How I Think We'll Interface With the Digital World in the Future

A prediction I have is that interfacing with the digital world is going to look very different in the future. You're going to talk to an agent, your agent is going to talk to other agents, and most of what you do on your phone right now is going to get done without you looking at a screen. Think of how people used to go through a stockbroker or a doctor. That's closer to it than anything on your phone today. A lot has to line up for that to happen, and it starts with software getting almost free to make.

AI is going to build the software

Companies like OpenAI and Anthropic have most of the compute and the top AI models. Everyone is using their tools to build things, and those models are training on much of the same data the entire software world is training on. So they could essentially just build everything. We're already at the stage where, if you want a website or a particular app built, AI can produce the slop to build it.

But the slop still has to be fine-tuned and made better. Better prompts get you better apps, and the models keep improving. The open problem is taste. There's a lot of work going into aesthetics and front-end quality for AI, and it's still a big problem, but I don't see it as an unsolvable one. Once it's solved, anyone can have their own version of whatever software they need. The catch is that all of that generation gets expensive, in cost and in electricity.

That cost is why I think the next step comes from the companies at the forefront of AI. If they build a framework so people can easily tailor what they want on the spot, a truly agentic OS becomes possible, running on a model small and efficient enough to live on an edge device. That's going to be the future of software.

The interface changes with it

Because software is getting so easy to make, the interface is going to change too, and I think it moves toward wearables. Speech-to-text has gotten so good that you really don't need to type anymore, or at least not nearly as much, and it's a lot faster. So wearable devices with microphones, like smart glasses, smart bands, and rings, are going to become more popular. If you had smart glasses with a camera and speakers, something like the Meta Ray-Bans but a lot better, you could do almost everything.

Now, you don't have a screen, so how do you solve that problem? Most tasks won't need one, which I'll get to, but for the ones that do, the Apple Vision Pro was a great step, because it let you augment your surroundings with screens. Mechanically and logistically the device had enough issues that it never became commonplace. Once those are sorted out, a simple device like that becomes a no-brainer compared to always buying the latest phone and charging it more often. I think the phone has peaked as a form factor.

When it comes to interfacing with software and services, there will be an AI agent on each end. Say there's a Vercel agent. Your agent could interface with it and transfer data, as long as both are following the same data protocols and security exchange keys, and the Vercel agent could spin up a website using information from your agent. The guardrails would be important here, to make sure no agent is acting up, and any agent that doesn't follow protocol gets reported.

The first version of this is already here. Meta just came out with Muse, its personal AI agent, which can interact with Instagram, Facebook, and WhatsApp, and which Meta says is coming to its glasses. SpaceXAI has Grok Bot, which can interact with specific services as well, and which is built as a team of bots, often with one managing the others. I'm sure the other software behemoths will come out with their own agents that interface with their proprietary data and tooling. Each of these agents can do essentially the same thing: manage your life and manage your information across different services. But a single person doesn't really need more than one agentic service managing their life. Maybe one for personal life and one for business, but not every company's agent needs to be the one you use.

People are on a lot of different services, though, and when you want to interact with a specific one, its own agent is the most convenient way in. So the interface goes back to what I was describing. Each person has their own edge agent, on whatever device they carry, and that agent communicates with the Muse agent, or Grok Bot, or the agents behind every other portal and service they use. Muse is clearly trying to be that one agent for everyone. It will be the gateway to all of Meta's services. Most platforms in the future will have agents in front of their services. For now that agent lives in the cloud, which is where Muse and Grok Bot actually run, and it moves onto the device once the hardware gets small and cheap enough. That is how I believe the future of digital communication is going to happen.

Defense may be easier than offense

The main thing to be careful about in all of this is data privacy and security. Those have to be built in, solidified, and extremely robust, especially once we get into breakthroughs in quantum computing. This is where I think a lot of the smartest models should be put to work. If every company and every government has its own superintelligence, some of them are going to be pointed at each other, one attacking and one defending, but I think defense has the advantage. Defense is a community-driven thing, with everyone working on it together, while offense is usually a small group of actors. Whatever one intelligence finds as a way to attack something, another intelligence will patch. So I expect security to be a solved problem, unless we find out that nothing is actually secure, in which case we'll have to find different methods for sensitive data. But that's something for the future.

A lot of these tasks don't need a visual interface

We're already spending too much time on our devices. Ordering specific items, checking email to see if there's anything interesting, filtering out spam, reading through work emails, writing messages: it takes a lot of time, and for some people it's fatiguing. What we've seen with something like OpenAI is that a lot of people enjoy telling their AI to just go ahead and do these things, and once you do that, you realize a lot of these tasks don't need a visual interface. You can choose to look if you want to, but there's enough inference that the AI can usually do it correctly, and then you're only shown the things you need to see. Some people are hesitant because it's not good enough yet and makes too many mistakes, and that's true right now. But once it gets better, this is the direction people are going to go.

We're glued to our devices because everything is an app, and we go to separate platforms to do separate things. Once it's all brought together, we interact with one agent and that agent interacts with the other apps. Even if we never end up on one universal platform, that's fine. We're still interacting with one thing, and that one thing is interacting with everything else. It's another way out of the app fatigue and screen fatigue we have now.

A giant hub of software

Say there's a container in the cloud that already has something like Photoshop on it. You clone it, use it, and stop using it, and the next person gets the same compute with no memory of you. That kind of ephemeral, shared computing could become really important, and you can already see the pieces of it. There are apps that bring your terminal, web browser, ChatGPT, plots, and code together in one master app. The operating system basically doesn't matter anymore because the interface is so customizable. So I think the differences in platforms and compatibility slowly go away as more things execute in containers. An AI agent just finds the appropriate container where something can run, and if it needs backwards compatibility, it develops an adapter on the fly.

What that leaves is raw hardware optimized for compute, and then really good software with a lot of pre-built components, or components that AI is already building and sharing online. There could be a giant hub of software for that. GitHub is kind of like that right now, but GitHub has raw code, not raw software, and you still have to download it, install it, and figure out how to use it. Maybe that's okay, because an AI agent can go on GitHub, find what it needs, install it in a container, do what it has to do, and return the result.

A pool like that only works if everything in it is highly vetted, scanned, and checked for security. There would need to be a lot of AI agents doing that, along with people manually checking and validating software before it gets approved. It makes the most sense for these repositories to be hosted by the people who own the agents, like OpenAI, Anthropic, or SpaceXAI. It's a little trickier for the open-source models, but they could use a community-vetted pool of tools. A lot of what's in there could be bare-bones, with an easy customization or UI layer built on top.

The SaaS system is going to go down

Right now, Chinese LLMs like Kimi and GLM are top-tier and pretty close to the American frontier models, but they're way cheaper, and some of them are open-weight. Hardware is scarce at the moment, but once China comes out with really good, cheap hardware, or someone has a breakthrough, that problem gets solved. Then all that's really going to matter is: do you have the hardware, and can you pay for the electricity to run these AIs? Wherever electricity is cheaper is where this becomes most valuable.

Then what are people going to do? The value of Anthropic and OpenAI probably moves from the models themselves to the frameworks and tool hubs around them, but that's not the important part. The important part is that the whole SaaS system is going to go down. People are already pointing Fable or Kimi at a UI and having it reproduce that UI. I saw one example, and I won't name the tool, where it copied the app UI of Flighty, and it did it easily. Copying a UI doesn't require taste. That's why it's already easy while original design is still the hard part.

I know not everyone is going to be a developer, but these systems keep getting easier. Right now you can go from speech to code. Eventually you'll go from speech to application, and once that's truly achieved, business owners and the people who actually use the software can create what they need. It doesn't make sense for developers to continue making tools and selling them to other developers because those same developers can easily make it themselves with just a single prompt in the future. Or goes to an open source version, so there's no redundant token burn. A lot of companies are paying huge amounts for these tools. I know a company that was paying over a hundred thousand dollars just for Slack, and that's not going to keep happening. Companies are going to have their own internal messaging systems, configure their own security, and have an agent, or a set of agents, manage and host it.

All of that is going to exist in some kind of agentic software-hosting society or platform, and I'm curious whether there's going to be a term for it. Maybe agents as a service, although the acronym would be AAS, which is horrible and, honestly, kind of funny. I guess it's still SaaS, but just at a lower layer. So the money doesn't disappear...it just moves from per user software to compute or electricity to whoever hosts that agent. But once the hardware is cheap enough, people will just host that agentic layer themselves. So it truly becomes a system where my AI is talking to your AI. My agent talks to your agent. That's essentially what the system is going to be.

No one knows what's real

The big thing that comes with all of this is AI fatigue. People are going to go on social media, go online, and there's going to be so much AI everywhere that no one knows what's real and what's fake. Everyone is going to start building slop to the point where no one really cares anymore, and in that case authenticity becomes a lot more important.

I think people are going to want more in-person interactions and more of their lives away from screens and AI. You can already see it. They're trying to live healthier, they're feeling the drag of the whole dopamine craze that comes with being online, they're leaving social media platforms, and dating apps are crazy. But we still need software, and we still need to interact with it, and that is where edge intelligence comes in.

There's another direction this could go, and it's already happening. People get completely lost in the digital world, and some will prefer that. Virtual reality and gaming are going to become way more immersive and way more enticing in terms of what they do for dopamine. Even pornography addiction might get worse when people can generate things based on their exact fantasies. And once robotics comes along, some people who don't truly want to reproduce may feel they don't need a partner at all, because a robot can outperform a human in the things they care about. I don't think that's the long-lasting direction. Video games will still be there, with more immersive versions coming out, and that's fine. That's high-quality entertainment. The difference is whether it's entertainment or a replacement for people.

What stays human

There are also things AI isn't going to reach on its own. Emotional intelligence develops and is expressed differently in different cultures, which is another reason you might need different AIs trained in different knowledge scopes. There could be obscure information in an Indian village or an Indonesian village that AI is never going to know, because that information just isn't available anywhere. So how do you go learn that information and bring it to AI? That's a job. Exploring, learning, and documenting, gathering knowledge that's never been collected. The archaeologists, the knowledge people, whatever those roles end up being called, are going to become more important, because there's so much out there to collect and because it has to be truly understood, which isn't something AI and robotics can do entirely on their own.

Interpersonal communication is similar. Facial expressions, observation, the things that still rely on our senses: AI and robots aren't going to pick up all of that, because they don't have the biological functions. We could say we'll give them the biological functions. But a true human is probably going to react differently to that, and there's going to be a separation there.

People love the human in the loop

Once your agent is doing things for you online, the other side needs to know there's a real person behind it. That's why I really like the idea of Worldcoin. You use your biometrics to create a unique identity while staying anonymous, because I think anonymity is going to be important in the future. Ideally you could produce multiple identities from it, an anonymous one when you want it and a public one that's verified through the biometrics, and no one should be able to relate the two to each other. Some countries might pass laws requiring that all linked identities be identifiable. I doubt the U.S. does, but who knows. Worldcoin, or something like it, proves a real person is actually interacting and not some bot army, and it would need to be run by a global organization, almost like how the internet is a collaborative effort by the world.

I think, or at least I hope, more of the world heads in the direction where people are more personable, outside more, and interacting more, while the work and online interactions get handled agentically. There just need to be secure protocols and order for that to happen. Digital maintenance, verification, and monitoring can be handled by agents, and different parties can have their own variations with their own guardrails, checking for the things they care about in the way they want them checked.

It's going to be almost like an API. Instead of an application programming interface, it's an agentic gateway interface, or just an agentic gateway, whichever sounds better. Maybe there will be a better term. It's almost like how, back in the day, we did things through a broker. For the stock market or real estate, we went through a stockbroker or a real estate agent. For tests, we went through a primary physician. Now we'll have agents to go through. There can still be a human in the loop. People love the human in the loop, but that human might have a slightly different role.