Not quite the direction I expected that to take, but I do agree with the broad strokes. LLMs lowering the cost of code isn't exactly solving software development, and is definitely not solving product development.
The only people with higher expectations of the AI boom than the optimists are the pessimists. The forecasting I've seen [0] is that AI will be in a position to vibe-code things like Photoshop with some human assistance by around 2027-2030 if the current trends continue. Maybe fully autonomously in the mid 2030s depending on how many human-hours a basic clone of Photoshop takes to build.
Same question I ask, it’s free money, why isn’t a million people asking Claude to build it?
My guess is the lack of training data and understanding of the problems it solves, but AI was supposed to fix this already?
Happy I can use Claude for more trivial things, but if the AI labs wanted an actual benchmark, Photoshop or Davinci Resolve would be it.
The biggest reason they don't exist is that you can buy them today. Why pay thousands of dollars and spend hundreds of hours to vibe code a photoshop when you can use the real, existing photoshop or one of its competitors immediately for a fraction of the price.
People are not upset because "Level 1" was taken away from them.
For artists and creatives, _all levels_ have been taken away. And for most, their very basic needs (i.e. income) are being taken away.
And all that, just so we can get past "level 1"? Not even "level 4"? What kind of trade off is that??
We're supposed to be engineers. We spend days/weeks discussing how a table should be designed so it doesn't bite us in 6 months time. Yet here we are discarding almost _everything_ just to make "level 1" a little bit more convenient.
His argument is the equivalent of saying in ~1910 when the 1st mass produced Ford car came out "Where are the carbon brakes? Where is the hybrid motor system? Where is the ABS? Where are the rear parking cameras"?
I think one of the major reasons why it’s not here is because most AI tools are great for getting a prototype built up but undertaking a program like Photoshop which we can assume has a couple millions of lines of code is actually not easily replicable by vibe coding.
Can it be done? Probably. But the token cost to do so would be astronomically expensive. You still need a human (read prompt engineer) to steer the AI. Even small ideas I’ve thrown at it to test viability is often plagued by code that just simply doesn’t work. I am constantly having to send it back and say “this doesn’t work” and it will eventually figure it out but that is more wasted tokens just to get something that doesn’t throw an error during the build process.
One-off specialty programs are absolutely going to feel the heat. Take a multi-tone generation application (think Motorola pagers). You can now ask an AI to create that program for you, complete with tone generation and .wav recordings to use later.
Large programs are relatively safe for the moment just because of scale. But any small application that is usually behind some sort of paywall or license is absolutely going to be threatened since it’s no longer difficult to throw together a program in an evening and have 100% of the features you want and none of the features you didn’t need or want.
Second sentence form the article, "If vibecoding is what people say it is, the world should be drowning in vibecoded artifacts right now." My answer: it is. Maybe we don't have a bunch of photo editing apps but it sure seems like we have a lot of vibe coded commercial and non-commercial projects being created. If i have to see one more vibe coded agent harness I might actually lose it. And I wouldn't call it all slop
If Photoshop can be vibe coded in a couple of weeks, that's superintelligence.
I think of "tokens" as units of intelligence spent on a problem.
Producing software like Photoshop, Linux, or Excel literally cost billions of tokens (human intelligence back then, and now both human and artificial intelligence combined). A lot of deep intellectual problems were solved along the way, and the token cost of solving them accumulated. The true cost of Photoshop today should be measured in billions, maybe trillions of tokens.
AI can definitely generate some tokens, and it has accelerated everything involving intelligence. But it still can't produce 100 billion tokens worth of intelligence in a week. When that day comes, that's superintelligence. We'll go to Mars, we'll make fusion work. That's the singularity.
> Level 3 is the deciding. What to build at all [...] AI lowered the cost of Level 1. It did not touch Levels 2 or 3.
Actually where I get the most impressed working with AI is kind of at Level 3, where I ask for a feature and AI will suggest going further with it, or doing it in another, sometimes better, way.
Sure, I'm still the one _deciding_, but AI suggested things that I would never have thought about.
I don't think that's the direction this is going to take.
Replacing a mature app with an incredibly wide audience and a million use cases may or may not be possible, but what's actually happening is that people are making an app that does exactly what they want, and using it to solve their own needs.
Previously you might use Excel to take raw data from various places and analyze it, create charts, reports, extract findings. Now you can have AI write a script to produce the exactly report you want from the original data. No Excel required.
Photoshop is a good example for that, actually. How many people are just using ChatGPT or Gemini to get the image edits they want instead of reaching for Photoshop? I don't know if this is showing up in Adobe's subscription numbers, yet, but I expect it will eventually.
Yeah, the pain/reward ratio is against vibecoded replacements for mature tools. Piracy is cheaper than tokens.
But over the next 5 years I expect a growth in Blender-like open-source projects aiming to take on the big closed-source elephants. Code is cheaper now. The main downside of LLM coding, unmaintainable spaghetti code, can be mitigated effectively with discipline and coordination.
You still need maintainers to uphold contribution standards, but people will throw tokens at you. A small, disciplined team can go a long way, make a decent enough product, and then attract the institutional money (like Blender did) and hit that growth curve where everyone rallies you and you've won.
Lots of companies would have a vested interest in reducing these dependencies to Adobe et al., or have a more customizable product. Competitive professional tools, more like Blender and less like GIMP, but in other areas, like DAWs, CADs, and others.
100%. This is what’s happening with agentic coding. We develop individual solutions tailored to our needs. We don’t need one-stop shops anymore.
For instance I have created a snippet manager, because I didn’t like the one in Raycast. Or I have created a journal viewer because the daily notes in Obsidian are not easily viewed across days. And so I have dozens of smaller solutions that are tailored to my needs. And I am guessing lots of vibe coders do the same.
That sounds correct to me. I vibe-coded an app that creates a stereo pair from .mpo files for viewing on an antique stereoscope [1]. Is there anything more niche than that?
This is such a bizarre misrepresentation of the author's point that I have to assume in good faith that it's based on a misreading of the post. Otherwise, you are unintentionally proving the author's point by describing (now) trivial uses of AI as some sort of counterpoint, or simply making an unrelated point that doesn't have anything to do with the author's main argument.
The author's point is not specifically about Photoshop. It's right there in the second paragraph:
>Where is the vibecoded Photoshop. The vibecoded Excel. The vibecoded Maya. The vibecoded Blender. The vibecoded compiler that compiles itself. The vibecoded database, the vibecoded OS, the vibecoded anything-that-requires-architectural-judgment-to-hold-together
What do all these examples have in common, besides requiring "architectural judgment-to-hold-together"?
They are software created by mid-large companies or organizations with a large of numbers of contributors. It's not about Photoshop, it's about developing complex software with high quality control and a continuous mature and continuous release cycle to a demanding set of hundreds/thousands of users. So to take that it in and then say, "Well, I don't think AI going to be used fo build that stuff, it's going to be directly used by non-technical end users to bypass existing more complex legacy software like Excel," is only proving the author's point.
Cause at the end of the day, what are all these companies "token-maxxing" actually building and selling to their customers? That's the larger point the author is making. Just as you imply, a large swath of apps and SaaS are being wiped out by end-users directly using AI. So the only viable path is to raise the bar on what is being built. Someone should be vibecodinfg a phone OS and vibe-building an actual phone alternative to Android and iOS, if it's actually possible. Someone should be vibecoding the Metaverse and vibecoding hundreds of Apple Vision Pro apps to make VR/AR actually viable (again). And on and on.
The only honest counterpoint to the author's post, unless you can cite vibecoded software at the level of scale and maturity they are seeking, should be:
Yeah, 2026 was the year the coding performance of LLMs passed a certain maturity threshold adoption-wise so we should realistically give it a few more years to see if it is actually possible to vibecode continuously released ambitious complex software used by demanding users on the scale you mention.
Yeah, you can use some ai models to completely reinterpret your photos with different styles, or just skip the "having your photos" as a source in the first place and generate them synthetically.
not some iteration of a markdown reader, that's easily copied of GitHub. or some useless React app - that doesn't interact with external data like most businesses do.
look around u - even the biggest companies can't ship software that works. n u expect an LLM to do so.
118 comments
[ 4.2 ms ] story [ 75.9 ms ] threadThat might be a letdown for some.
[0] https://metr.org/blog/2025-03-19-measuring-ai-ability-to-com... does nice charts
Happy I can use Claude for more trivial things, but if the AI labs wanted an actual benchmark, Photoshop or Davinci Resolve would be it.
For artists and creatives, _all levels_ have been taken away. And for most, their very basic needs (i.e. income) are being taken away.
And all that, just so we can get past "level 1"? Not even "level 4"? What kind of trade off is that??
We're supposed to be engineers. We spend days/weeks discussing how a table should be designed so it doesn't bite us in 6 months time. Yet here we are discarding almost _everything_ just to make "level 1" a little bit more convenient.
It's pathetic.
Moving the goalpost.
Can it be done? Probably. But the token cost to do so would be astronomically expensive. You still need a human (read prompt engineer) to steer the AI. Even small ideas I’ve thrown at it to test viability is often plagued by code that just simply doesn’t work. I am constantly having to send it back and say “this doesn’t work” and it will eventually figure it out but that is more wasted tokens just to get something that doesn’t throw an error during the build process.
One-off specialty programs are absolutely going to feel the heat. Take a multi-tone generation application (think Motorola pagers). You can now ask an AI to create that program for you, complete with tone generation and .wav recordings to use later.
Large programs are relatively safe for the moment just because of scale. But any small application that is usually behind some sort of paywall or license is absolutely going to be threatened since it’s no longer difficult to throw together a program in an evening and have 100% of the features you want and none of the features you didn’t need or want.
I think of "tokens" as units of intelligence spent on a problem.
Producing software like Photoshop, Linux, or Excel literally cost billions of tokens (human intelligence back then, and now both human and artificial intelligence combined). A lot of deep intellectual problems were solved along the way, and the token cost of solving them accumulated. The true cost of Photoshop today should be measured in billions, maybe trillions of tokens.
AI can definitely generate some tokens, and it has accelerated everything involving intelligence. But it still can't produce 100 billion tokens worth of intelligence in a week. When that day comes, that's superintelligence. We'll go to Mars, we'll make fusion work. That's the singularity.
Sure, I'm still the one _deciding_, but AI suggested things that I would never have thought about.
Replacing a mature app with an incredibly wide audience and a million use cases may or may not be possible, but what's actually happening is that people are making an app that does exactly what they want, and using it to solve their own needs.
Previously you might use Excel to take raw data from various places and analyze it, create charts, reports, extract findings. Now you can have AI write a script to produce the exactly report you want from the original data. No Excel required.
Photoshop is a good example for that, actually. How many people are just using ChatGPT or Gemini to get the image edits they want instead of reaching for Photoshop? I don't know if this is showing up in Adobe's subscription numbers, yet, but I expect it will eventually.
But over the next 5 years I expect a growth in Blender-like open-source projects aiming to take on the big closed-source elephants. Code is cheaper now. The main downside of LLM coding, unmaintainable spaghetti code, can be mitigated effectively with discipline and coordination.
You still need maintainers to uphold contribution standards, but people will throw tokens at you. A small, disciplined team can go a long way, make a decent enough product, and then attract the institutional money (like Blender did) and hit that growth curve where everyone rallies you and you've won.
Lots of companies would have a vested interest in reducing these dependencies to Adobe et al., or have a more customizable product. Competitive professional tools, more like Blender and less like GIMP, but in other areas, like DAWs, CADs, and others.
For instance I have created a snippet manager, because I didn’t like the one in Raycast. Or I have created a journal viewer because the daily notes in Obsidian are not easily viewed across days. And so I have dozens of smaller solutions that are tailored to my needs. And I am guessing lots of vibe coders do the same.
[1] https://github.com/EngineersNeedArt/Stereographer
The author's point is not specifically about Photoshop. It's right there in the second paragraph:
>Where is the vibecoded Photoshop. The vibecoded Excel. The vibecoded Maya. The vibecoded Blender. The vibecoded compiler that compiles itself. The vibecoded database, the vibecoded OS, the vibecoded anything-that-requires-architectural-judgment-to-hold-together
What do all these examples have in common, besides requiring "architectural judgment-to-hold-together"?
They are software created by mid-large companies or organizations with a large of numbers of contributors. It's not about Photoshop, it's about developing complex software with high quality control and a continuous mature and continuous release cycle to a demanding set of hundreds/thousands of users. So to take that it in and then say, "Well, I don't think AI going to be used fo build that stuff, it's going to be directly used by non-technical end users to bypass existing more complex legacy software like Excel," is only proving the author's point.
Cause at the end of the day, what are all these companies "token-maxxing" actually building and selling to their customers? That's the larger point the author is making. Just as you imply, a large swath of apps and SaaS are being wiped out by end-users directly using AI. So the only viable path is to raise the bar on what is being built. Someone should be vibecodinfg a phone OS and vibe-building an actual phone alternative to Android and iOS, if it's actually possible. Someone should be vibecoding the Metaverse and vibecoding hundreds of Apple Vision Pro apps to make VR/AR actually viable (again). And on and on.
The only honest counterpoint to the author's post, unless you can cite vibecoded software at the level of scale and maturity they are seeking, should be:
Yeah, 2026 was the year the coding performance of LLMs passed a certain maturity threshold adoption-wise so we should realistically give it a few more years to see if it is actually possible to vibecode continuously released ambitious complex software used by demanding users on the scale you mention.
not some iteration of a markdown reader, that's easily copied of GitHub. or some useless React app - that doesn't interact with external data like most businesses do.
look around u - even the biggest companies can't ship software that works. n u expect an LLM to do so.
leave the LLM to be a better search.