Ultraviolet Catastrophe

AI is about to make the public internet useless

The term “Ultraviolet Catastrophe” was created in the late 19th century when physicists noticed that electromagnetic theory predicted that the night sky should be white, not black. This was because classical theory held that things with a temperature above absolute zero (thus including matter seen by looking in any direction in the sky) would ‘glow’ with larger and larger amounts of energy at shorter and shorter wavelengths, meaning we should all be getting barbecued by hard radiation from deep space. It turned out that an early conjecture from the emerging field of quantum mechanics - the idea that energy had to be transmitted in discrete chunks - saved us from the sunburn.

Sadly, the design of the internet has no such limit: It allows anyone to send an infinite number of packets (the internet version of a photon) to anyone else, whether the receiver wants those packets or not. If I know your IP address, I can send you packets. This defect is behind the DDOS attacks that constantly plague all widely-known service providers: If enough people send packets from different locations toward your server, you get knocked offline and there is nothing you can do about it. This design flaw happened because ARPANET wasn’t made to be used by billions of people - it was made to be used by hundreds of military, government, and academic personnel, all of whom knew and trusted each other. If anyone abused the system, they could easily be taken offline by changing the routing rules… no one would forward packets received from them anymore. Danny Hillis gives a great TED talk with a similar message called “The Internet Could Crash, we Need a Plan B”, in which he actually brings onstage a modest-sized paper book containing the names, IP, and email addresses of literally everyone using the internet in 1982.

This critical flaw was then replicated in later layers built on top of TCP/IP: If I have your email address, I can send you as many messages as I like, without asking. If I have your social media handle or phone number, I can do the same.

Social media, with it’s ability to create accounts which cannot easily be tracked back a responsible party who can be subjected to penalties, has been an early-warning demonstration of the consequence of having no cost for sending messages.

But, until now, the creation of a widely-broadcasted message has almost always required human authorship. Someone, somewhere, has to type the message (or record the video) that is to be sent. That means there is still a substantial minimal cost in human effort to create a believable message. Unfortunately, those days are over.

AIs are now easily able to create accounts, establish IP addresses, and send plausibly important or interesting messages at effectively zero cost. This means that the internet will soon be saturated with websites, social media, YouTube, and Twitter accounts creating messages at millions of times the rate at which humans can create content. Worse yet, many of these messages will deceptively claim to be from real people. If you have elderly parents, you are feeling it already. You are already thinking about when you will have to disconnect their computers to keep them safe, right? But you may be also suffering from the fantasy that these messages for some reason won’t fool you. In sufficient quantity and quality, they certainly will. And even when they don’t, they will “flood the channel”, making it impossible for you to find actual useful information.

An example: As a tech executive, I now receive dozens of messages a week from people asking whether I’d be interested in selling one of my companies. These messages come by email, LinkedIn, and even SMS. When I, somewhat out of morbid curiosity, explore the people and companies behind these messages, I find a mirage: LinkedIn pages for financial service companies with recommendations and founders even with vacation pictures on Instagram accounts… that on a bit of closer examination don’t exist. They are completely fabricated - pictures, friends, and all - by AI. As a result, the internet is no longer a place where I can receive messages from legitimate business people wanting to talk about M&A (sorry if you are one of those people to whom I have not responded). Fake people trying to buy companies are canaries in the coal mine… we are seeing them first because it’s still a bit hard to setup these types of AI-powered scams, so they are targeted at high-value transactions first. But within a few months, you’ll be getting video messages from college friends you haven’t talked to in a while wanting you to meet them later today to try out the great new gym just down the street, and so on.

How will we respond? At a high level, we will simply have have to stop using open/public services like email, TikTok, YouTube, or social media which have a low cost to send messages. Is this a bad thing? I’m not sure. We’re already overloaded with low quality information, why not just tear off the band-aid? In a couple years we may look back and laugh at how ridiculous it was that we tolerated spending so much of our days skipping over solicitous junk.

What seems likely to rise from the ruins of Internet 1.0 will be a different low-level architecture, in which IP addresses are replaced with encrypted channels. You can’t send packets (or messages) to a destination address. Instead of giving someone you want to stay in touch with your e-mail address or phone number, you will touch your phones together and create an encrypted channel that both of you can use indefinitely to send messages to each other. Messages on these channels will be stored and forwarded by routers in a manner similar to how they forward packets today, but you won’t be able to send a message to a channel without knowing its encryption key. It’s a bit like how you pick a specific channel on a two-way radio to communicate with a group of friends while skiing. There are already prototypes emerging, like activityPub (used by Bluesky) and and NOSTR, that do something quite similar: allowing people to create and use public channels that are forwarded between servers. You subscribe to the channels you are interested in, and receive messages only for those channels. If you stop checking the “mailbox” (stop subscribing to the channel), you stop getting messages. These channels can be 1:1 or used by huge groups (as with Bluesky).

Summary: AI is about the flood the internet with messages and render it useless for many tasks. To scale to billions in a world filled with AI agents capable of typing 1000 times faster that us means a complete overhaul of our aging internet architecture. We will have to turn off anonymous public services, replacing addresses with channels - or in the language of graphs - replacing nodes with edges. There will be turbulence during this process, so fasten your seat belt and find some good books to read during the down-time.