Perhaps, but then my comment is even more true, as these people are just misinformed. To me, vibe coding is merely conversational interaction with the AI...instructional based direction, for it to write what I'm telling it to write, which of course requires a ton of review. But if I ask it to do something like add a new property to a database model, and add everything required to support it from the back end to the front (with a much more specific prompt), it can do that in one pass, without a mistake. That's impressive, and not something any coder should ignore.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Xcode 26.3 Lets AI Agents From Anthropic and OpenAI Build Apps Autonomously
- Thread starter MacRumors
- Start date
- Sort by reaction score
Non-Euclidean
macrumors 6502a
The rise of Moltbook suggests viral AI prompts may be the next big security threat
The rise of Moltbook suggests viral AI prompts may be the next big security threat
We don't need self-replicating AI models to have problems, just self-replicating prompts.
arstechnica.com
You are correct. Don't let it compact. The Claude Code documentation refers:
Claude’s context window fills up fast, and performance degrades as it fills.
Do X task, /clear, next task, etc. If the task is big, split it into multiple smaller tasks. This is the same methodology devs used long before LLMs were a thing.
You should not be keeping one conversation for a project ad aeternum. I see a lot of people do this (especially outside coding) then they complain their ChatGPT becomes "delusional".
bmustaf
macrumors 6502a
Hey, there are still people out there who shave the sheep, clean the wool, produce the thread, and manually weave the full sweater and everything else they wear and sell. And good for them, I have no need or desire to change their mind, and if they'd like to continue to self-select themselves out of the economy and production as it shifts, I mean, it's just better for the rest of us who do understand how to adopt new technology properly & productively, so I honestly am OK with them making that choice and holding that attitude.
Ref: https://en.wikipedia.org/wiki/Luddite
thefredelement
macrumors 65816
In that case, you want code that documents itself, no comments (only why not what, if needed), simple methods (no more than 1 real logic fork), easy to understand names, etc.
This way the LLM can at least ad hoc interpret your codebase instead of getting filled with markdown files, etc.
ProbablyDylan
macrumors 68040
That's brave. I never want to assume that future me has any idea what current or past me was doing.
Andreas from Zurich
macrumors member
IvyKing
macrumors 6502a
One thought going through my head was Vibe Coding would be a great way of spreading malware. OTOH, AI may be a great hope in spotting malware.The rise of Moltbook suggests viral AI prompts may be the next big security threat
The rise of Moltbook suggests viral AI prompts may be the next big security threat
We don't need self-replicating AI models to have problems, just self-replicating prompts.
right with you... been coding forever. I am so done with typing... 'vibe coding' is like my personal assistant that just types out my intentions perfectly. I am still the architect. You are correct, there are those that will be left behind.
Arislan
macrumors member
Exactly. Prototype code being used in production apps. Almost impossible to upgrade without a complete rewrite and every single point of failure multiplied due to AI meddling. I love using these things for selling ideas, but actually shipping a product and expecting to continually update it is insane.
gimmesomemo
macrumors regular
I’m sorry, I genuinely can’t follow what you’re on about. I can’t even if you intentionally quote replied me or not. If you did, it’s nonsense.
Why would non technical stakeholders be looking through code documentation? I should also clarify my anti doc statement. I meant in code docs, as in, if you need to add documentation to a function to explain its parameters or output, or explain what a class is doing, for other developers to understand - then that’s the smell I was referring to.
As uncle Bob would say *wince* the only thing worse than bad code is stale documentation. Or something along those lines
I use it to learn code or explain me why something happens in the code. At the same time seeing how good it is I do wonder if there is any sense in learning to code. AI kind of makes manual coding obsolete…
Maybe idea is to learn to code AI models🙈
But also using it extensively every day I see it disables thinking. I do have entire life before AI boom so I have reference to the reality and can verify nonsense but I do wonder if our kids raised by AI in one or another way will be able to differentiate between real and made up.
…Knowing it will not go away… I hope every day it will…
Maybe idea is to learn to code AI models🙈
But also using it extensively every day I see it disables thinking. I do have entire life before AI boom so I have reference to the reality and can verify nonsense but I do wonder if our kids raised by AI in one or another way will be able to differentiate between real and made up.
…Knowing it will not go away… I hope every day it will…
It's really giving me an advantage at the moment over people who still don't understand what it is, and how powerful it is. Thing is it is only really powerful in the hands of an experienced coder. A novice, or someone with no experience, can't use it. You need to already know what you're doing in order to direct it properly and build something real.
This is very true. I do hope though that it will help the other way around, helping novices who actually do want to learn to do it. This could be a democratizing factor in allowing people more of their own software freedom. People that felt the barrier to entry was too high now might make custom software for themselves that otherwise wouldn't exist.
That's how I want to see it anyway. The Internet itself didn't exactly pan out like we hoped...
Exactly. People don't get this. It's not perfect but neither is anything else. If you do know enough to know what to trust from it, what to question, and what to research further, it can be a huge time saver.
But the exact same thing could be said for a human resource. People can get things wrong or misunderstand or make mistakes. They sometimes even mis-remember and make things up without realizing it.
I'm also one of those skilled people (researcher, scientist, and academician) using them to help speed up my research and statistics. I spend much less time coding than I used to, about the same time looking for bugs, and my overall productivity is much higher.
I should clarify that while I spend less time personally coding, I have LLMs generate a lot more code than I used to generate. There are things that used to take me a couple hours or days or weeks to code -- all of this is generally bespoke processing and analyses of my data that would only ever be run once -- that I can do in minutes to hours. It's remarkable. There is some back and forth sometimes with the LLM to get it to do what I want, but the whole process is easily 100X faster than it used to be (and I wasn't slow before).
For my dissertation (many years ago now) I generated thousands of lines of code over a period of months. I could recreate the same code now within a few hours. Everything would be commented better and follow better coding practices (I'm not a programmer but learned coding on my own during graduate school).
While I like coding, this means I can spend more time with the science and less time making the code.
The trick is always knowing what the output should be (which I do, because I've been in my career for years) and what to ask the LLMs to do.
cjsuk
macrumors 68020
I'm actually directly responsible and experienced at cleaning up the mess left behind. The only potential this has is disaster. I'll take empirical evidence over faith thanks.
The problem is it gives people agency who have no agency. They are still idiots, just faster more dangerous ones.
cjsuk
macrumors 68020
Nope. This is not what is happening. You put it in the hands of skilled people and they stop thinking about how to solve problems and just solve bits of them and stick the bits together. What you end up with is scraps of badly thought out crap, some of which is entirely incorrect, which can't end up facing a customer because it'll cause reputational damage. You want to see some of the garbage really good people are churning out. And it's mostly because they are good at creating stuff but have no idea of what is conceptually "good". It literally enables the worst behaviour in the best people and takes away their higher level organisational and conceptual modelling skills.
As for corporations forcing you to use it, yeah they are all over it. They have dashboards that tell they have AI adoption for their marketing packs. But the average corporate monkey runs out of tokens after 2-3 days and they aren't paying for any more. That's realistically what is happening.
9 months down the line when the little token ring (tee hee) network that works moving cash between the tech companies go down the crapper those tokens are going to get rather pricey. At which point it'll be seen as an unsustainable operating cost and suddenly everything will revert to where it was before, quietly without fanfare.
But that's not the problem...
The big problem is most of the problems people solve with LLMs don't need to be solved with LLMs. They are just pasting over crappy engineering decisions, bad processes, tools and incompetence and people don't want to deal with it because it means they will be have to be accountable for screwing up in the first place.
cjsuk
macrumors 68020
Yeah and no. I've worked with academia a lot.
I almost guarantee you are using the wrong hammer to solve that problem without even looking.
Most of the time I see academics thrown in the deep end with scipy/numpy and matplotlib all wrapped up in Jupyter notebooks. It's like being thrown a bag of wonky tools and expected to throw it all together and produce something. Everything is knee deep in mutable state with no idempotency, nothing is shared in a working state and the outcome is a reproducibility nightmare.
As mentioned above, LLMs just paste over the problem that our tools are made of ****. Another layer. That is all.
We need better tools not LLMs.
My favourite analogy is really the front and LOB business app developers. I will watch people do a 3 month project for a data collection forms interface in React on microservices which could have been done in 30 minutes on Microsoft Access back in the 90s. Surely we can do better?
Abydos
macrumors 6502a
my experience was with neural nets searching data sets for correlations for further study, and pseudo deterministic models to find optimal sample locations (to reduce model uncertainty - or error - depending on your paradigm ) not sure an llm would have helped or is a viable tool