Zo Computer

A personal computer in the cloud operated conversationally

A few months ago, our team decided to spin down the AI infrastructure company we’ve been working on and start building a new consumer product.

We’ve been working on Zo Computer. The best way to think about it is: an always-on personal computer in the cloud, which you operate through conversation. But what does that mean, why would I want one, and why now?

Cloud Computers

Let’s start with just computers in the cloud. These are seriously powerful computers with excellent CPUs, terabytes of RAM, effectively unlimited disk, and the best internet connections on earth. When coupled with modern container technology, these machines are what we colloquially refer to as “the cloud”—a flexible industrial scale computing environment that runs virtually all the digital services we use today.

While this compute is generally provisioned and consumed by big corporations, that's not because it’s inaccessible or too expensive for regular consumers—it's just that they're hard to use. Unlike your laptop, these computers don't have graphical interfaces that make sense to humans. Rather, you operate it by typing a kind of computer-speak to a "shell", which, while powerful, is tedious and unforgiving. The "programs" that you run on these machines are not apps made for humans, but generally bundles of code downloaded from public repositories, which require arcane knowledge to configure and run. Many of today's valuable software companies can be thought of as efforts to package the right mix of hard-to-use free software (databases, media tools, ...) into an experience that solves common problems. This is not to discount the value of good user experiences, but from a capabilities perspective many great software products are relatively thin cosmetic or workflow layers on top of free community-driven software.

The thing is—AIs are exceptionally good at translating natural language into this computer speak. Zo Computer is a way for regular consumers to enjoy some of the nice properties of "owning/operating" a cloud computer, while interacting with them in an intuitive way.

Computers are the tool—pointers and windows and LLMs as interfaces for operating that tool

Zo enables users to operate personal data-center computers without needing to learn computer-speak. Not-coincidentally, finding a way to let people use computing without having to learn computer-speak was the same problem that sparked the first boom in computer use.

In the late ’70s, people like Alan Kay realized computers needed to go beyond a black screen with green text to get mainstream adoption. Their solution was to generate a spatial illusion of windows, icons, and a mouse pointer that facilitated computer operations. The thinking was that humans naturally operate spatially, and so the spatial affordances and metaphors around “navigation” offer an intuitive way to use computing.

LLMs now offer a similar thing: humans are very good at talking, and LLMs can translate intention into the formal language a computer speaks, finally giving us the inevitable interface to computers that we have seen in countless sci-fi worlds. Visual interfaces changed the world by enabling non-specialists to use computers more easily, conversation is a similarly important computing interface.

What using Zo feels like

Having a Zo Computer means you have access to a private always-on computer in the cloud, and you operate it mostly by talking to an AI. You get some familiar graphical affordances—a file browser, document editor, a terminal—but the heavy lifting (installing headless tools, running simulations, writing code, hosting websites, etc.) is delegated to an agent (though advanced users can obviously jump in here where they'd like).

Why this design matters

Cloud computers are quite different from personal computers like a laptop, and have features that are useful but generally unfamiliar to most people. They also happen to be the perfect environment for AI operated computing for a few reasons.

1. Full AI autonomy
   The computer is the tool. Rather than bolting a dozen rigid mini-tools onto an LLM (MCP), we just let the AI use the full system directly, in the language computers already understand. A raw computer connected to the internet is the most flexible tool ever conceived of; given enough time and effort it can basically do anything, now we have a conversational way to make things happen.

2. Safety and rollback
   Giving an agent full access to your personal laptop is a bad idea: LLMs are imperfect and exploitable; you are vulnerable to getting your information stolen or your system being broken. Zo lives on a layered container stack, so you’re in control of what data enters, and you can snapshot/rollback instantly if something unintentional happens, like if it deletes files, makes incorrect edits, or breaks some programs. OS level “undo” is not something that happens on our laptops, but it’s a common, cheap feature of modern systems. Having this rollback option is essential if you also want the capabilities that come with full AI autonomy. Zo keeps a rigorous, searchable audit log of every action, so you can always inspect what has happened.

3. Hardware portability
   Because containers separate software from physical machines, the entire system can teleport between machines worldwide. Day-to-day you might run on modest specs. When it’s time to render a video or train a model, move the whole system onto a powerful GPU machine, finish the work, then move back. With cloud computers you pay for just what hardware your software uses. This is perhaps one of the biggest differences between physical personal computers and virtual ones. If you edit videos, you may pay a premium for a top-of-the-line laptop, usually many thousands of dollars which you might replace every few years. Most of the time this powerful machine is either off or doing basic tasks like email. With cloud computers you always have access to computers many times more powerful than even the best consumer computers, while being able to use them for just short periods of time, at generally less than a dollar an hour. This is because the industrial compute market is much more efficient than the consumer one. For consumers though, there just haven’t been good ways to access them.

4. Datacenter quality by default
   All of this happens in a datacenter built for scale, redundancy, and speed. You inherit all of that for free. Your computer is on 24/7 so you can e.g. host websites from there, run a local database for personal projects, make yourself some custom tools and automations.

The bigger picture

While it’s cool that LLMs can encapsulate a lot of facts and trivia, one of the more significant impacts of this wave of AI is that it offers another interface to computing: conversation. In a lot of ways, it’s the digital computer, not AI that is, and has always been, the great invention. You can do virtually anything with a computer. Up until now it’s just been very tedious and difficult to put that truth into practice. The result is that specialized organizations produce large-scale generic things that serve billions of people. The internet today has become an extremely optimized monoculture of platforms and content because of computing’s high barrier to entry.

Early internet users will remember certain aspects of how the internet used to be more diverse, organic, and ultimately humane. The modern internet is basically IKEA: an institution that can do the market research to find out what people generally want, produce it at scale, and sell people cheap veneered commodities because they can’t make it themselves. IKEA is fine, and even important in many ways. But Zo is more like working with a custom furniture maker who has all the industrial tools to make quality, just-for-you furniture, while being affordable. As more people have the ability to make computers do what they need, the digital world will become not only more useful, but cozier and more nourishing.

We think that there’s a world where the physical computers we use will continue to be lighter, and more portable, reaching out to a more dynamic cloud computer for most tasks. The offline continuity and data privacy of local, personal computers will still be important, but most utilitarian tasks should be handled on more robust, dynamic infrastructure. Having a cloud workspace will be like having an email account or using the electrical grid. It will be a place to host personal projects, execute tasks and research, organize documents, and manage automations—all things that you can already do on your laptop in some capacity, but to a much greater degree.

Private beta testers

Zo is currently in limited private beta. During this period it is free to use while we collect feedback. If you think you have a good regular use case for a product like this, request an invite using this link.

======

notes / examples / extras

The video example: User finds out that trying to put a bunch of videos on their website doesn’t work on Safari. With AI it’s easy enough to find out that the problem is that safari doesn’t support the specific codec and the solution is to use a tool like ffmpeg to convert it. But if you imagine the very very best way to accomplish this task, it goes like:

• Recognize the solution is to transcode all the videos

• Write the code that can do that

• Realize that each one takes ~10 minutes on regular hardware, and that there are ~50 files

• Recognize that a single GPU machine could do this in less than a minute, and costs about $1/hour

• Spin up the entire system on ~50 GPU machines (each starting up in ~3s), run the code to convert all of them across these 50 computers, save them to a shared volume

• Spin down the replicas, user sees all the new videos on their basic CPU machine

All of this might take 45 seconds and cost about $5. The alternative would be to fumble around with the computer, wait a long time, or maybe use a big service like Youtube to do it incidentally and custody the videos on their ad platform, which would still take hours to upload and transcode.

The point is not that you can get this specific thing done. It rather, illustrates that there are scenarios where consumers could really use some of the computing patterns offered by data-center compute, which large corporations use all the time to very valuable effect. Things like this won’t be common necessarily, but because they are possible at all means that consumer computing can look and feel quite different than it does today. Just as significantly, all of this can happen very quickly, just by saying “Hey, these videos don’t work in Safari, can you help me fix them?”.

Consider a small business owner who wants to automate customer onboarding, inventory tracking, or sales reporting: "Zo, every time I make a sale, log it, update my inventory, and send the customer a thank-you note." A Zo Computer conforms exactly to your needs over time, and any automation it has created in the past can be re-run immediately, or updated as things change. As the last two decades have demonstrated, software offers a lot of leverage on operational capability; now, if you can describe it, Zo can make it happen.