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.
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.
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.
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.
Wow that's slow. I can do the same in Google Maps in seconds. Can you tell me what I'm missing? I'm definitely missing something.
> just drive easy for a couple minutes.
Fortunately our industry has been getting people used to these for years.
Sometimes it is wild to me how two people can look at the same information and come to completely different conclusions.
As best as I can tell, the only constraint the really matters right now is availability of compute, or as a buddy of mine put it "GPU GO BRRRRRRRRRRR!"
So, any Apple computer?
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).
With the caveat that literal today's models are not allowed to run code themselves but instead give you the code in the chat window and ask you to run it yourself for security and safety reasons. But the capability is there.
A tool for generating a report for September 2026 is no more expensive than an agent generating a report for September 2026. But that tool is useless for you in October 2026. You'll need to generate a whole new tool for October 2026, and generating 2 tools costs as much as generating 2 reports so you're not saving anything.
Making a tool that can generate a report for any single month can easily cost 10 times more than a single report, entirely due to the extra code flow analysis required to extract the right variables in the right places. And that's without any UI. If you want UI that's another several times more tokens because UI is hard. Also it will suck compared to a simple chat window where you can ask the assistant for anything and it will understand. If you want UI that's better than the chat, that alone costs you more than everything you spent so far.
All in all, your permanent solution starts paying for itself in about 4 to 6 years, assuming you never had to update how the report is generated in the meantime.
Or you could have a very simple, very cheap script that only gets run once then gets deleted (or even better, kept for future reference to make future report scripts even cheaper to generate).
Note that this is all in the hypothetical future where AI is actually capable of one shotting such a task. If it makes mistakes, then human-made apps are still the most cost-effective option and this whole discussion is moot.
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.
Sometimes you really want the latter even if the former is less effort.
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 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.
Can you build a new hospital? Up to the local government to decide. How much is your property tax? Only a tax assessor can tell you. Which trees get trimmed and which are left intact? Decided on case by case basis with no strict rules. Which convenience stores are getting audited for tax and food safety compliance this week? There's a middle rank officer who makes that decision every week, my mum used to be one. How high is the bail set in criminal proceedings? Judges are famously very inconsistent on that and it's a good thing. In countries without constitutional gun rights, it's common that the final step in the process of obtaining gun permit is some government employee having a good gut feeling about you, and if they get bad feeling you get denied. And so on and so forth everywhere you look.
Manufacturing is a different beast entirely. I'm not saying we should put Siri and Cortana in charge of bolting down Boeing fuselage. But for most office and bureaucratic tasks, demanding perfect determinism is not only unnecessary but actively harmful.
Nondeterminism is actually a bug in many cases, and training, process and procedure, is how we mitigate it in humans.
Defining every single step of every single task and making them perfectly repeatable works in an assembly line. But as an office assistant - and doubly so as a personal assistant - it creates inefficiency, makes dealing with edge cases very painful, and greatly increases maintenance burden of keeping all procedures up to date. Instead, giving people some leeway in how they approach their tasks and deliberately underspecifying processes to let them exercise their judgement is shown to work much better. Surely you had it covered in your training materials?
Computers can do that!
That said, my point is that determinism is incidental to how current generation of compression algorithms work, not an explicitly desired quality. And even that isn't strictly true, because e.g. x264 in certain multithreaded configurations already produces non-deterministic results that look as good as single-threaded encoding. For this use case - and thousands of others - losing determinism when switching to AI-based solutions really isn't a big deal.
The LLMs empower for building, the determinism improves output-expectation alignment. It’s fundamentally an on device software factory, hypervisor, and storage system.
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...
Think of all those bad sci-fi movies where aliens take over Earth, enslave all humans and treat them like cattle. Very competent cattle that's put in charge of local governments, scientific research etc. but very much still cattle where individuals are entirely interchangeable and killing them is just everyday work. This is me and AI assistants. (I'm the aliens.)
Pros and cons to code being cheap to generate now, it’s always about tradeoffs. I don’t want to build cars, I want to drive cars.
People suck at "good enough". All major engineering disasters were something "good enough" for somebody. People burn down their houses with "good enough" electrical work.
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.
Most of those websites have been complements for Google Search rather than standalone services whether they like it or not, and if chat bots replace Google Search, they'll replace the entire ecosystem.
Not that this matters, because Google search alone disproves the claim that people only use phones for videos and social media.
I do personally think that we'll see regular users interested in new, different workflows than the automation ones we associate with power users, and the image generation use case you mention is a good example. Chat-bots do seem to be owning those rather than entirely new apps.
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.
“Go play video games for me”
“Go create music for me”
Idk. Seems like there is still use for apps.
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.
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.
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).
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...
I don't understand why people are still thinking that clicking on stuff will still be a desirable thing in the future.
What happens after an input is still the best part of a UI. We like seeing/hearing/feeling things in different ways.
I don't want to talk to my computer to use it.
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.
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.
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.
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.
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.
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?
If AI keeps getting better, then it is inevitable that voice conversations will become the dominant way in which mostbpeople use most computers for most tasks most of the time. It's just part of human nature and there's no way around it. The only way this doesn't happen is if AI never becomes good enough to support it.
Personally, I'm not a fan of voice input. But I want AI to get better. And the future where voice conversations work well enough to be used as the main HCI method is inevitably also the future where voice conversations are used as the main HCI method by most people. I'm hopeful for technological progress, that's all. It's just that you cannot have this progress but not have people take advantage of it in this exact way.
Compare it to the future where most people vibe code their own apps, which has zero chance of happening no matter how good AI gets.
"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]
In other news, switching from gas guzzlers to electric cars had the unfortunate side effect of highways disappearing.
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.