I wish we had real cloud computers. I want to be able to write some code and deploy it to the ether, without having to maintain it. If other people find the program useful, they can tap in and run my code without me having to manage the application's uptime, bandwidth, and costs. The less control I have over it, the better. In fact, I'd love if the mere act of uploading the program also relinquishes it from my control, placing the burden on the users who will operate and use it. And, importantly, I'd like this system to be a web primitive—something that lives in a shared and open spec, like the way that HTTPS and HTML works, rather than relying on a single private company to manage everything (*cough* Amazon).
One analogy I can think of might be a network of floating vending machines. You spend some time crafting and coding your little machine, and then you launch it up into the sky, connecting it to the rest of the "cloud network" and making it discoverable to the world. People might use your machine to perform some specific type of computation: they put in some coins to execute the operation, and get back a result based on their inputs. Perhaps you, as the creator of the machine, collect a small fee on each transaction as a means of monetization, or maybe you've gifted this machine to the world as a free and public domain utility that is permanently accessible to all.
Some of the open source tools I've built over the years are entirely client-side, like spectrum, a real-time WebAudio specturm analyzer. This kind of app is pretty close to my vision. In 2016, I wrote some code, deployed it to a static HTML website, and it's still running and providing a public utility several years later (in Chrome, at least) despite me not having to maintain or pay for it. The deploy was about as painless as possible: a single terminal command, thanks to surge.sh (which, let's be honest, is probably just a thin wrapper around Amazon's services).
There's a few gotchas with this approach. One is that it's incredibly tedious to create graphical user interfaces (GUIs) compared to command-line interfaces (CLIs). This is why most of my tools are CLIs, like image-sdf, which takes an image file and converts it to a signed distance field for resolution-independent rendering in games and graphics.
There is also a cost to hosting static websites, although it's usually pretty negligible. But, over a long period, or with particularly popular applications, this cost might exceed what the host is willing to pay. I had to nuke Audiograph as its bandwidth was consuming nearly all of my free Netlify tier. As the annual hosting and/or domain costs become too much to bear, most hosted websites have a tendency to close up shop over decade-long periods, leaving behind a graveyard of 404s and broken applications.
Of course, there is also a limited type of application that can be run entirely client-side. This has become clear with the recent AI boom, dependent on large language and latent diffusion models. Most users do not, and will never, have the hardware required to render the latest and best AI models, and so most computation is concentrated to and owned by a handful of major players (OpenAI, Stability AI, Microsoft, Adobe, Google).
I recently published a free Figma plugin that crops and rasterizes layers into a bitmap, to better optimize PDF exports. The process of developing and publishing this plugin ticked some of the boxes I'm looking for: it was quick to code, easy to publish, cost me nothing, and I didn't have to muck about with building a complex and rich user interface just to support a "core" set of features that most users have come to expect in graphical applications (e.g. managing images, zoom, history). It's also monetizable: if I really felt like this product was worth selling, I could gate it behind Figma's payments API, which might provide an incentive for me to continue to maintain it.
Unfortunately, this is a rather centralized system, and works only within Figma's walled garden. Ideally, some parts of this vision could be brought outside of the scope of a single company and product, and into a sort of open source "App Store" on the web. Yet, the web protocol does seem to be gradually getting there, turning into a massive operating system that supports a wide range of features previously impossible without an expensive web server (FileSystem, WebCodecs, WebGPU, WASM).
An obvious question here is: who pays to host all of these "machines in the sky"? Simply: the commons. Like a vending machine, the user must first insert some coins to operate the mechanism. If the machine runs out of coins, it doesn't magically disappear, like a web server or domain typically would. When somebody wants to use this tool, they might have to deposit some nominal fee as part of their interaction (such as 5¢). Unlike a real vending machine, this digital system could be topped up by anybody else in the network. Maybe the author decided to release the tool with an initial deposit of 5,000 free uses, or maybe a dedicated entity (e.g. Figma) has an interest in keeping the tool free for their top tier of paid subscribers.
These systems also do not need to be built in a way that extracts more rent than necessary. If you deposit $10 USD into your account, but only use $1 credit for an operation, you can decide to withdraw the remaining $9 if you no longer need the tool.
The next question might be: who runs all of this compute? Again, the commons. Any user could become a validator in this distributed compute system, providing compute in exchange for some small fee of transactions or other reward structure. This is fairly similar to Folding@home, except with a stronger built-in reward and incentive program—to create a more sustainable network of operators—and hopefully much better UX.
Some might also wonder about the privacy, security, and fault-tolerance of such a system. If a network operator has control over what is sent to the user, they could lie about their computations or attempt to attack the network for some financial gain. A distributed system like this would depend on the latest advancements in cryptography, verifiability and consensus, which are still a long way from achieving high performance and ease of use. Ongoing research in zero-knowledge proofs, fully homomorphic encryption, and consensus shows us a possible path forward.
Many of these thoughts occur to me as I attempt to generate an image with a large AI network. This is a task that, theoretically, might be possible in a future distributed system for the cost of a few cents, and without needing to share any private or personal data. Meanwhile, here I am, signing up for a $12 USD monthly subscription to Midjourney, and sending away my home address and credit card details, perhaps just for this image.
Large models will probably be the defining technology of the next decade, and I am concerned with the sharp concentration of revenue and data-collection toward a small number of private monopolies. I'm also noticing how subscription models for APIs are beginning to change on the web, perhaps in reaction to data collection and anti-spam policies brought on by LLMs (see: Reddit, Twitter).
These concerns have led me to blue-sky thinking: not that I imagine the system of distributed compute I'm describing here is practically feasible in the near future, but in that I hope it one day could be.
Aside from the AI-generated image produced by Midjourney, this post was created and written by the author's own hand.
