Comment by Xirdus
1 day ago
In my opinion, this post misses the forest for the trees. Yes, the model of an OS running various off-the-shelf apps for various tasks will soon become obsolete. But it's not the idea of the OS that's outdated - it's the idea of apps. Why would you ever want to ask an AI to make an app for you to complete some task, if you can instead ask the AI to complete the task directly? The future isn't a bunch of extremely personalized apps - the future is a single AI assistant doing all the work.
I've seen a few posts of the form "why would you ever want to use anything else than AI from now on?" by now. My counterpoints would be:
Unless you have a local model and the appropriate hardware for it, your agent is Somebody Else's Computer. Do you really want to send all your data and make your entire computing experience dependent on whatever OpenAI or Anthropic or whoever else is planning in this moment?
That's not even starting with inference time and token cost. Despite all the incredible advances in inference, it still takes more time than most non-AI computer functionality. Do you really want to wait a few minutes and pay money for something that you could also do with a few clicks fully locally on your PC?
But the most important thing: User interfaces. Right now, we're basically cramming everything you could possibly want to do at a computer into a chat interface. But there are lots of applications fir which specialized graphical interfaces are much more suitable. Why would you want to get rid of them?
Somewhat connected to that: Repetition. If you have to do the same task again every week or every day, it seems wasteful to ask the AI for it every time: Not just are you wasting a lot of time, energy and tokens, you're also at risk of getting inconsistent results, if the agent from today's session will interpret the requirements slightly differently than the agent from yesterday.
You can circumvent all those things by having the AI write you a custom app, but then using the app without AI to do the task.
Those are all techbical limitations that will be solved with time. In 5 years I would be surprised if we do NOT have an AI as smart as today's frontier models, that runs completely locally, and has a real time voice interface.
I'm not a Treky but I think Star Trek has the right interface here. For basic operations you just use voice and tell it your intent or what you need to know "Computer, what's the fastest route for me to get home?". "Madison is fastest right now, it's 3 minutes faster than Lake, it traffic is building up".
This isn't just whatever Waze gives you, it knows about your constraints and preferences. So you need to program that somehow. Imagine taking it one step further "Computer, what's the fastest route to pickup my daughter and get home?" or even it just responding to the first prompt "Take Roosevelt, and today you need to pickup your daughter from volleyball".
In Star Trek they often do this when they create "simulations". They tell the computer constraints and then it refines it. No writing programs, but effectively that's what you are doing but with the natural language voice interface.
"Computer let's create a simulation to figure out the fastest way for me to get home. I want to take a route along the expressway and know whether to get off at Roosevelt, Madison, or Lake. I need to pickup my daughter when she has volleyball." You don't need to repeat the whole instruction each time, and behind the scenes it is probably turning it into some deterministic code, so next time you trigger that program it's not going to hallucinate some nonsense.
Afair Star Trek AI is more like 'Ok Google', and generally doesn't really seem like very agentic, more like a computer with a voice interface, able to execute simple commands and answer simple questions.
1 reply →
Why don't you just use the transporter instead of your car?
And yet, even on Star Trek we can see people typing things into screens or onto pads. The helmsman doesn't fly the ship by talking to the computer.
My point is that text-based interfaces - and voice is just a convenience around text - are not the best interfaces for a lot of tasks.
We live in a "capitalist" society. Moats will be built, regulations will be enacted to ensure the American Corporation controls the market segment. No one has been able to unseat Amazon, no one has been able to unseat Google, no one has been able to unseat Microsoft, no one has been able to seat Twitter, Facebook, etc. The modern market is segments that are controlled by a single entity. There will be no local models.
> "Computer let's create a simulation to figure out the fastest way for me to get home. I want to take a route along the expressway and know whether to get off at Roosevelt, Madison, or Lake. I need to pickup my daughter when she has volleyball."
Imagine saying that to a random dude with a map and that should give you a good idea on how many info you're leaving out. You can say that the AI will infere those but that seems to be the common fallacy of AI enjoyers: They're always assuming the AI will magically find out the missing information from their prompt somehow.
I have Organic Maps installed on my phone and within a few minutes (faster than my car warming up) I can set a multi stop route (you can bookmark places). Training a random user to use such apps is also equally fast. The same happens with various pro tools: once trained, a user can be very fast with them.
9 replies →
Loss of control/privacy, poor performance at high cost, bad UX.
Fortunately our industry has been getting people used to these for years.
Are you criticizing vibe coding or assistants here? I really can't tell.
We are talking about a hypothetical future where coding AIs are good enough, cheap enough and ubiquitous enough that not just expert programmers, but everybody is capable of vibe coding exactly what they need, to the point they actually reach for coding AIs and not the app store with ready-made apps. That future is still far away, at least 5 years, probably closer to 10-15. In that hypothetical future, most of the problems you're talking about would be solved, or else we wouldn't have that future yet.
Man, you seem like you're thinking linearly... the capability growth does not look to be linear... 3 years tops IMO
2 replies →
Empirically, nobody has a problem with their stuff running on someone else's computer, with chat interfaces, or with repetition, or with repetition.
> Unless you have a local model and the appropriate hardware for it, your agent is Somebody Else's Computer
So, any Apple computer?
> if you can instead ask the AI to complete the task directly?
Going to entirely depend on the nature of the task and where AI pricing goes in the future.
Would you like to spend $0.10 on every task or would you like to spend $1.00 to make an app that does it free for you forever?
I don't think it'll be one or the other.
But also I don't see things like video games ever being replaced by 1 off AI. If for no other reason then the fact that a lot of games are fun because of multiplayer capabilities. That's not something you can do a 1 and done with AI (unless you only want to play games with family and close friends).
If it costs $0.10 to do the task once, then you can be sure as hell it costs orders of magnitude more than $1 to vibe code a whole damn app that does that thing, especially if you want good user experience which is the whole point of custom vibe coded apps. Whereas the App Store will have an inferior, mass-market version of what you want actually priced $1.
That's not necessarily true. Vibecoding a simple JSON -> YAML converter to translate a 2MB JSON is going to be much cheaper than including that JSON in a prompt and burning hundreds of thousands of tokens. Besides, there's lots of cases where the math simply favors a bigger one-time cost rather than paying fees in perpetuity.
3 replies →
What if all computer games from now on are AI Dungeon?
I imagine the future of computer interaction could evolve into a split model combining foundational capabilities that just make sense to be pre-built, tested and shared deterministic code - codecs, numerical solvers, geometry kernels, databases and transactional tools, compilers, theorem provers, cryptography, etc. - with a fluid interaction layer that accesses either local or cloud-provided versions of those foundational capabilities and is capable of synthesising any tools it needs where it would be helpful to build it first - either as a one-off or something to cache - for anything that can’t just be done as a task.
Applications are compressed bundles of some assumptions about how something should work and be interacted with. So you wouldn’t need VLC, it’s a shell around a foundational codec layer. Instagram is just a view over data in the Instagram social network, and you could customise it as you see fit. Transmit is a lovely application but you can delegate that to a fluid task layer that can do one-off tasks or make something you can re-use.
Everything in the fat ‘middle layer’ of software that is connecting a user to something foundational through interface assumptions might disappear, unless it’s more useful for it to exist as a permanent tool. Things that a lot of people share or that are more complex, Blender for example.
I also imagine there’d be a lot more focus on making explicit things like canonical records or schema, contracts, provenance, versioning, access policies, some kind of fixed views. Things that your local fluid interface would rely on the structure for when it came to connecting to your bank and doing whatever you want with your finances.
Not all software as we think of it might disappear, but the things that are lightweight layers over something more stable could all be subsumed with tools synthesised on demand.
Perhaps, at least.
An AI assistant can give different answers for the same inputs. An application written by an AI assistant, if it doesn't just call an LLM at runtime, will mostly likely produce the same answer for the same inputs. That is, unless there's something intrinsic to its business logic that makes it nondeterministic.
Sometimes you really want the latter even if the former is less effort.
And when applicable, the program would be much more efficient than the LLM.
So it would really be beneficial for the LLMs to be able to have their own environments to write and run code, but also to store those programs for later use automatically.
A *human* assistant can give different answers for the same inputs. At some point, we'll need to stop treating computers as instruction executors and start treating them as autonomous agents. Non-determinism isn't a bug to be fixed, it's just the reality of working with this new kind of machines, and operators of those machines need to embrace that fact, like they did for human assistants.
And we started adding rules for the humans to follow, until we have complex bureaucracy. We invented tons of formal notations systems and error detecting and correction mechanisms. We invented deterministic machines. All because we are deeply unsatisfied with the non-determinism! Because it means you can never trust some process, optimize it away and stop thinking about it. You always need to check, and there is always something missing.
A lot of our manufacturing ability improvement during the last 200 years, as well as our engineering ability improvement during the last 500 years, is not just the raw power we unlocked, but that we made it deterministic, reliable and reproducible and hence automatable and automatic.
1 reply →
So would you want a genius figuring out how to do everything you ask it to from scratch, or do you want to train someone how you want something done, so they do it the same way every time and you don’t have to explain yourself over and over?
Nondeterminism is actually a bug in many cases, and training, process and procedure, is how we mitigate it in humans.
1 reply →
But sometimes you do just want the same answer each time. Not this human woo crap, where the answer depends on god knows what, but a deterministic function of some kind, reliably producing a given output for a given input according to a fixed set of rules.
Computers can do that!
5 replies →
I want an agent studio and execution sandbox where my LLM driven assistant builds and maintains my apps with a receipt or transcript like history of what it did to create each app. Determinism built per use case with non deterministic assistance orchestrating the build and management layer. Extra credit if I can trivially share and collaborate with others per “deterministic silo” via a shared link. Something like Claude Code + AWS Lambda|(Docker|Podman) + Tangled.org with LLM code generation for my phone and workstation.
The LLMs empower for building, the determinism improves output-expectation alignment. It’s fundamentally an on device software factory, hypervisor, and storage system.
> my LLM driven assistant builds and maintains my apps
Remember when everyone was talking about treating computers as cattle rather than pets?
This seems like a weird inversion. Programs as pets instead of cattle. Toxoplasmosis, ahoy...
3 replies →
> Yes, the model of an OS running various off-the-shelf apps for various tasks will soon become obsolete.
I doubt it. Most people would rather stab at their phone to download an app that just does what they want than explain to a chatbot over a period of minutes/hours and several reiterations to get it to spew out what they want, let alone what kind of software they want it to make for them. Most people can barely articulate what they want in the first place.
The market for applications isn't going anywhere and neither is the OS. I can't imagine booting a computer directly into a chat window where I have to explain to a bot what I wish would show up on my screen, where on my screen, and how it should look. What a total nightmare that would be!
Even if what passes for "AI" these days was as smart and capable as the computers on Star Trek (and it very much isn't) you wouldn't want to ask it for everything. Even the Enterprise had an OS that people interacted with. What they also had though was an OS with configurable interfaces that people could customize according to their current task and preferred workflow.
That sort of thing is maybe possible in the future for "AI" empowered operating systems we might live to see, but 90% of the people using computers aren't going to bother with any of that, just like they've never seen the settings/options in most of their applications or their operating system right now. Not because it couldn't make their time working the computer better, but because they just can't be bothered.
The writer of this piece remembers how, as a child, computers promised to let him write his own software and games and it seemed like there were endless possibilities until the reality that programing takes effort killed a little of that enthusiasm (although not enough that he didn't eventually become a programmer), but many people never see a computer as anything more than a means to an end or a device for mainly passive content consumption.
The cognitive dissonance is astounding in AI circles. The vast majority of people have a personal computer in their pocket and they use it the vast majority of the time to scroll Instagram, play games, and occasionally interact with their bank and other social services. People who have laptops and desktop computers have them for specialized tasks they already do. None of the vast majority of users who buy their phone when the screen cracks and the battery stops working care about the agentic AI world because their phone already does everything they want to do.
I think the rise of Claw's, Grok Bot and Muse is a good counterpoint to this. Yes they are still niche specialized interfaces right now, but they are basically allowing non-technical people to do things that have been easily possible with scripts and cronjobs for a long time. Its basically the Ford and faster houses analogy.
11 replies →
Okay, I went too far with that one sentence. Can we all pretend I have instead said "this model will be obsolete for most, not all, but most productivity apps and most, not all, but most office tools"?
“Go enjoy my family photos for me.”
“Go play video games for me”
“Go create music for me”
Idk. Seems like there is still use for apps.
> Yes, the model of an OS running various off-the-shelf apps for various tasks will soon become obsolete.
Define "soon".
Banks in the UK have thirty million customers each. They have moved away from web interfaces to apps on well-controlled platforms.
Social media doesn't use the mobile web; it's not good enough.
Neither of these applications for a phone are simply going to surrender to agentic environments on a phone that de-emphasises apps.
They both have a need to fully control that experience, and yet they are polar opposites in terms of seriousness.
We can't even stop voice assistants acting on commands they mishear from TV shows, youtube clips, radio.
Absolutely agree. At the start of the year when Opus 4.6 came out, I remember my co-worker showing me a bunch of apps he made for himself to make himself more productive.
Fast forward a few months, and most of those apps are abandonware. Not that he's not as productive, he just skipped the middlemen (the apps), and started directly asking the agent to do those things.
It is not clear whether models will be both capable enough and cost effective enough but not as capable to wipe us all out. The outcome is a tiny sliver of possible futures and even then likely short-lived.
Current capabilities are not there. You need software engineers to keep it going. And even then as intoxicating as it might be to be able to bring to life all the side projects you often had no energy/skill to finish, today it is still may be cost / “cognitive load” -effective to buy off the shelf apps that somebody else has to maintain (even if they are not as personal / less customized for your use-case).
> Why would you ever want to ask an AI to make an app for you to complete some task, if you can instead ask the AI to complete the task directly?
Given the current trends in predictive language models, once users are able to articulate their needs with clarity and precision, conventional software development may become obsolete.
This reminds me the car industry touchscreen craze: if you can paint anything and control anything via the touchscreen, then the only thing you need is the touchscreen, right?
Exactly my opinion too.
Apps will become barely visible data providers. Some will provide specialized widgets for those cases where a specialized UI is needed. And data will be designed to inter operate.
I have reached this conclusion from a specific example I'm working on, more details in this blog post if you pardon a link to a website of mine: https://cloudywithachanceoflatency.net/blog_02-how-llm-accel...
Like browsing TikTok or messaging your friends? People want their devices to work like their friends’ devices do. Your prediction seems to be out of touch with how and why people actually use computing devices. Nobody actually uses them to "perform tasks".
Maybe we need an app store just for agents to make purchases. Imagine a world where startups have their valuations determined by how many agents are using their apps. Those startups... Also operated by agents. We don't even need users anymore!!!
Why would you ever want a nice word processor to write a letter to someone when you could just ask AI to write the letter? Why would you ever want photo editing software when you could just tell the AI to make your picture look good? Why would you want audio recording and mixing software when you could just tell the AI to output the next Billboard 100 hit? Why would you want a camera when you could just ask AI to generate photos for you?
It seems like you're going for a rhetorical here but there are people whose usage of word processors and photo editing is and has been trending down to 0 because of AI.
This is the part that many people are missing. Decades ago I heard "the ideal UI is a search box," and I agree today even more.
I don't understand why people are still thinking that clicking on stuff will still be a desirable thing in the future.
Inputs for the ui might be me talking, or typing, or having someone else talk or type. They might also be GPS, temperature, BLE devices nearby. Or maybe a button I smash because I don't feel like re explaining something.
What happens after an input is still the best part of a UI. We like seeing/hearing/feeling things in different ways.
The fun part about AI assistant is that you can straight up tell it how you want data presented, and it will do its best to present it that way (within limitations of its capabilities and available tools).
I agree - we should have many more good fully-keyboard-accessible UIs!
I don't want to talk to my computer to use it.
What you’re describing is basically a terminal. We’ve had those for a while.
I agree, which is why I don't understand why these are treated like magic tools by the laymen.
It’s funny you say this because Apple resisted an App Store for apps. They originally expected web based apps. The market demanded an App Store. Remember web snippets where one could clip a portal to view a small section of a website?
The current market very clearly demands AI assistants that can do everything. See ChatGPT app, or the new Muse app.
There is a big misconception in your answer in that you seem to believe that an LLM would always produce (machine) code that does what the user intended, in a correct and safe manner. Neither of these assumptions is true. If you knew how LLMs are built and operate, you would know they are not reliable at all. What you might ask from an AI interpreter OS might be unique and thus might be absent from its training set and might not follow a pattern inferred from its training set either.
LLMs are the first machine learning models that blatantly and regularly produce incorrect output and we have been brainwashed into accepting that. An application on the other hand can be exhaustively verified. There is no comparison.
First and foremost, I'm comparing specifically just having a suite of heavily personalized vibecoded apps like in OP blog post, versus using an agent directly. I'm comparing just these two options and nothing else.
If we don't assume up front that AIs are good enough for at least one of those things, then there's no conversation to be had. Personally I'd rather have a conversation than not have a conversation, but you do you.
Had this thought this morning after seeing DHH announcing that Basecamp isn't writing code by hand anymore and they're using AI to write Rust for Hey's backend.
To me they're failing to see the even bigger picture that maybe even Hey itself is outdated in the Ai era and not only is its Ruby framework antiquated... the whole Hey service is too.
> The future isn't a bunch of extremely personalized apps - the future is a single AI assistant doing all the work.
I suppose if you break your glasses, squirt real hard, and drop down about 30 IQ points, then sure.
We're extremely far from anyone, even those with the most interest and capability, giving a practical demo of this future you're suggesting.
What we all see irrefutably happening is some new people are just barely able to do slightly more than they used to. That's the impact we'd expect from a better search engine and a better "I'm feeling lucky" button which, of course, is what this era of AI is.
> We're extremely far from anyone, even those with the most interest and capability, giving a practical demo of this future you're suggesting.
Yes, absolutely. This is a very far future. But that future is (maybe, possibly, hopefully) coming, unlike this other future which is absolutely never coming.
Why "hopefully"?
Having AI "do the task directly" sounds just as miserable as anything else people keep insisting is the future. We can't rely on probabilistic methods. We will still need humans in the loop to pin down the deterministic parts and also the parts we intended (often the same, but not always).
What else did you have in mind?
1 reply →
> the model of an OS running various off-the-shelf apps for various tasks will soon become obsolete
"ChatGPT, please send <username> a message with the following contents: <content>. Oh, what platform? Create the platform to send the message for me. That username doesn't mean anything on this brand new platform? Uhhhh" [insert step where the AI-psychotic decides to give up on communicating with humans and tells ChatGPT to roleplay as the person they intended to message, I guess]
I hate that web platforms and other low level cloud infrastructure are called "apps" these days. Makes it impossible to talk about actual end user apps, vs. all the other code that's also running in addition to actual end user apps, which will remain in place even if actual end user apps become obsolete.
In other news, switching from gas guzzlers to electric cars had the unfortunate side effect of highways disappearing.
> the future is a single AI assistant doing all the work.
Sounds more realistic than what this blog post is proposing. Meta also tried releasing a phone, how did that go?
Muse on the other hand is the polished version of what OpenClaw should have been and Meta appears to have given a lot of thought about it more than this blog post.
I don't think we need yet another phone.
The author is missing fundamentals. Exactly what to expect from someone who thought the built is BASIC was lame.