yikes.
3 Likes
You could make the same argument about books (or “the internet”). “Who’s going to remember things now that all the things we used to remember are now in books (the internet) ?” . The books (the internet) are going to remember.
1 Like
The skills needed for vibe-coding will most naturally reside, right now, as the things that experienced coders picked up over the years.
Anyone who didn’t come up through that route will find vibe-coding a ‘good’ robust app with a long life, harder … not necessarily impossible … but they’ll need to to learn to develop an audit trail for what they’ve decided (ADR) write tests that explain the objective of the code (TDD/BDD) avoid duplication (DRY), don’t start to build anything you don’t need yet (YAGNI) isolate speculative/spike/incomplete feature code (branches/feature flags) … and many more I just apply all the time without thinking about it when vibe-coding.
The skills needed are now just the higher level of what the whole job used to be.
4 Likes
…yep and nope…
i’m the artist…i’m the producer…i have final decision…always.
but an audioengineer at my side i can talk to, which is no living person but an digital assistant will be just the next step in progress…and progress is the only constant thing in the whole universe…