Your App Subscription Is Now My Weekend Project
61 points by anup
61 points by anup
For those who vibecode and have a better idea of the costs: is it cheaper to vibecode a markdown editor than to just pay the one-time 15$ license fee of Typora? Ignoring the time spent by the person making it.
Either way, it's sad that money that would normally go to smaller companies instead go to big tech.
Either way, it's sad that money that would normally go to smaller companies instead go to big tech.
This, 100%. Saving pennies building a crappy product, putting small mom & pop shops out of work in the mean time.
I could buy a cake from a mom & pop shop but often I bake it myself using ingredients from a supermarket (giving money to a mega corp instead of a small shop). Should I stop?
But this isn’t like “baking a cake yourself” it’s like “getting a cake Doordash’d”.
More corporations involved, more exploitation of workers, for a worse result slightly faster.
Oftentimes the result is globally worse, but locally better - it has exactly the features you care about but not the nice overall polish of an app someone has been sweating over for years
You are assuming people are actually paying for stuff like that. I see this more for people like me who usually don't.
Are you implying you get "free" access to LLMs? Either way, your time has value as well.
Yes. I've never once paid for an LLM tool. I assume they'll stop giving it away eventually...
In my own case I'm not implying anything, if this works with local Ollama I'd probably try it for stuff I don't want to build myself.
I've had GitHub Copilot free for over 2 years now because I had an open source project with enough stars or what ever. Now I have Gemini Pro "free" for a year when I bought my new Pixel phone, I didn't even realize it was available free until after I bought the phone.
This answer is interesting & relevant now and will become more interesting later once the subsidies stop.
I don't think assuming that LLM prices right now are subsidized and will go up in the future is a safe bet. It might hold true, but given the constant theme of optimizations and labs figuring out how to speed things up I would personally bet the other way.
This sounds wishful. Once a corp had determined the market will accept a price point they rarely, if ever, lower that price point unless forced to due to regulation or competition. Neither of which appears to be on the horizon for the big players.
The big players are in competition with each other?
That hasn't happened for cloud services. Over time they have trended more expensive, despite more competition and more powerful hardware.
When you have a couple of players controlling a market, you don’t have competion, you have a cartel.
The subscriptions of Cursor Ultra, Codex Pro and Claude Code Max all cost $200 / month today.
Thankfully, there is a lot of competition in the foreseeable future and inference appears to be quite profitable.
https://martinalderson.com/posts/are-openai-and-anthropic-really-losing-money-on-inference/
It is entirely possible to use current open models to get very close to state of the art results, so the answer will not vary much.
On typical consumer hardware? My impression was that open weight models are close, but still behind and usually require GPUs with 16-32+ GB of VRAM which are - to a typical user - prohibitively expensive. Has this changed recently?
The best open weight models - the ones that are competitive with the best closed models from ~6 months ago, models like Kimi K2 and GLM 4.7 - need more like 600GB+ of VRAM, maybe even 1TB.
You'll need to spend at least $20,000 (probably a lot more) on hardware to run them at home - $20,000 buys you two Mac Studio M4 Max each with 512GB of GPU-accessible RAM.
No, but you have hosting providers running inference on open models at profit right now, so there is no subsidy to speak of.
FWIW the SotA comparable open models require far more in excess of 32GB, sometimes as much as a terabyte of memory.
I think the open-weight models that are close to the proprietary ones are at least in the 256GiB RAM + 32GiB VRAM range (or even higher), which is indeed expensive on the consumer side, and you need to run it in the talking-to-itself mode raising performance needs higher (but given that independent hosters don't seem to have a path to a monopoly, it is posible that hosted execution of those models is already done at profit).
For mere 32GiB VRAM for the entire model, and direct generation mode, one option for a desktop GPU is apparently a laptop CPU. Although there is a comparative loss of performance and all, but on a used not-that-fresh-generation mobile Ryzen I run models up to 32B parameters and for some things it seems to be fine. But yeah, fully coding an app will be slow in such a setup, and the first attempt probably won't work, so even if it can fix what it writes the time needed exceeds my patience.
If your time is worth nothing, then for sure it may be cheaper.
One fun thing about coding with agents is that for small well defined tasks, you can have them work on things while you go do literally anything else. I’ve had Claude work on things while I’m sleeping.
I'm always worried that it'll either go way off track, or stop five minutes in and "waste" the remaining off-time. How do you compensate for this?
If it goes off track, I literally didn't waste any time on it, so I can throw it away and try again.
Sort of same with stopping. I have experimented with saying things like "i am about to go to bed. keep working until you're done, and use your best judgement, as you will not be able to ask me questions about what you've done. After you're done, if you've made decisions you're unsure about, make sure to describe them in the summary" and I'm not sure if it helps or not, I haven't done this in a little while and so I can't say it's actually effective or not.
This works pretty well for me too. On the other end of the spectrum I also use a lot of “do not make any changes until I say it’s ok, just tell me your thoughts” and then go back and forth a bit before “ok this sounds good, i’m going to sleep go ahead and implement”
For instance using the Ralph Wiggum plugin, essentially a loop with an escape hatch and a "success condition".
I build a lot of my own tools now and for my the freedom to live with my own bugs and bugfixes is great. It does what I need and doesn’t break because someone else needed something specific that I didn’t want.
But that's always been the case. I've written plenty of tools that have existed when I wrote mine. Or is it now it appears to be cheaper to do that?
But that's always been the case.
From my experience it has been fun to build tools, but the maintenance burden was so high that it was very rare for it to be worthwhile doing it unless your problem was truly not solved.
now it appears to be cheaper to do that
Cost of maintenance went to fractions of what it was before.
is it cheaper to vibecode a markdown editor than to just pay the one-time 15$ license fee of Typora
If you're already paying for the $20/mo subscription for any of the major providers, the delta this will make is negligible, so I'd say it is cheaper. What that means for the future of software business models, I shudder to imagine.
What that means for the future of software business models, I shudder to imagine.
Probably very little. You can cook meals yourself yet restaurants still exist.
Considering Typora is a one-time payment, I also don’t think the financial savings matter that much in this case.
To me, it’s a little different for subscription-based apps, however. For some that I use, the recurring subscription fee doesn’t really feel justified for the value that’s being provided (beyond the initial version of the app itself). If it wasn’t for work, I wouldn’t pay the subscription in these cases and either resort to a free alternative – or maybe look into such vibe-coding approaches as outlined in the article.
A lot of people doing vibecoding have a subscription like Claude Max, so the marginal cost in $ is pretty small. That said, if I were to pick the 20% of Typora features that seemed most relevant to me, I think I could recreate it for ~$5 of tokens (but several hours of work.)
This becomes a standard build vs. buy situation: you get more customizability in exchange for more maintenance work.
Wouldn't you vibecode it on a subscription you already pay for? The marginal cost of each app is effectively zero at that point (ignoring time).
I think time is not particularly bad either, since as others have mentioned, you can have the agent run in the background and occasionally check in on it rather than actively managing it through the entire process.
Typora is a one-time license of USD $15. (money well spent, too).
Once again the value of vibecoding is not compared to the value of not-vibecoding.
For one-time licenses it makes no sense to try to vibecode an alternative.
For subscriptions with no server-side value-added then it does, but only because the subscription was out-of-place already.
Now if besides the vibecoded app you also need a long-running server with extra monthly costs and maintenance burden then everything is questionable.
I would not trust my vibecoding enough to make these into products. If something goes wrong, I don’t know how to fix it. Maybe my LLM friends can, but I don’t know. But vibecoding is 100% viable for personal stuff like this: we now have apps on demand.
As a generally anti-vibecoding/LLM person, I kinda have to agree here. The difference is that I enjoy solving those problems on my own. At least usually. Development on Windows is often bad enough that I might be tempted.
I'm still much happier if there's a good FLOSS application that I can use, thank the authors, and maybe collaborate to improve.
I especially like to vibecode when it's something I have no desire to learn at all. Like some PowerShell script for example. I don't want to know PowerShell and I don't want to internalize its syntax. I much rather ask an AI to write it for me, so I can spend my time on what I actually care about and want to learn
Is this a common occurrence? You don't use PowerShell enough to learn it, but use it enough to vibe code it? I use Linux/Mac-OS X exclusively, yet I haven't bothered with learning to program in shell---I just use something I know (you know, a real programming language <joking! I'm joking!>)
but use it enough to vibe code it
If your task is simple enough that you can talk free version of whatever service into vibecoding it, where would the lower limit come from? I guess in general the idea is that PowerShell is not the only thing in the category of «not used enough to prioritise above better things to learn»
No, it's only once in a while that I need to make PowerShell scripts to run in a CI/CD pipeline on Windows Server. So "vibecoding" is perhaps a bit excessive to call it since it's quite infrequently. And it was just an example of an area I'm not particularly interested in to bother learning. I mean, I would rather spend brain power on learning Elixir for example, rather than PowerShell.
There are other similar tasks as well, that are either repetitive or boring, that I also like to use AI to solve for me.
I think somebody already did that and I use it as my daily driver: https://github.com/kitlangton/Hex
It's worth pointing out that some services, like Pushover, can't be replaced by a vibecoded project because of the distributed trust required. This is not quite the same as the issue of trust that's already brought up, because it is about the trust invested in the complete end-to-end infrastructure rather than the trust in the app as a hermetic package. Incidentally, Pushover charges a one-time fee rather than a subscription fee, but I'd hesitate to conclude that infrastructure is cheaper than personal trust.
I feel like this is a somewhat crazy valuation of my own time. My experience with 3rd party Mac applications is typically they provide way more value than the dollar amount associated with them. I use Dash around 200 times a week, $25 bucks a year to keep all those doc sets updated and Dash running with the yearly macOS releases is a simple decision for me. SoundSource with having to make constant video calls and not have to keep up with the millions of API changed in the macOS sound ecosystem is a no-brainer.
LLMs have no memory of why they built stuff in this way, which means whenever you encounter a problem and ask them to solve that problem, they're coming at it from a completely fresh perspective. I've seen this in my own testing a thousand times, where in order to fix a small bug in an existing code base they suggest refactoring a massive percentage of the total code to effectively make their solution work. Just because they can make this app once for you doesn't mean it'll work in six months and it also doesn't mean they can fix it when it breaks.
My day is spent fighting with technical problems in the space that pays my bills. The last thing on Earth I want to do is start fighting with a totally different problem space just to save a completely trivial amount of money compared to the value provided. If you want to make your own applications and just deal with the frustration when they break? Go with my blessing. For me, half of the appeal of the Apple ecosystem IS the amazing third party community of developers. I would never sacrifice that for such small amounts of money.
"Reel" looks interesting, would like to see some visual examples of the app before I use it though.
One can't vibecode a screenshot.
You can have LLMs take screenshots, via tool calls.
Do you know if it's possible for the AI to run the app through an MCP like Playwright and take screenshots and post them in a PR?
All of these $10/month apps are suddenly a weekend project for me.
I would rather spend the time contributing to a FOSS alternative for that software than spend time vibecoding something that I would only use myself. I would both learn something new and contribute to a project that will be of a higher quality. Sure, maybe like the author I don't know Swift and MacOS programming. But fixing some docs or some bugs here and there would be a perfect excuse to learn. And I wouldn't need to spend money on highly subsidised rent-seeking LLM subscriptions either.
Maybe your app subscription should be our weekend project, and learning experience.