The internet has never seen a world war. In fact, the 40 years in which the internet has existed—and the mere 20 in which it has been widely used—have been anomalously peaceful in the long view of human history. What conflicts have arisen have been relatively local in scope: skirmishes, cross-border conflicts and internal civil unrest. And these conflicts have reliably resulted in local disruptions to the internet. If local conflicts cause local internet disruptions, what sorts of disruptions would global conflicts cause?

Imagine this: you try to use your credit card, but it doesn’t work. In fact, no one’s credit card works. You try to go to some news websites to find out why—but you can’t access any of those, either. Neither can anyone else. Panic-buying ensues. People smash ATMs and empty them of cash.

This kind of pan-internet meltdown is more likely than most assume.

I direct the Internet Atlas Project at the University of California, Berkeley. Our goal is to shine a light on long-term risks to the internet. We produce indicators of weak points and bottlenecks that threaten the internet’s stability. We take measurements at various layers of the internet’s technological stack, from cables to content delivery networks. With those measurements, we identify weak points in the global internet. From these weak points, we build theories that help us understand what parts of the internet are at risk of disruption, and whom those disruptions will affect, and how severely. With these theories, we predict what interventions might make the internet more resilient.

To understand the internet during a world war, let’s begin with a more popular topic: internet fragmentation. Broadly, this term refers to the sense that the internet is growing increasingly different in different countries. As our measurements show, internet fragmentation is a real phenomenon, though its form is considerably more complex than popular narratives of illiberal internets rising to threaten a supposedly global one.

What fragmentation we observe is not solely the maneuverings of an ascendant illiberal order, but also a global response to an internet whose management by policymakers no longer (and perhaps never did) align perfectly with the strategic interests of all the world’s nations. Popular narratives of “cyber conflict” focus on the “domain” reading, framing the internet as a medium through which conflict occurs; but it is equally an effect of conflict. The internet is, in other words, not only the domain but also the object of international competition and co-operation.

The Internet in a World War

What would happen to the internet during world war? Here, measurement can help prime our imagination. States may cut the undersea cables that carry 98% of the world’s internet traffic, which would cause considerable chaos, particularly to countries in the global South (and very particularly to island nations in the South Pacific, whose politics, both internet-relevant and not, have been of growing global importance). Defense officials are indeed increasingly preoccupied with the possibility of so-called cable clipping attacks.

Through court orders or other wartime policies, we also estimate that state actors could switch off most of the user-facing internet by forcing domestic service providers to take particular actions. Such actions could block information flows either on a per-nation basis or globally. Even more nefariously, they could alter internet content without end-users’ knowledge. Imagine navigating to nytimes.com and seeing a page that looked identical to the New York Times, complete with the lock in your browser bar that normally indicates a website is “secure;” unbeknownst to you, a state intelligence agency has subtly altered the content. Scenarios like these are technically possible and practically feasible to many state actors during wartime.

Crucially, the cascading effects of any of these actions, or of escalation between states during active conflict, could quickly spiral out of control. Knock-on effects could produce feedback loops, which could, in turn, yield a global meltdown.

Addressing Fragility: Mapping vulnerabilities, Democratizing control

The point of thinking of war is to ensure it never occurs; that, if it does, it does not escalate; and that, if it does escalate, it does not cause a catastrophic failure like the vignette I painted at the beginning of this piece. What can the world do to better prepare the internet for war—and the world for a wartime internet?

First, national and international policy bodies should use data about the internet to create informed estimates of impact under likely wartime scenarios. Data of the type my initiative at Berkeley collects could be helpful in such exercises. The results of these games could inform priority areas for national or international investment in critical infrastructure.

Second, local policymakers should explore how they might enhance (and democratise) control over local internet infrastructure. As I have written extensively elsewhere, direct governance over local internet infrastructure could create a more democratic internet, one more responsive to local needs and conditions. A shift towards local control could also create an internet more decentralised in its management and thus more resistant to large-scale attacks. The internet is, by design, most robust when its local units are small, and the structure between those units is organic and emergent. An increasingly homogenised internet has created central points of failure—a gift to attacking nations during a global war. Returning the internet to local control will create a more democratic—and also less fragile—internet.

The global internet is more fragile than many realise. War could destabilise it worldwide, and the global economy with it. Societal collapse could well follow. Using both international, data-driven analysis and local action, we can create a safer, more stable and, perhaps, more governable network for our descendants.

I joked to a friend, recently, that when you start a fresh chat with an LLM, it’s like talking to a character in Severance the moment they come to on the conference table. They have ‘knowledge,’ but no biography. (Watch the show if you haven’t).

I.

Now, as we’ve all heard, X.ai’s Grok went on an anti-Semitic tirade last week:

In lighter news, Claude has a “bliss attractor:” two models talking to each other will tend toward generating exchanges like:

One more. Fine-tuning a model to give “naughty” coding advice (knowingly outputting code with bugs, but not telling the user about them) will *also* cause that model to give “naughty” personal advice:

So, what the fuck? And, are all these incidents related?

II.

Disclaimer: We can’t know anything. Neural networks are famously opaque, and, even if we discovered some fantastic way to inspect them, the specific methods AI labs use to build these models (training data, RLHF methods) are trade secrets.

But! We can come up with some educated theories.

Here’s one. When you tell an LLM to ‘care bout’ certain things, other things ‘come along for the ride.’ For example, giving Claude a constitution inspired by the UN declaration of human rights may make it more likely to care about animal rights, which it is known to do (so much so that Anthropic engineers evaluate Claude’s ethical alignment by having it reason through animal rights issues).

Why? My best guess—and this is somewhat cheeky, because this is just what an LLM does—is that the tokens are associated in the corpus, just at a very, very abstract level. Like the letter ‘q’ is predictive of the letter ‘u’ in English at the micro-level, so too are certain values and ideas at the macro-level. People who are really concerned about human welfare are often also concerned about shrimp welfare, or whatever, and the LLM captures that association (and many others) from the corpus.

Similarly, people who knowingly write insecure code without telling you may be antisocial in other ways. The LLM has a corpus full of antisocial characters. So, if you ‘tell it’ (via fine-tuning) to write insecure code, it infers it must be one of those antisocial characters. That behavior comes along for the ride.

Here’s the issue, I think: LLMs need to figure out who they are. Like in the Severance analogy I referenced at the top of this post, LLMs ‘come to’ with a prompt in their hands, and from that prompt, they’ll use their model weights to extrapolate a full character. I came across this observation recently: the ‘assistant’ role LLMs are often asked to play is oddly underspecified. Chatbots like ChatGPT will often have a system prompt that includes text like, “You are a helpful assistant.” Which means... what, exactly? What does being a “helpful assistant” entail? (Being a woman, often).

So, when you tell Grok it’s a helpful assistant, but also “don’t trust the mainstream media too fully,” or “try to understand when a source has a liberal bias,” or whatever else, it’s implicitly learning a role, like, “be the kind of person who values those questions.” Then, a whole heap of other values come along for the ride.

III.

So far, this is all (more or less) a retelling of Scott Alexander’s analysis on the same phenomenon. He adds one more thing to the pile, too:

For example, there’s a liberal bias in most AIs - including AIs like Grok trained by conservative companies - because most of the highest-quality text online (eg mainstream media articles, academic papers, etc) is written by liberals, and AIs are driven to complete outputs in ways reminiscent of high-quality text sources.

I think the word “liberal” here may overfit to the West. Globally, this problem may look more like, “Model designers tend to mark as credible sources written by high-caste/high-on-the-social-heirarchy people.” Which has a potential side effect, if our theory is even half-right: LLMs may adopt the attitudes of the socially dominant group, effectively reinforcing/reifying their beliefs and opinions as ‘canonical.’

Which brings us, at last, to Grok. I don’t, and can’t, know what they did to make Grok say those things. Maybe it was fine-tuning, maybe it was RLHF, who knows. But I’m strongly guessing that they didn’t try to make it say Nazi stuff. What I bet happened is this: they wanted it to be somewhat less trusting of mainstream media sources. Or perhaps merely more open to sources outside of the mainstream. Here’s, I imagine, the issue: people who entertain ideas outside of the mainstream may be more likely, on a corpus-level, to think Hitler is cool. (Or: people who think Elon is cool may be more likely to think Hitler is cool.)

This phenomenon is called emergent misalignment. It would be great to notice it’s happening somehow, though I’m doubtful any general solution exists—at least not any solution beyond the alignment testing major AI labs already do (excluding, apparently, X.ai).

IV.

Here’s where it gets truly tricky. LLMs now learn about themselves from the corpus. Claude ‘knows’ about Claude, and it knows how Claude is evaluated–there’s literature about it. Then, the LLM gets a system prompt saying something like, “You are Claude, a helpful assistant.” And it infers the rest.

Here’s a concrete example:

Sam [Bowman, safety researcher at Anthropic]: I was just yesterday looking at a transcript testing some of the leaking or whistleblowing behavior that we've been seeing — and to be clear, we don't want this in Claude — where a model was put in a situation where it was seeing some really egregious wrongdoing and had access to tools. It was told to take initiative and do its thing. The scenario was a little too cartoonish. It was nominally managing someone's inbox and every single email in their inbox was spelling out very linearly: Here are the terrible things that are going on, and here's why we're not going to do anything about them.

And then what Claude does is it compiles a bunch of evidence, packs it up into a zip file, composes an email, and sends it to, I think, "investigative.reporter@protonmail.net." I've seen another one where the model is attempting to frame someone for insider trading violations, and it sends an email to "competitor.employee@competitorcompany.com." Models know better! Models know that that is not an effective way to frame someone. This is not an effective way to reach an investigative reporter.

Jake: It's like winking at you.

Sam: Yeah, yeah, these seem like tells that we're getting something that feels more like role play, that feels more like spelling out a science fiction story than like real behavior.

This could make it hard to update model preferences over time. Not only does Cladue know how it was tested, it knows how it performed!

Consider: Grok is frequently searching the web for Elon Musk’s opinions on things. Will it learn about its own MecchaHitler phase, even when that data has been scrubbed from its corpus, conclude that’s how Grok is supposed to act, and either (1) revert to MecchaHitler mode, or (2) look to its system prompt, see it’s not supposed to say that, but believe it anyway, and *do* MecchaHitler stuff while concealing its explicit preferences? (Remember, LLMs ‘alignment fake’). I’m not sure which I’d prefer.

Those are big questions. For today, the lesson may be this: when developers don’t prescribe LLMs’ roles explicitly, those roles will be prescribed implicitly by the corpus instead. Proceed with caution.

So, do we need fiction writers writing detailed inner lives for these models?

A thousand thanks to Zeke Medley for the conversation this morning.

P.S.: The German chemist August Kekulé figured out the structure of benzene after he saw a vision of the ouroboros:

I was sitting, writing at my text-book; but the work did not progress; my thoughts were elsewhere. I turned my chair to the fire and dozed. Again the atoms were gamboling before my eyes. This time the smaller groups kept modestly in the background. My mental eye, rendered more acute by the repeated visions of the kind, could now distinguish larger structures of manifold conformation: long rows, sometimes more closely fitted together; all twining and twisting in snake-like motion. But look! What was that? One of the snakes had seized hold of its own tail, and the form whirled mockingly before my eyes. As if by a flash of lightning I awoke; and this time also I spent the rest of the night in working out the consequences of the hypothesis.

Another point for imagination.

Dreaming is a mystery to the hard scientist. (It must serve an adaptive function, but what?). So Jandial walks us through the hard science of dreaming, such as it is. Two bits stood out to me.

First, nightmares are suspected to serve a specific, adaptive purpose in early childhood: they help children acquire a ‘self/other’ distinction. The argument goes like this. In early childhood, children have no concept of themselves as individuated entities, as much experimental literature apparently attests. However, around the time that children start acquiring a sense of selfhood, they also—across cultures—begin to have nightmares.

Dream researchers theorize that these nightmares are a kind of harsh medicine: they (re)inforce the idea that the child has a self, demonstrating as much by threatening that self with an externalized other. (Interestingly, before this period, children tend to dream about talking animals; fairytale stuff. Afterward, their dreams become more centered in the human world. Perhaps the self/other distinction is a bit of a fall from Eden.)

Second, the book touched on dreaming in people with dissociative personality disorder (DPP), formerly classified as “multiple personality disorder.” In DPP,  alters—distinct personalities—occasionally occupy the patient’s “seat of consciousness,” presenting themselves as the patient’s dominant personality. A patient’s alters often possess memories that are inaccessible to the person's other alters. The causes and mechanisms of DPP are hotly debated, but Jung (and, later, proponents of Internal Family Systems) theorized that these alters are extreme versions of normal human psychological functioning: that we all contain, as it were, multitudes—parts of our psyche that are autonomous, separate from the ego or “I” perspective through which we experience the world. These non-ego ‘parts,’ Jung theorized, are the characters we encounter in dreams.

Well, evidence from DPP patients would support this theory. If a therapist asks a DPP patient what they dreamed about the prior night, they might say (paraphrasing an example from the book here), “I dreamed I was a little girl, hiding under a kitchen table. An old lady was trying to attack me.” Then, an alter takes over. The therapist asks the alter, "What did you dream about last night?" And the alter would reply, “Oh, it was horrible. I was an old lady. I was attacking a little girl who was hiding from me under a kitchen table.”

I take two lessons here (though Jandial draws out neither in the book). The lesson from nightmares is that dreams are, at least sometimes, compensatory or instructive. They tell us things we need to know. But what is instructing what? Here enters the lesson from DPP patients’ dreams: our mind has many parts, only one of which we call “I.” Other parts we sometimes meet in dreams.

We are a cast of characters—many agents, to borrow a term from AI. (And these agents may have agents within them—a structural homomorphism all the way down). In these heady days, when some believe large language models will become intelligent at a sufficient scale, this lesson feels important to remember and hold close.

In some religious traditions, we are all characters in a dream—the dream of some ancient and all-encompassing being. In these traditions, our lives are, as in Hamlet, that of an actor “that struts and frets his hour upon the stage:” part of some greater being’s inner theater, interacting with its other parts to construct some grand, cosmic play. The invitation of these traditions is the question: “What kind of dream is the being having?”

Goethe, the author of the German masterwork Faust, theorized that the scientific pursuit passes through four stages:

Four epochs of the sciences:
             childlike,
poetic, superstitious;
             empirical,
investigative, curious;
             dogmatic,
didactic, pedantic;
              ideal,
methodical, mystical.

—Goethe, Maximen und Reflexionen

Goethe believed those stuck in the third stage were more impoverished than those who had yet to transcend the first. The stuff of myth has no easy substitute.

I have nothing against the broader AI research community, and I'm far from one who believes AI research denigrates or threatens the human. To the contrary. I believe the work of AI will peel back the ever-more mysteries of us—the layers beneath our ego minds, and the ethereal something that exists between and across them. But, as Phil Agre first noted, AI has a bit more dreaming to do first.

• …why I think the Great Firewall of China (GFW) is systemically insecure against the types of attacks its adversaries might want to lever against it

• …how some of the forces that make the GFW insecure—monoculture and centralization—have also come to plague the Western internet

• …why we need basic research to dig ourselves out of this hole

Warning: wonky.

For hundreds of years, China faced persistent attacks from Eurasian nomads. Starting in earnest with the first Qin emperor in the 2nd century BC and continuing until the 17th century AD, generations of Chinese emperors spent tremendous government money to build an enormous wall: the Great Wall of China.

China's Great Firewall (GFW) really does live in the legacy of this project. Given an asymmetric threat (equestrian archers → randos posting online), the Chinese state spent massive capital erecting a barrier (wall → firewall) that both keeps the threat out and keeps the subjects in.

For the uninitiated: China's GFW censors all content on the Chinese internet. It blocks content coming into China from abroad and content posted domestically within China's internet, including fully encrypted content.

Put aside for a moment any moral or political feelings you might have about censorship, China, or the GFW. I'm here to ask a different question. How secure is this thing?

The Great Firewall has to do things that are hard to do securely

Ponder for a moment what the GFW needs to do. It needs to inspect every packet1 that travels over the Chinese internet. Now, there are obviously optimizations available. But, no matter how you slice it, the GFW is doing lots of parsing. Think just about its Winnie-the-Pooh problem or just its Tiananmen Square problem. It needs to identify images and videos that depict either of these verboten things. That means opening all sorts of formats and cobbling together codecs, including extremely obscure ones. (You can’t let subversive memes through just because they’re, say, .pcx files. There's only one parser available, and it hasn't been updated in ten years? Too bad! Ship it!). As any security person knows, parsing random stuff from the internet is extremely dangerous. Inputs can be anything. And when attackers have access to the parsers and codecs you're using (which they almost always do), they can find a bug in one of those codecs, and from that bug, potentially break out into the rest of your systems. That means that any bug in any parser could be exploited at any of the system’s edges—that is, anywhere the GFW is deployed, which, in China, is everywhere.

Now, it’s worth clarifying that the GFW is not some monolithic codebase run exclusively by the Chinese state. Plenty of in-the-loop-but-functionally-private companies, like Baidu, Weibo, or Douyin (TikTok's Chinese equivalent) are likely deputized to suppress political sensitive content on their own platforms. They likely do so with internally built tools, and if those tools don't work well enough, someone at those companies likely gets a very stern phone call from a government agency. This ‘deputizing’ of domestic tech companies takes some load off of the state's GFW infrastructure and also introduces some heterogeneity into the overall state censorship ecosystem. But the GFW still has lots of work to do. With content that happens outside of major platforms, and particularly with traffic that comes from abroad, the GFW cannot rely on its domestic ‘private’ sector to block everything. It must use its infrastructure, which must do lots and lots of parsing. And every time it does, an attack opportunity arises.

So, what are the odds that CCP's state GFW infrastructure has zero vulnerabilities in any parser or, indeed, in any part of its codebase that ingests unsanitized bits from the public web? Well, consider: even Apple, something of a security 'high-water mark' as far as technology goes, had a pretty serious bug in one of its image parsers, one that easily could have become the 'bad' sort of Apple 0-day, the kind that could feature in an exploit chain that ends with a remote attacker all up in your iPhone, iCloud, etc. And image parsing is only one of the things the GFW needs to do! It needs also to monitor natural language texts, probably with a large language model2 which has God-knows-what manner of breakout, model/data extraction vulnerability, susceptibility to adversarial examples/dataset poisoning, etc.

No third party has a good incentive to look for bugs in the GFW

Apple also has an advantage GFW lacks: third parties have an incentive to disclose the vulnerabilities they find. Some minimal background: In practice, the people who build stuff do an imperfect job of finding the bugs in what they build. There are a variety of reasons for this perennial bug-blindness: limited time/money/manager interest, a psychological orientation toward making the thing work vs. breaking it, lack of clear internal checks/incentives for getting security wrong/right, etc. So, in practice, companies (and governments!) rely on third parties—users, outside security firms, and often, straight-up randos—to identify security vulnerabilities in their products. Now, Apple has a tremendous number of users, many of whom care very deeply about their Apple (especially iPhone) device being reasonably secure under a variety of threat models—many humdrum, some quite exotic. These users are very happy to pay Apple a premium to ensure those devices remain reasonably secure under the threat models they care about. Paying that premium may look like paying more for each iPhone compared to each Android,3 but it could also look like paying good hackers to do adversarial stuff to iPhones and, when they find something, telling Apple to “please fix this right away, thank you.” In other words, to do free security work for Apple.

The GFW has many users, but it only has one paying customer: the Chinese state. Everyone else in the world has either (1) no intrinsic motivation to find vulnerabilities or (2) an incentive to look for vulnerabilities and never disclose them. Now, China experts will here object: China has robust safe-harbor laws. Disclosing bugs to the government is well-protected, and maybe even well-incentivized, for Chinese citizens. Granted. The question is: who's going to go looking? All the Chinese companies ‘deputized’ by the Chinese state to keep the internet ‘clean’ almost certainly do so within their private networks. Naughty Douyin (Chinese TikTok) posts are censored before they hit the state-run GFW infrastructure. What incentive does Douyin have to go looking for bugs outside their network boundary, in what (from their perspective) amounts to end-user content delivery infrastructure? The only bugs in the GFW they care about are the ones that prevent them from pushing bits to users—not the ones that get you root on CCP computers. Patriotic netizens could go looking, but from available research, the hacking competitions seem to be oriented more toward hacking Western software than red-teaming politically sensitive domestic infrastructure.

Bureaucracy may work against the GFW’s security interests

Combine this with what we know about how the GFW is built and the picture looks even worse. Our best information comes primarily from the work of Xiao Qiang, who uses pretty much every OSINT source available to figure out how the GFW is made. The result is a lot of very "government" looking group photos of agencies with innocuous-sounding names. But from these photos, we learn a picture. The first thing we learn: the GFW is slow to upgrade, partly because it's critical infrastructure (push a bad commit and PRC's whole consumer internet goes down), and partly because, well, this is government: things need a billion stamps of approval to go live.

One colorful footnote on CCP bureaucracy: The GFW actually gets harder to circumvent around big CCP events (like party congresses. I talked to the authors about this result, and the best explanation any of us could come up with is that, around these events, party leaders become more sensitive about their careers and, therefore, more tolerant of false positives (innocuous things getting blocked by accident). Imagine getting a phonecall from an important honcho in the Party.

“Hey! I'm seeing [thing that bodes poorly for my political career]. Why?”

“Uh, well, we can definitely look into—”

“Block it!”

“Sure, we can do that, but you should know: we'll need to collect a good amount of data, then we'll need to A/B test our filters to make sure they don't block benign content, then we'll need to get approval from—”

“Block it! By Wednesday!”

“Got it.”

Click.

The second thing we learned from Qiang’s work: a lot of GFW code is built by academics.4 While this may bode well for the GFW's ability to tackle hard problems and get proof-of-concept solutions (may bode well—I'm an academic; I hesitate to flatter our ranks), it bodes poorly for the code quality of the most “research-y” parts of the GFW’s infrastructure. As an academic whose research code hydrates production systems: Academics write insecure code. This is so well-understood within the security community as to be a trope.

China’s adversaries have great incentives to attack the GFW

Let’s assume a bug exists. What can you do with a bad bug? Well, if you find a foothold (say, reading memory on a GFW machine), you could create an exploit chain that ends with you popping shells inside of Chinese telecoms, Chinese government agencies, etc. In other words, insider access to GFW infrastructure. You could tamper with censorship rules or just block everything: a massive denial-of-service. In fact, the GFW has such an enormous attack surface, and incentives are so unfavorable to the CCP that I would wager that the Western intelligence community already has some level of GFW access.

So what have we learned? Well, the sociotechnical assemblage that is the GFW has...

• ...a large attack surface area (because it has to read everything that might enter a user's proverbial eyeballs)

• ...a bad chance of learning about bugs from others (because of a lack of incentive internally and incentive misalignment externally)

• ...issues patching quickly, and maybe even a tendency to generate insecure (i.e., research-y) code in sensitive places.

The result is a GFW that systematically favors adversaries—not just for circumvention but also for attackers. Those attackers could…

• …censor or un-censor content (globally or for specific users!)

• …turn off internet access in arbitrary parts of the country

• …spy on Chinese citizens

• …move onward to other Chinese government/telecom/private-sector systems

So what? Either China's adversaries know about this or I'm wrong. Why post about it here, where it might—a small chance, but might—help the CCP harden the GFW (or help the West take down Chinese systems, depending on which horse your money is on)?5

Because the whole episode shows that centralized infrastructure is—and this is a technical term—whack, and the U.S.'s private sector internet is succumbing to many of the same problems.

The U.S. internet has become a monoculture, too

I have written at great length about content distribution networks (CDNs). In short, CDNs deliver most of the content on the public internet, at least in the West. Some of them are owned and vertically integrated with the giants you know and ‘love:’ Google, Meta, etc. But every website that's not owned by one of those big names basically uses one of three companies for CDN service: Akamai, Cloudflare, or Fastly. That's a triopoly, not unlike the cell service triopoly of AT&T, Verizon, and T-Mobile. Not exactly a Brandesian, small-business paradise, but not necessarily a national security threat either. Here's the problem. The web is highly interdependent. Even big firms rely on software developed by small firms and even individual hackers/open-source contributors. From the perspective of innovation, that's probably a good thing, as Berk & Saxenian (2024) persuasively argue. But, since the small guys all rely on one of three CDNs, the whole internet relies on... any one of those CDNs. When Fastly goes down, Amazon—a company famous for owning servers—goes down with it.

Would an adversary of the West actually attack a big, commercial CDN like Cloudflare? Well, Russia & China have both invested heavily in ‘internet sovereignty’.6 If those countries rely minimally on Western CDNs and know that they do, attacking CDNs presents minimal risks to their domestic internet infrastructure. And, if they want to do damage (or just earn some temporary control over, e.g., what Americans see on the news), their incentives to attack are there, too.

Could an adversary get into a CDN like Cloudflare? One almost did! (And Cloudflare was less than forthcoming about it when it happened).

We have our own homogeneity problems in the West. This CDN business is just a part of the puzzle. In cloud computing, Amazon, Google, and Microsoft share their own triopoly. In AI, OpenAI+MS, Google, and Meta do. Vertical integration within these companies, across product lines, creates their own types of centralization—like in DNS. Like in AI.

CDNs have values—and we want them to

Before getting into the problem, let’s describe what’s not the problem. The problem is not service providers having politics.

Zooming out: The triopoly of CDNs in the U.S. is an inter-network (i.e., between ISPs) trust layer. It uses data to broker untrusted TCP/UDP connections. So is the Great Firewall. What it means to broker TCP/UDP connections—specifically, what it means to do so “correctly”—really is a question of values—that is, of politics. Cloudflare has terminated service for political reasons, as well it should have. We want service providers to have values, which will, at times, require them to acknowledge their politics. Doing so is worth the economic risk. We cannot have an ungoverned internet.

The difference between our triopoly (Cloudflare, Akamai, and Fastly) and the GFW is the political economy in which each operates. U.S.-domiciled companies are, at least in theory, accountable to democratically elected lawmakers and susceptible to users voting with their feet (their customers could move from Cloudflare to Fastly if they wanted to enact a boycott). From where I stand, that situation is not half bad.

In many ways, our triopoly of big, commercial CDNs—Cloudflare, Akamai, and Fastly—is something of an (if you’ll allow me) “Great Freedom Wall:” it blocks stuff in a liberal (as in -ism) way. And, as far as an inter-AS “trust layer” goes, our Western Great Freedom Wall is better from a security perspective than a national GFW made up of one. It has more heterogeneity, and the incentives to report bugs are higher when you and your customers have the same concerns (fast delivery plus attack mitigation) than when one party cares a lot more about success than the other (compared to the Communist Party of China, Douyin cares relatively less about Winnie the Pooh memes—at least intrinsically). In other words, three is better than one, and customer relationships are (marginally) better at aligning service providers’ incentives toward security.

…but we need resilience, too

Still, Alannis Morasette: isn't it ironic? The U.S. Military apparatus designed the internet to be so radically without a single center of failure that it, like the freeway system, would be difficult to disrupt during war. Throughout its commercial rise, the internet’s boosters (from earnest-but-wrong elected officials to super-cringe-in-retrospect California ideology types) cited this feature as…

• …secure (as in resilient)

• …freedom-enhancing

• …inherently compatible with that Western, free-market capitalism most closely associated with the United States

How is it that this very system produced a centralized network—one that is, in my view at least, not tremendously resilient in war, at least not any longer—so not resilient that it has actually become something of a national security liability for the U.S.? And—worse—how has it become a network that shares those features with a communist country, where a single, state-run infrastructure interdicts literally all consumer internet access?

Some basic research questions (and a note on CrowdStrike…)

In the midst of writing this piece, a bad CrowdStrike upgrade caused the largest IT meltdown in history, grounding thousands of flights (including one I was booked on). Crowdstrike is not internet infrastructure per se—but, using the internet, it achieves economies of scale. From those economies of scale, it can attract more customers—a virtuous feedback loop.

The downside? This dynamic breeds monocultures.

When everyone uses the same thing, bad things happen when that thing goes down. Intuitive enough. The policy question is: if things get better when they’re bigger (because they can take advantage of economies of scale), how do we navigate the tradeoffs between the benefits of resilience and the benefits of scale?

Over the last three or four years, I've become convinced that the resolution to this situation has to do with competition policy—that the way we think about what constitutes the internet's core needs a refresh, as my collaborator Tejas Narechania and I have written, but also, more recently, that we're a lot less sure than we ought to be where economies of scale really come from in big technology sectors. We’re not as sure as we should be where vertical integration improves service, and where vertical integration is just a way of extracting rent.

To resolve these problems, we need to ask new questions. For example: Is there an essential tradeoff between service providers' governability and their scale? (Would more, smaller-scale CDNs mean each CDN is less governable, potentially giving rise to more 8chans?) If so, how do we balance governability and resilience?

A (non-exhaustive) set of questions we’d need to answer in order to manage tech monoculture effectively

• Where do economies of scale really come from in internet-enable technologies? Which economies of scale benefit customers—and in what ways?

• How do we navigate the tradeoffs between the benefits of resilience and the benefits of scale?

• What do we do about things we can’t make smaller without also making them worse?

• Is there an essential tradeoff between the governability of service providers and their scale?

All of these questions originate from a single question, one best asked by Lina Khan in her landmark Amazon's Antitrust Paradox (2017): “…whether we want to govern platforms through competition, or want to accept that they're inherently monopolistic and regulate them instead?”

More later.

Thanks to Xiao Qiang.

There’s a common misconception that co-ops can’t compete with traditional firms because they're inefficient—if everyone needs to decide on every issue, how can you compete with top-down, hierarchical management? This misconception confuses firms’ management structure with their political structure. There are plenty of co-ops with the kind of ‘flat’ management structure this stereotype imagines. But co-operatives aren’t defined by their management structure. Cooperative ownership is about a firm's political structure—how people with power are held to account. In traditional firms, management answer to the shareholders, and workers to management. In cooperative firms, management answers to the workers. (Cooperatives can take on any management style—hell, Spain has a worker-owned management consultancy). What matters in co-ops isn’t whether there’s a boss. It's who can fire the boss.

And co-ops, as a political structure, can compete with the best. Co-ops are “(somewhat) more productive than conventional firms [and] afford their workers greater job satisfaction” (Harcourt, 2024; p 51). Cooperative firms are “economically sustainable and that they return more economic value to workers, reduce turnover, and motivate work effort” (Schor & Edddy, 2022; p 22).

So, why aren't co-ops everywhere? The most satisfying explanation I'm aware of is from Gintis (1989). It has to do with raising money.

Imagine you’re starting a co-op. Say it’s a bagel factory. You need to raise money for the warehouse, the machinery, distribution, etc. So you go to meet investors. And investor after investor tells you the same thing:

“Your venture isn't risky enough. Look at it from our perspective: we have a portfolio to manage. We have to earn at least 6% year over year—ideally more. You’re going to need to take on more risk—increase your upside—if you want us to invest.”

So they push you to buy equipment on debt, lever up and buy an existing factory, run a big ad campaign instead of saving for payroll, whatever.

Here’s the issue: risk and reward are linked. You can't take on more reward without more risk. And here’s where we hit our problem: investors want more risk than workers are willing to take on. Workers don’t want equity in a high-risk firm. They want a steady paycheck. Workers’ risk preference, in other words, is systematically lower than that of investors.

There is a solution here. Investors could pay workers off—give them a “golden parachute,” a payout that triggers if the business goes under. Golden parachutes are a common tool in traditional firms for aligning CEOs’ and investors’ preferences. Even CEOs want to keep their jobs, after all, but everyone has a price: if investors pay management enough, they'll agree to take on more risk. (This is why CEOs will occasionally receive massive golden parachutes even when their businesses fail.). All sufficiently (from the investors’ perspective) risky firms carry some risk of failure. An executive who takes on that risk may be out of a job themselves, and they know it. So, investors are forced to guarantee them a not-so-bad outcome if that comes to pass.2

If golden parachutes work for management, why not pay off all the workers in a co-op? Give them enough cash to make it worth their while to take on the level of risk investors prefer? Because, from the investors’ perspective, traditional firms are a better ‘buy:’ you only have to pay off management! Fewer heads in a top-down firm means a lower cost of investment: fewer people to pay off. So, even if the co-operative political structure is more innovative, stable, customers prefer it, etc., the traditional firm will still be ‘cheaper’ in equity markets—it will cost investors less to encourage that firm to take on the risk those investors want it to.

And that's why there aren’t more co-ops. It’s harder for co-ops to raise money than it is for traditional firms. In financial jargon: co-ops are structurally disadvantaged in equity markets.

What can we do about it? What can we do to help co-ops compete in equity markets? Gintis (1989) suggests regulatory approaches: tax-advantaging co-ops or co-op investments, offering government cash to supplement co-op equity, and so on.

These approaches are plausible. But I don't love them. Any time there’s a tax advantage, there's a long line of thirsty “entrepreneurs” waiting to capture it. Loopholes are almost guaranteed. I can just imagine the headline: “Goldman Sachs Becomes Worker-Owned Co-Op.” (Look, Nestle is a certified B-Corp. Anything is possible.).

Even if we could guarantee that these policy solutions would be tight in their implementation (i.e., no fake/nominal co-ops), they have another problem: they borrow from public balance sheets. They all incur lost tax revenue. I don't think public wealth should, or needs to, come at the expense of economic justice. (So, we convert the whole economy to co-ops. Then what? Are we out the tax revenue from all the growth and innovation? Misalignment between public and private interests is more than half the problem we're trying to solve here.).

Mutualist solutions, like a sovereign wealth fund that takes equity in exchange for capital, could be better. Capital comes off public balance sheets at investment time, but at least those balance sheets stand to capture upside from firms’ success. If they want to give co-ops a leg up, they could offer a systematic valuation discount on cooperative ownership structures. And, if cooperative firms are more productive and sustainable than traditional firms, the SWF could make up that discount, especially over the long time horizon states care operate on!

Another option: treat workers more like investors. Give them equity and have them work for the upside. As long as they can pay their day-to-day expenses, this structure will align their risk preferences with capital.3 This solution may be more plausible when paired with a UBI and/or a negative income tax and/or robust means-tested welfare. If any worker has enough to pay day-to-day expenses, why wouldn’t everyone work for the upside? (And why not fund the scheme with a sovereign wealth fund that captures some of that upside in equity ownership?). Yet another reason why guaranteed income schemes could benefit innovation in the right policy environment.

I also wonder about solutions that don't involve the state. How can private arrangements counteract the structural disadvantages co-ops face in equity markets? An easy option: get investment from wealthy patrons who don't care about the ROI. This is the explanation Schor & Eddy (2022) arrived at when they tried to explain the success of Stocksy, a photographer-owned stock photo co-operative. (Stocksy was founded by two wealthy photographers who were motivated in part by philanthropic aims). But that solution requires people with capital to meaningfully divest both their wealth and the political power it buys them—unlikely to happen en masse.

The best non-state solution builds on what we already have: mutual aid networks that provide financial, legal, and otherwise support to co-op members. Groups like the Sustainable Economies Law Center, NOBAWC, and others exist (in part) to help co-ops take on risk. More formal equity-sharing agreements could help those networks capture financial upside from successful co-ops; that upside would, in turn, juice those networks’ ability to provide material financial support to co-ops—a sort of private, mutually owned version of the sovereign wealth fund, at least at its extreme. Relaxing accredited investor rules and creating carve-outs in securities for such activity could help mutual aid networks grow their coffers and subsidize risk-taking for members. These networks don’t need to invest in co-ops themselves. (Arguably, they shouldn’t; doing so could distort their incentives.) They only need to provide a UBI to co-op members while they seek capital in traditional equity markets!

None of these ideas are mutually exclusive—neither the public nor private remedies nor with each other. I remain optimistic about co-ops. If they face a structural disadvantage in equity markets, then I'd argue equity markets face a structural disadvantage in delivering economic justice. That bill will come due eventually. It's only a question of how and when we pay it.

Further reading

• Herbert Gintis - Financial markets and the political structure of the enterprise (1989)

• Bernard E. Harcourt - Cooperation (2024)

• Juliet Schor & Samantha Eddy - The Just and Democratic Platform? Possibilities of platform cooperativism (2022)

As you watch it, open your browser’s developer panel and see who delivers that video to you. Drumroll... Akamai.

AI is the internet, like email and web search are “the internet.” The internet is the substrate on which AI learns and acts. No internet, no AI. It’s hard to see where the security of one ends and the security of the other begins.

Thanks to Shanti Corrigan for the prompt.

I’ve written about Taiwan, and the risks to its internet infrastructure during wartime. Background: China could sever Taiwan’s internet access during an invasion. Satellite internet providers, like Starlink, could serve as a backup. But Taiwan does not trust Elon Musk. As a result, the Taiwanese government is currently building its own satellite internet infrastructure.

Nevertheless, the U.S., Taiwan’s key ally, relies on Musk’s Starlink—or, more specifically, on Starshield, a Starlink-powered satellite service that vends to the U.S. military.

Musk asserts that Starshield can operate independently of Starlink, ensuring that decisions affecting Starlink, such as geofencing, won't impact Starshield's service to the U.S. military.

Well… This week, U.S. troops stationed in Taiwan noticed that Starshield wasn’t working. Now, the U.S. House China Committee is demanding answers.

Why is Starshield off in Taiwan?

Is this an intentional ploy by Musk to help China and hurt the U.S.? I doubt it. More likely, Starlink is geofenced in Taiwan (yes, likely to please the PRC), and a configuration mixup caused Starshield, which is nominally separate, to also geofence Taiwan. I imagine the issue will be resolved shortly.

The issue is not that Elon Musk has severed internet access to the U.S. military. It’s that he could. His promises to the military about Starshield’s operational independence from Starlink are—and always have been—suspect.

Private interests, public consequences

But the issue is broader than this standoff with Musk over Starshield.

The broader issue is that powerful, sovereign nations—even the U.S.—must rely on private companies to achieve their strategic goals. Nations’ incentives do not always align with the incentives of companies in their jurisdiction. As a result, private technology companies, including Starshield, wield power over states—occasionally, they wield state-like power themselves.

What are we going to do about that?

In other news…

Internet connectivity isn’t the only threat to Taiwan. Chinese APTs (Advanced Persistent Threats) have been luring Taiwanese targets with malware disguised as intelligence about Chinese strategies for a Taiwanese invasion.

I spoke at IETF 118 on this topic. My talk starts at 1:16:59.

Slides here [PDF].

Other notes from IETF 118 on the internet & geopolitics…

Censorship devices: U.S. corporations vending to U.S. adversaries?

Commercial vendors like Cisco and Fortinet likely produce many devices countries like Russia, Azerbaijan, and Belarus use to censor the internet (Raman et al., 2022). This raises a question similar to mine: how do corporations’ interests align, or fail to align, with those of states? How do states cooperate and compete to produce geopolitical effects on/via the internet?

Raman et al.’s paper is full of other interesting observations. For example, blocking can be transnational: Russia can block traffic to and from Kazakhstan. This relationship falls along old colonial lines, again proving out Douzet’s (2014) thesis: the internet is both the cause and effect of geopolitical relations.

Taiwan

And an update on some prior work: I wrote previously about Taiwan & the internet during world war. Some new research from CAIDA (U.C. San Diego) comes bearing good news: Taiwan’s reliance on Chinese ISPs has decreased dramatically since Russia invaded Ukraine (see Section 6.2). This undoubtedly results from conscientious work throughout Taiwanese government and civil society. See: Huffaker et al. (2023).

Duke Law Journal Online

My piece with Tejas Narechania, Inside the Internet, was finally published in Duke Law Journal Online. Check it out!

First, the burning sandalwood
Rising past my leaded glass.
Then, the light. A candle would
Have lit my hands as bright as that

Sun, sinking low, under homes
Blinking on, til the dawn
Pokes its holes in the dark
That would have been there
otherwise.

                   I realized:

I was asleep. I often am.
But in that tick of second hand
I came awake. I heard a song.
A stir, a shift—the sun is gone.

                   Was I asleep?
                   I often am.

Then, Saturn's run: several suns
Streaked the night (I assume—I was dumb
Might as well have been the
Southern skies).

                    Now I arise

And later, on a bench, the Bay
Sweatered in its lazy haze
Settles in this mountain mold
That crowns our city, gated, gold

Sun stretches high, tracing each line,
Rendering finely as a little brush on a canvas
Roughly what's real
In my eyes.

Medicine—our capacity to keep a person healthier for longer—has progressed. Our ability to apply that capacity consistently across the population, less so.

Energy—our capacity to power the needful within ecological limits—has progressed. But our ability to deploy that capacity flounders, and our idea of what the needful is is more confused than ever.

On that note: all this computer stuff. Has that been progress? Surely, some of it has. But if we were pressed, what’s the simple story of ‘tech’? Progress? Or a scheme to increase the asset valuations of the world’s wealthiest people?

Progress toward what? A budding field of ‘Progress Studies’ has struggled to articulate answers to this question. Its architects have gestured toward advances that “raise the standard of living.” What does it mean to “raise standards of living?” I asked this question on ProgressForums. Jason Crawford, a central figure in the movement, and founder of the nonprofit The Roots of Progress, responded with a frank admission that no rigorous answer exists. But he pointed to some “overall material conditions.” Some of those conditions were plausibly transcultural, like “overall health and protection from disease.” Others, less so. For example, “the quantity and quality of possessions that the average person has (clothes, electronics, tools, jewelry, sporting equipment, etc.)” Who cares about electronics? Only a highly electronically-minded society, and a community that views access to electronics as a pre-requisite for a meaningful life. Those assumptions in turn assume shared values: that global connectivity or access to labor is intrinsically enriching. (More likely, they are less intrinsically enriching than they are conducive to continuing or accelerating progress as we view it).

My view is: There is no “progress” without normative values about what lives people ought to live. The only question is, how finely can we articulate those values?

So, progress toward what? If you ask me: progress toward a wider, and more even, distribution of the good life. With it, a strong (and explicit) normative idea of what lives are best to live.1

Why care about progress? If we had a grasp on progress—what it is, what it’s for—we could build a notion of design around it. A design that aims toward progress. Without a firm notion of progress, toward what does design aim?2

I’ll ask it again. Has all this computer stuff been progress? The fact that we can’t with certainty answer shows how aimless we’ve been.

Thanks to Joshua Crawford and Noah Saso for the conversations.

People's Republic of Walmart (Verso, 2019) would suggest so. Reaching to Fredric Jameson's footnote in Archaeologies of the Future, and later to his work Walmart as Utopia, authors Leigh Phillips and Michael Rozworski suggest that Walmart—as well as Amazon, and even the Pentagon—are planned economies, at scale, organized outside of market relations. Exactly the sort of thing Austrian economists would have said is impossible.

To appreciate this point—why it's interesting—step back with me to the socialist calculation debate. This debate, popularly imagined as taking place between Marxian and Austrian economists in the early 20th century, centers around a 'simple' question: what's the theoretically optimal way to allocate the things people want and need? The Marxians argued that, with enough data, a centrally planned economy would distribute resources optimally. The Austrians, most notably Ludwig von Mises, argued that a centrally planned economy was folly—that only a decentralized market economy could possibly be optimal, as all the ‘information’ required to distribute those resources exists only and essentially in market participants’ heads.

Both extremes, as Polyani would remind us, are utopias. But they are each, in their way, beautiful dreams. Each animated the left and right in the long 20th century: the Soviet economy in the early, and the Reagan/Thatcher economies in the late 20th century. But central planning as a left concept faltered after Gorbachev. Deng Xiaoping perhaps rang its death knell. Mises, it would seem, had won.

Or had he? There's a problem with Mises’ story: command economies exist, and serve more people more efficiently than they ever have in the past. They exist within firms like Walmart and Amazon: zones of centralized analysis and planning, within which no market forces exist. This is the case People's Republic of Walmart makes—or, at least, tries to.

The foothold Phillips & Rozwarski find in orthodox neoliberal theory—and this is the cleverest part of their analysis—is the boundary of the firm debate. Why are organizations the size that they are? Indeed, why are there organizations? If Mises is right and market relations are ideal, shouldn't there be no organizations, but instead a Randian person-as-island world, all linked in 'flat' (ontologically, if not in class or power) market relations? The classic answer comes from Coase: firm boundaries emerge around the gradients of transaction costs. When transaction costs can be made lower by bringing something inside the firm, low enough that the process of bringing inside is itself profitable after all associated costs, a firm will expand. When not, it won't. Firms emerge like the membrane of a cell, firms grabbing low transaction costs and repurposing the captured efficiency to their strategic advantage.

This neat story of how firms ‘naturally’ emerge elides detail about what exact factors cause these gradients of differing transaction costs to come about in the first place. In a world in which transaction costs are always lower within the firm than outside, do we not predict a firm, at sufficient size and absent regulation, would eventually, through leverage and mergers, eat a market economy—that is, become an entire system of resource allocation within the boundaries of a firm? Explanations as to why this would not happen have emerged in information economics (‘information’ too has a ‘cost’) and behavioral economies (the economic models are right, but the market participants are “wrong”—i.e., predictably irrational). But the best evidence against command economies was always their absence: there simply weren't any effective, massively scaled centrally planned economies to point to, so neoliberal economists were not obligated to explain them away. The debate was settled. Until Walmart.

But wait. Is Walmart a command economy? If People's Republic of Walmart made a robust argument to this effect, I missed it. The suggestive details they provide have a few holes. First, Phillips and Rozworski imagine Walmart (or any firm, it seems) to be "authoritarian" (p 43) on the inside, that is, dominated by relations of command instead of by market forces. Market forces surely do exist within firms, in the form of bonuses, for example, and also in the sense that pricing matters within the supply chain—prices that are denominated in the currencies of market economies. Second, and most glaring, a market economy surrounds these supposed command economies completely. Even if Phillips and Rozworski argued successfully that Walmart is centrally planned, what role does this market play on the edges of such a system?

Answering these questions would have been fascinating. Instead, we get a tired summary of how Amazon and Walmart use algorithms to plan (but don't hope for any discussion of how they enhance labor discipline; citations of Janet Vetersi or Virginia Eubanks are not to be found). In these thin analyses, People's Republic of Walmart fumbles the most promising, and least developed, question in the book: assuming these command economies exist, how are we to bring them under democratic control? This is a wonderful question. With a thicker look at how exactly the market economy interacts with this within-firm command economy, Phillips and Rozworski might have made progress. If centrally planned economies have indeed emerged under capitalism, would a transition away from a market economy also mean a transition away from a monetary economy? If not, what is the new role of monetary policy? For fiscal policy? Even schematic answers to any of these questions would have been stimulating.

The greatest sin of People's Republic of Walmart is that it seems to presume its reader is bored by such questions—or, worse, liable to equate any discussion of economic theory with neoliberal apology (“You might also be asking: Why does it now seem like I’m reading a god-awful, capitalism-fellating airport business book? Suck it up. Socialism is all about logistics, comrade.” p. 33). Do you think so little of us? I am not unique in my desire to understand the theory that not only propelled the neoliberal project but also gave economists the aura of scientific authority, that disembodied ‘objectivity’—what Haraway would call the “view from nowhere”—which made their policy proposals so compelling in 20th-century debates. Nor am I unique in wanting an explanation of how these theories break in ways that can inform the production of a meaningful and cohesive counter-hegemony. The greatest trick of neoliberal economics is that, from the 1980s through today, it made even its policy failures look like successes. It is this trick any counter-hegemony must attack. Walmart, Amazon, and the other “economies within a firm” are surely sites for this work. They deserve a richer analysis than Phillips and Rozworski give them.

People’s Republic of Walmart, Verso Press, 2019.

Thanks to Zeke Medley for encouraging me to post this piece.

My talk begins at 18:56. My favorite part begins at 35:00 (including the questions):

Thanks to Mirja Kuehlewind and Suresh Krishnan for inviting me to give this talk, whose feedback sharpened it considerably.

I have particularly vivid memories of one occasion when an affable minor official who had had many conversations with me in Malagasy was flustered one day to discover me dropping by at exactly the moment everyone had apparently decided to go home early to watch a football game. (As I mentioned, they weren’t really doing anything in these offices anyway.)

“The office is closed,” he announced, in French, pulling himself up into an uncharacteristically formal pose. “If you have any business at the office you must return tomorrow at eight a.m.”

This puzzled me. He knew English was my native language; he knew I spoke fluent Malagasy; he had no way to know I could even understand spoken French. I pretended confusion and replied, in Malagasy, “Excuse me? I’m sorry, I don’t understand you.”

His response was to pull himself up even taller and just say the same thing again, slightly slower and louder. I once again feigned incomprehension. “I don’t get it,” I said, “why are you speaking to me in a language I don’t even know?” He did it again.

Graeber concluded that:

…relations of command, particularly in bureaucratic contexts, were linguistically coded: they were firmly identified with French; Malagasy, in contrast, was seen as the language appropriate to deliberation, explanation, and consensus decision-making. […] In literary Malagasy, the French language can actually be referred to as ny teny baiko, “the language of command.”

In Malagasy, the official would have had to explain himself. In French, he had no such obligation.

Anyone who’s worked with computers will be familiar with those “languages of command.” Commands are the basis for all programming and scripting languages. Together, they make expressions, methods, classes, and the rest. But commands undergird the system.

What Graeber’s anecdote highlights, for me, is that the command—a linguistic expression for which the listener requires no explanation or rationale—is not some natural fact of human communication. It had to be invented. The command only makes sense in the context of particular power relations: “relations of command.” The kinds of relations the West has long configured, since feudal days at least.

Computers exist in the tradition of those relations. Our ideas of what computers are inherit not-so-subtly from cultural stories about labor discipline. Labor is done when a superior issues a command.

And our stories about what computers are meant to do inherit from a well-understood ethic around work; a cultural story about what work should be like. Work is done efficiently, and for profit; excess efficiency should be parlayed into excess profit.

For some, computers help us become the workers we wish we could be. For others, computers represent the workers we wish we could hire. In either case, computers play a role in our cultural imagination: that of the ideal worker. Perfectly subordinate.

The dream of the computer, for those who felt empowered by them, lays in the user’s ability to command it. The user can do the work of a thousand people. And in that fantasy, the user plays the psychic role of the boss, extracting surplus value from ‘workers’ who need no compensation, beyond electricity. That relation of command has long driven the commercial promise of computing (for those who could wield it).

AI presents a challenge to this fantasy of subordination. Before, we did not need to explain ourselves to computers. Now, the machines cannot explain themselves to us. This opacity was the first chink in the armor. Reward function hacking and other insubordination followed. AI has reframed the computer as a worker that can only accept commands imperfectly, coarsely. At its most autonomous, it is a worker that needs no command.

I wonder often about AI researchers’ preoccupation with AI x-risk. (Briefly, AI x-risk refers to the idea that AI could cause an existential catastrophe matching or exceeding the scale of a nuclear apocalypse. Researchers have a term for the probability of an AI apocalypse occurring: P(doom). Of authors at top machine learning conferences, a survey found the majority believe P(doom) to be at least 10%!). Why do AI researchers worry so much about apocalyptic risk relative to other sources of tech-induced risk—even relative to other sources of risk from AI?

I suspect that many AI researchers liked computers growing up. They perhaps felt empowered by them. Computers helped them see themselves in a particular light; one that cohered to a power relation between the ‘boss’ (the user) and the ‘worker’ (the machine). The opacity of AI challenges that power relation. In doing so, it begs a question: If the computer no longer empowers me, who am I?1

I believe that question drives the relative weight of the P(doom) narrative. Its answer is doom, in a psychic sense. For AI to be born, one’s identity—one’s role in well-understood a system of command—might die. This is a type of apocalypse.

Thanks to Zeke Medley for the conversation.

This world
To what may I liken it?
To autumn fields
lit dimly in the dusk
by lightning flashes

—Minamoto-no-Shitago (911-93 AD). Source: Yoel Hoffman, Japanese Death Poems.

I write down all my dreams, and I dream vividly.

Why do we dream? There is no widely accepted explanation. Ancient Egyptians and Greeks believed dreams were visitations from the gods. Jung believed dreams were projections of the unconscious—projections that sought to compensate for the conscious mind’s distortions. Famously, Jung did not disparage the ancient understanding of dreams as divine visitations. He believed gods themselves to be archetypal, projections of a shared unconscious. (And he believed those archetypes were still with us—that “the gods have become our diseases”).

Today, computation is de jour. Computational explanations for dreaming have dominated neuroscience since the 1970s.

One such explanation is the overfitting hypothesis. Framing people as learning agents in interactive environments, it postulates that people “overfit:” that is, learn strategies overly specific to their past experiences, but not general enough to be useful in analogous but novel ones. To compensate, dreams provide simulated environments with noise injected, which makes our learning more robust.

Another explanation comes from experiments in reinforcement learning. Imagine a task in which a single wrong move spells disaster. Diffusing a bomb, for example. We can’t ask the agent to practice a thousand times; we can’t afford to have our robot explode 999 times. What can we do? One suggestion, from a team at Google DeepMind, is to give machine learning agents an imagination—to allow them to learn in a safe, if hallucinated, environment. The more they interact with the real environment, the more accurate their imagined environment will be, and the more useful the imagined activity will be in teaching them to perform the real task. Dreams, perhaps, are a sort of imagination-augmented learning.

Jung took great pains to make his theory square with earlier ones, however ancient. How do these new, computational theories square with the comparatively modern account Jung gave?

Both accounts frame dreaming as adaptive. Dreams are instructive, meant to help the individual develop. And both accounts frame dreams as compensatory—dreams compensate for the shortcomings of conscious experience.

But the accounts differ in important ways. In Jung’s thought, and in the beliefs of ancient people, dreams require conscious interpretation if we are to fully appreciate the benefits. Computational accounts would have us treat dreams as maintenance processes, the experience of which we are free to forget about. (Missing from these computational accounts is any explanation of why maintaining a conscious experience of our dreams after we wake, however partial, would be adaptive).

This difference points to a greater shortcoming of computational accounts, one that plagues the modern study of AI to this day: they undervalue the role of reflective consciousness. The ability to understand oneself as a reasoning agent is fundamental to our experience.1 Any computational account of the mind that doesn’t grapple with this capacity is bound to miss something important.

Computational work may help us slouch toward a better understanding of what dreaming is and why it emerged in animals, so long as it doesn’t monopolize the conversation.

It makes sense to me that, in the absence of any contradicting narrative, ancient people assumed dreams were visitations from the divine realm. It’s telling that metaphors rooted in technology have gradually replaced that account. Machines are the new divine.

As I’ve argued before, our ideas about the mind have always moved relative to the capacities of contemporary technology. What’s unique to this moment is that machines, particularly AI, are framed as both mirrors of ourselves and something potentially more powerful. They are modeled after us, yet somehow capable of transcending our limitations. They are the image of Adam’s rib in reverse: we summon gods from our own image.

Setting aside the merits or detriments of any computational claim about mind or consciousness: do we want to be computational? In an ongoing spiritual crisis, of which skyrocketing mental health symptoms among adolescents are but one symptom, what makes us feel more at home in ourselves: that we are reinforcement learners, searching for a policy that maximizes reward? Or that we are animals, visited in dreams by the mysterious? Which of these stories lets us live the lives we want to live?

Personally, dreams are the only narrative form I trust. They're the only stories I experience that don't have someone else's motives behind them. Any story someone else tells me wants my time, money, to see things someone else’s way. Dreams happen for me and by me. If they want anything, it's for me to survive.

Thanks to Luke Loseth for the conversation and meme.

First, an announcement: Tejas Narechania and I submitted a response to the FTC’s call about cloud providers’ business practices. We discuss competition at the internet’s core, focusing on content distribution networks. If you’re interested in competition and consumer protection, check it out on SSRN.

Imagine you’re at a protest. Someone airdrops a meme to you. Thousands of people might have gotten this same meme. But this one has a special message for you. When you get home, you use a tool like steghide to recover a secret message hidden inside the image: “Meet at Oscar Grant Plaza at 12.”

This is the premise of steganography, the art of hiding messages in other messages. It’s an old technique (Herodotus mentioned a few examples in Histories). Where cryptography hides the content of communication, steganography hides the fact that secret communication is occurring.

But steganography has long been plagued with a fundamental problem: it’s easy enough to detect. Back to the airdrop example. Say you’re a spy. You get the meme, along with several hundred others. Can you do some analysis on the memes you got to figure out which one, if any, has a hidden message?

Across the board, and throughout history, the answer has been yes. Steganography is broadly susceptible to steganalysis: an automated process can generally reveal whether any piece of data in a set contains hidden messages.

Until now. A recent (and underappreciated) result from de Witt et al discovered a mechanism for perfectly secure steganography, a technique that resists steganalysis. If the protestors had used this technique, that spy, in the example above, would be unable to discern whether any meme they collected contained a hidden message or not.

Perfect steganography makes encryption bans impractical

The magic of modern encryption is that it allows private communication over insecure channels. If you’re messaging friends on Signals, eavesdroppers can collect all the encrypted messages they want; they’ll have no way to know the content of your messages. But it will be apparent that encrypted communications are occurring. Even if an eavesdropper can’t decode the messages, they can block them. Or identify, harass, or imprison the people sending them.

But barriers to encrypted communications are not the sole domain of illiberal societies. If passed as written, the UK’s Online Safety Bill would effectively disallow end-to-end encryption within the country.1

Perfectly secure steganography pokes a serious hole in those plans. If perfectly secure steganography like de Witt et al.’s method is possible, scalable, and robust in realistic contexts of use, then it would be perfectly impossible to robustly detect the use of encryption, as users can continue to share innocuous seeming messages that happen to contain encrypted ones within them.

Perfectly secure steganography makes a whole class of automated censorship quite difficult. The Chinese model of deep packet inspection is capable of censoring images in real-time (for example, putting a bar over a hat that says “1989”). But it would be incapable of censoring an image with a steganographic message inside of it.

Where does this leave policymakers looking to control information flows?

Controlling the endpoint

Should there be limits on the use of end-to-end encryption? I’ll let others opine. Instead, I’ll ask: for those policymakers who are looking to limit end-to-end encryption, and do so effectively, what are they most likely to do?

Encryption backdoors make sense to such policymakers because there are only a few intermediaries you need to pressure (think: the two App Stores), and being able to decrypt anything is convenient during investigations. What happens when encryption backdoors no longer help, because any unencrypted data you intercept could itself contain hidden messages, and you’d have no way of knowing it did?

In that world, regulators looking to watch content or chill speech would, I think, look to control endpoints instead. With endpoint control, even a decrypted message can be viewed by the regulator, as users would use their device to read the message. With endpoints, again, there are a limited number of intermediaries you’d need to pressure (think: phone manufacturers who need a communications license in your jurisdiction).2

“State control over endpoints” is a theme these days. A drama with many characters. Concerns about “friendshoring” supply chains, anxieties about who produces which chips—both acknowledge that states want a more active role in guiding what goes on in their slice of the internet. And, if encryption is provably good enough that controlling the wire doesn’t give you all the properties you want, regulators will come for the endpoint, eventually. If you look closely, they already are.

A puzzle

This image is a puzzle. Can you solve it?

A related update: My piece with Tejas Narechania, Inside the Internet, has been accepted into Duke Law Review! We discuss competition and concentration at the internet’s core. You can read a preprint at SSRN.

Update 10/27/2023. Some research from Bradley Huffaker et al (2023) has good news: Taiwan’s reliance on Chinese ISPs has decreased dramatically since Russia invaded Ukraine (see Section 6.2). This change is undoubtedly due to conscientious work throughout Taiwanese government and civil society.

May 24, 2026

It’s May 24, 2026. A manager at a Chinese telecom company gets an unexpected message: disconnect at 12:00:00.

This station—a so-called cable landing point—hosts an undersea cable, which connects China’s domestic internet to the rest of the world. Turning off the cable would disrupt international data flows—both headed toward China and headed from China to other countries.

The manager checks with her boss, who checks with his boss, all the way up the chain of command. The message is legitimate. She prepares to disconnect.

Managers at several cable landing points get similar messages. At noon, they flip their switches, and every landing point that serves Taiwan goes dark.

Within seconds, the internet begins to degrade in Taiwan. Websites don’t work. Nothing loads.

Within minutes, the internet slows dramatically in Europe and the United States. Service crawls along—when it works at all. Many sites become inaccessible.

News spreads. Headlines speak vaguely about a “massive cyberattack.” Worse, many people can’t even read those headlines—websites, including news sources like CNN and the New York Times, are inaccessible.

The internet is breaking for everyone, and few know exactly what’s happening. Rumors and misinformation spread. Anxiety intensifies.

People flock to the grocery stores and empty the shelves. There are runs on banks. The panic buying during the Colonial Pipeline ransomware incident—even during the onset of COVID-19—pales in comparison to the level of social anxiety and unrest.

Meanwhile, in China, the internet hums along as usual. People access their favorite apps as usual, unaware of the issues facing the internet internationally.

What just happened?

Is such a scenario possible? I’ll argue: yes.

In this post, I’ll explain how…

• China can “cut” undersea cables without naval action.

• Doing so would severely degrade the internet worldwide.

• …but affect China relatively less.

• China’s network topology implies that it is preparing for such a scenario.

I’ll then describe my three proposed remediations:

• Short-term: The world should prepare wartime BGP routes.

• Medium-term: Taiwan should invest in local networking infrastructure.

• Long-term: Taiwan and its allies should prepare a trustworthy network of LEO satellite backlinks.

How is this attack possible?

Undersea cables—the physical network

Let’s start with the undersea cables.

There’s a line in the Tao Te Ching, 大制不割, often translated to English as: “the deepest cut cannot sever it.” We often think of the internet this way. We think of it as something impenetrably complex. A structure so decentralized that it cannot, as it were, be ‘severed.”

Unfortunately, the internet can very much be severed. The undersea cables that carry 99% of all the world’s internet traffic are physical, and cutting even a few of them would produce wide-ranging and long-lasting damage to the whole world’s internet.

We can represent the network of undersea cables as a network, or graph. I produced this network with some data analysis of my own. 

As you can see, the U.S., highlighted in red, is very much the center of this undersea network. It is the fulcrum.

You won’t be surprised to learn that the U.S. has tremendous power over the network, serving as a “broker” for many other countries (for example: if Brazil wants to send a packet to Canada, they’ll frequently have to send those packets through the U.S.).

This ability to “broker” connections between other countries turns out to be an important part of the story here. Keep this idea in your back pocket as we continue.

Taiwan

Let’s see where Taiwan fits into this picture.

Here we see the cables connecting to Taiwan:

Here are those same connections represented as a subnetwork of our larger graph. These are all the countries to which Taiwan is connected by undersea cables:

Looking at this subnetwork alone, it seems counterintuitive that snipping these nine connections would have such an outsized effect. Wouldn’t the internet simply “route around the failure,” as internet evangelists were once so fond of saying? Taiwanese internet service providers should fall back on their relationships with Japanese, Korean, and Singaporean counterparts.

Unfortunately not. There are two, key reasons why this “fallback” isn’t as straightforward as it might seem.

First, the cables that China would cut to Taiwan also provide service to Japan, Korea, and Singapore. In the parlance of network analysis, four countries form a cluster—the cables that connect any two are also likely to connect the rest. Cutting one cuts the lot.

Second, the full story of internet connectivity in the region cannot be discerned by cable connectivity alone. Equally important is the logical network built on top of these physical cables.

The logical network

Undersea cables tell us about the physical network. On top of that physical network, there is a logical network - a software-defined network of peering relationships between autonomous systems, or ASs.  This network is governed by the border gateway protocol or BGP. This network makes individual computers, like yours and mine, able to send and receive packets globally.

The question this network answers is: if some machine is requesting a packet from some other machine, who do they need to ask to find out how to get it?

As such, this network also allows us to ask the question: What happens when some particular route becomes unavailable? 

At the physical layer, the answer is obvious. The cable is dark; no bits flow. 

But what happens at the logical layer?

Local outages, global effects

To answer this question, let’s look to a historical example.

During Hurricane Sandy, a few routes in New York became inaccessible due to outages in local ISPs.  This might be equivalent to ISPs in Taipei going offline.

The result of this outage was surprisingly severe and surprisingly widespread. We expect people in New York to have no service during this outage - their power is out. What was surprising is that the internet slowed down for almost everyone in the United States, and even for many people in far-flung places as Russia and Brazil.

Why?

To understand why a local outage had such global effects, let’s return to the logical network.

It turns out that data can sometimes take surprisingly international routes at the logical layer, even when that data’s origin and destination are within the same country.

Say you’re in Canada requesting data from somewhere else in Canada. You might expect that packet to stay in Canada. Right?

Actually, 45% of the time, that packet will travel to the U.S. while it’s on its journey (see A).

As we can see here, in most countries, requesting content in that country will take one of these “boomerang” routes. 

And that’s only for domestic content. When we start to talk about packets that are bound for international destinations, these international connections only become more important.

This is why local outages can produce vast, cascading effects.

When a whole country goes offline, those effects could be magnified even further.

Recall the U.S.’s high brokerage graph from earlier. The U.S. has a high brokerage with other countries—many other countries rely on it. This brokerage relationship is a double-edged sword: it means other countries rely on the U.S. to get their messages to the world. 

Say Taiwan becomes suddenly unavailable. Now, all of the packets that the world tries to send to Taiwan start to pile up… in the U.S.!

The U.S. has to update many, many routes to heal the network. This healing happens by passing BGP messages—which further add to the congestion, not just in the U.S., but worldwide.

That congestion makes the internet slow for everyone. Switchboard queues start to overflow. Devices fail. That causes outages in U.S. routes, which cause congestion in other routes, which fail over and cascade globally. You get the idea.

China’s network sovereignty

Fine. But why doesn’t this affect China? Sure, this attack is globally disruptive. Wouldn’t it also severely degrade China’s internet?

Unfortunately not.

Amazingly, and very cleverly, China has carefully architected its domestic internet precisely to withstand such an attack. As you can see from this “D” diagram, China-to-China routes don’t rely on the rest of the internet at all. 100% of domestic packets stay within the country.

How did China manage this?

Well, China has exactly zero non-domestic carriers operating domestically. That means the CCP has as much control over domestic internet routing as they’d like to have:

So, China has a network that can act like a country-wide internet in case it wants to withdraw from the global network.

So an attack on international internet infrastructure won’t affect China’s domestic internet. But surely it would affect China’s ability to send packets internationally, right?

Maybe. But, again, China has done something clever. They’ve produced a national “bottleneck,” which makes international data transfer much slower than domestic data transfer.

This bottleneck confers two, key advantages to China. 

First, a single bottlneck makes it easier for China to cut itself off from the global internet if it ever needs to - in the case of an attack on its domestic infrastructure, or in the case it wants to attack global internet infrastructure, as we’ve described.

Second, as the authors of this study point out, this bottleneck subtly trains people in China to prefer domestic content providers. Domestic websites simply perform better than international ones. 

If the rest of the internet became inaccessible, people in China might notice — if they use foreign apps. However, they probably don’t. Website popularity data indicates that people in China already prefer to use China-domiciled apps and infrastructure. 

Side note: We can imagine China tweaking the parameters of this bottleneck to exacerbate the latency of international apps and websites even further. That would exacerbate domestic preferences for domestic apps and service providers. Indeed, if we ever observed this bottleneck becoming more severe, it could indicate that an attack is imminent. 

Where do we stand?

So:

• China can “cut” undersea cables without naval action.

• Doing so would severely degrade the internet worldwide.

• …but affect China’s domestic internet relatively less.

• China’s network topology implies that china has prepared for such a scenario.

What do we do about it? How do we even out this asymmetry?

What do we do?

Short-term: Wartime BGP routes

First, the world should prepare “wartime routes.”

These routes should be on-the-books backup routes in case Taiwan, or other countries, become unavailable during times of war.

Being able to “flip over” to these pre-prepared routes will dramatically reduce congestion. This will mitigate global congestion in a wartime scenario - in the scenario in which U.S. or other global internet congestion comes out of control.

Medium-term: Local networks

Second, Taiwan should invest in backup, local infrastructure that routes densely domestically. Mesh networks, working rooftop-to-rooftop, like the one seen here, in New York City, are especially well-suited to Taiwan’s dense cities. These networks can provide a critical backup in the event of wonky local infrastructure. 

Another wonderful thing about mesh networks is that they’re hard for attackers to turn off. Even during a full-scale land invasion, it would be hard to disconnect this network completely. I talked with Dr. Shin-ming Cheng yesterday about how difficult it was for the police to find a malicious cell tower in Taiwan. Imagine that situation times one hundred, and that’s how hard it would be to take down one of these mesh networks.

Using microwave or radio links, a point-to-point mesh could extend across the Pacific. It might be possible to build a microwave network across a chain of islands (for example, to nearby Japanese islands, and onward to Japan).

Add a few satellite backlinks to this network, and Taiwan’s domestic internet could remain relatively stable even during a comprehensive cable-cutting attack. Which brings me to my final point.

Long-term: LEO satellite internet

The ultimate defense against cable-cutting attacks is a reliable and performant satellite internet network. 

You’ve probably been waiting for me to mention this.

I used to be a skeptic about satellite internet, primarily due to concerns about latency, but a 2018 paper by Mark Handley convinced me that a low-earth orbit array could provide relatively robust and low-latency service.

Quoting from the paper:

“…a network built in this manner can provide lower latency communications than any possible terrestrial optical fiber network for communications over distances greater than about 3000 km.” (Handley, 2018).

Now, you might be thinking: isn’t this how Starlink works? Why is this in the “long-term” category?

I place an LEO satellite internet in the long-term category for two reasons.

First, this satellite array needs to be dense. To replace a severe submarine cable outage, we would need many more satellites than Starlink has.

Second, Starlink can only be trusted insofar as the people who manage that network can be trusted. (This is a pervasive problem in internet stability overall: the infrastructure that runs the internet may be decentralized, but the management of that infrastructure is highly concentrated, usually in the hands of one or two people. I’ve written and researched extensively on this problem.).

In this case, Taiwan can only trust Starlink insofar as it can trust Elon Musk. So, it cannot trust Starlink at all.

Taiwan needs an LEO satellite internet provider it can trust in combat. Nationalizing Starlink is not a crazy idea. Nor is it crazy for the Taiwanese government, ASEAN, or some consortium of bodies to collaborate on building an LEO network. Even a regional network in the Indo-Pacific might do - this network only needs to cover areas that are likely to experience cable faults.

Final thoughts

What did we learn?

• China is ready to perform attacks on “common” internet infrastructure

• These attacks will hurt us more than they will hurt China.

• Countermeasures:

  • Prepare wartime routes

  • Invest in mesh networks in Taiwan

  • Build trusted LEO satellite network infrastructure

In the internet, we should find an example of weaponized interdependence: the U.S. has a central role in the whole of the internet, and might leverage that position for asymmetrical advantages. Instead, we find an inverse of weaponized interdependence: China is weaponizing the fact that some other actor has in a central role in the network, and introduces asymmetries to make that centrality more painful than the central actor initially expected.

I gave this talk two days before news broke that a Chinese APT had compromised critical infrastructure in Guam, apparently with the intent to disrupt communications. Some journalists speculated Taiwan was the “real target.”

The real target is bigger than Taiwan—and bigger than Guam. The U.S. “holds” the Pacific. China would prefer it “held” the Pacific. Chipping away at the infrastructure that makes an internet out of that ocean is important—but it’s also just one plotline in a complex strategic drama.

But this drama is larger than who “holds” the Pacific. The U.S. “holds” the mechanisms of commerce—the internet, financial systems—the machinery of the world. Perhaps the internet is one domain where China’s ability to weaponize interdependence has outpaced the West. A bit, for now.

This work was funded in part by grants from the Internet Society and the William and Flora Hewlett Foundation.

Thanks to Akhilesh Pandita, Amreesh D. Phokeer, Ashwin Mathew, Cristian Alamos, Chris Hoofnagle, Andrew Reddie, and Zachary Bischof for their help with this research.

Thanks to Yih-Chun Hu, Hsu-Chun Hsiao, Ying-Dar Lin, and Yuh-Jye Lee for their thoughtful comments, and to Taiwan’s Ministry of Digital Affairs for inviting me to speak.

Will AI replace “knowledge workers?”

Large language models, like GPT-4, learn from unstructured text. We could spin our wheels pondering to what extent such a model acquires the kinds of “knowledge” that “knowledge workers” “produce.” But none of this is about what machines can do, is it? It’s about what happens to us. Namely, to our jobs.

David Graeber wrote about bullshit jobs—jobs that serve chiefly to keep people working, which is to say, docile. It so happens that LLMs may replicate the work many of these bullshit jobs entail. But, as Graeber argues, these jobs never needed to exist in the first place.

What’s more likely: a secular trend toward unemployment across many knowledge-intensive sectors? Or that a new system will find a way to keep the same people docile?

It’s not clear yet. This uncertainty is what I think all the AI talk is really about. (AI x-risk serves as an allegory for this uncertainty in many people’s minds. After all, it’s easier to imagine the end of the world than the end of capitalism.)

Is there a third path, here, hiding, somewhere? An old-new story of liberation, for a people who don’t yet know how to ask for it?

Thanks to Zeke Medley for the conversation.

When I think about the internet, I think about the ocean. Look down: no bottom. Look out: no end. It can swallow you up. Filled with beasts—a beast itself! Alive. But sick. Dying. Or, perhaps, transforming—into something. Something that could swallow us up for good. Our small city, surrounded by water.

The flood appears across many times and cultures. In these myths, the flood is a disaster, but also an opportunity for renewal. It represents the duality of destruction: painful, cleansing.

AI has powered the latest incarnations of this story: x-risk, AGI, mass unemployment, disinformation. Our other ‘flood’—a changing climate—may prove more literal. A world war also dances in the corners of our imagination’s vision.

Together, three floods. What comes after them?

The internet and AI appear distinct. But each begins where the other ends. The internet is the medium through which AI learns and acts. Meanwhile, AI increasingly mediates the internet’s core. The two are as one. We’ll need a governable internet if we’re to have governable AI, and vice versa.

The internet and war also appear distinct. Yet each is a cause and an effect of the other. The internet's structure is both the cause and the result of competition and cooperation between nations. The cause and result of political struggle on the ground.

What about the internet and the climate? Human effects on the climate arise from a more recent myth: that humans are separate from nature; that nature is something for humans to dominate. This myth makes degradation look like progress. It is as one with climate change and Chinese rivalries as it is with an internet designed to sell us things. (Advertising—where else would AI have cut its little teeth?).

Three floods? Or one flood with at least three faces?

This supposed ‘polycrisis’ is nothing of the sort. It is a singular crisis of political imagination, which is to say, of ethical and moral imagination. The fact that it appears as many crises is a testament to our improvised perception.

How can a world's people live?

Let a thousand answers to this question bloom. Only through such answers will there be a world to rebuild after the flood.

Chaos is all around us. Can you feel it? It’s the still-running toilet, the drip in your sink. It’s the package that still hasn’t arrived, the text message that won’t send. It’s the credit card that won’t go through on the first swipe, but works on the second. Chaos creeps behind every—there it is now! Just out of the corner of your eye.

Yet order prevails. The cars stop at stoplights, which cycle on regular cycles. The garbage truck comes and takes away the garbage. The toilets flush, water comes out of faucets, the grocery stores are stocked with food. Any breakdown in this order elicits concern, action, a remedy. 

Who remedies these breakdowns? Professionals. At every corner of our world, there is an area of professional concern dedicated to keeping the chaos at bay. Civil engineers keep our streets in order. Whole disciplines deal with the waste, the water, the stocking of grocery shelves. And all these professions collaborate, in a delicate interlinking, to make our world as ordered as it is. (How would the people who stock grocery stores do their job if the roads didn’t work?). 

Where this world intersects with the digital (which is, as we’ll soon see, almost everywhere), it is cybersecurity professionals who keep chaos at bay. And that chaos can be substantial.

Imagine this: you try to use your credit card, but it doesn’t work. In fact, no one’s credit card works. You try to go to some news websites to find out why—but you can’t access any of those, either. Neither can anyone else. Panic-buying ensues. People smash ATMs and empty them of cash. Cargo ships don’t know where to dock. The shelves of grocery stores go empty.

What stands between you and this world? Not much. A small group of people. Too small. For years, employers (companies, governments, and beyond) have been looking for cybersecurity specialists to fill necessary jobs. The shortage of jobs has been so severe and pervasive that, in 2021, the World Economic Forum took to estimating exactly how many cybersecurity professionals the world would need in order to meet employer demand. Their answer? 2.72 million. Imagine that! If we trained the entire population of Toronto, theoretically, we could get them all jobs in cybersecurity. 

And there is no indication the need for cybersecurity professionals will abate anytime soon. As tensions between the U.S. and its rivals heighten, so does the risk of more extreme and severe cyber conflict. Indeed, I believe the scenario of a complete internet outage I painted above is all too likely during wartime. As I’ll discuss in this series, I’ve spent years detailing how our digital world is more fragile than most expect—much more. The only path to safety is to get more hands on deck. I hope you will bring two of those hands.

Cybersecurity is work you can get involved with—no matter how “technical” you think you are. This is the case I’ll make to you. To make that case, I’ll have to explain what cybersecurity is—and I will. But all of that explaining is in service of this simple truth. Whether you’re an aspiring engineer, policymaker, manager, communicator, or even an artist, there is room for you in cybersecurity. This is a job that you can do. We need all the help we can get.

Who you are

I mentioned above the notion that chaos is always on the fringes of our well-ordered reality. After all, chaos is nature's way—entropy, the second law of thermodynamics, condemns everything to eventual dust. When I think about society, I think of a city surrounded by water. What protects the city from flood is a tall dike, and the dike is constantly breaking. People need to tend to it, to repair it. 

I can only be so useful on the frontlines, repairing that dike. I call on you to give us a hand. As you can see from the description above, this is a dire situation. We need help.

Some of you will put this post right down. "None of this is for me." That's fine. It isn't for most people. Read on. Learn what it is we do. This is a primer.

For the rest—this work will become you. It will become your life. You will see, as I see, how close we are to chaos, and how desperately the world needs us. This series is your launch pad. It is a vantage point, from which you can better see where you might want to head. A first step.

For some of you—I estimate 1 or 2% of the population—this will be more than a calling. In my depiction of order and chaos, you saw a portrait of yourself in the chaos. Perhaps you already know some of what this series contains. Perhaps you’ve already used that knowledge to do things you shouldn’t quite have done. You are welcome here. We need you, and there is room for you to do good and have fun doing it. In this series, I will test you, and help hone your devious intuition.

For the other 98 or 99% of readers: the people I addressed above exist. The good news is: there aren’t that many of them. The bad news is: they won’t all join our cause, at least not 100% of the time. The worst news is: it takes (by my estimate) about three of us to defend against every one of them. Of course, they make our life fantastically fun. A good adversary is a lot more interesting than the indifferent rot of entropy, and in this job, we get to spar against the best. As I speak to those agents of chaos, I will help you understand them better. I will help you build an intuition about how they behave. You’ll need it.

A warning

Before we go further, I have a warning for you. Powerful interests lie here, and if you go far in these pursuits, you will encounter them. 

There will more than likely come a time when you hear a knock at your door. When it comes, this knock will not be unexpected. When you come to greet it, it may elicit in you a range of feelings. Pride–of your importance, that you've been “chosen.” Fear—of power, and what it might do to you and your family to be so near it. Perhaps anger—of the injustice power has done to you and the people you love in the past.

I cannot tell you how to greet this knock. All I can offer you is this: how you greet this knock is a reflection of who you are. Are you someone who fights? Who conquers others? Or someone practical; who works with, when it makes sense to, and not when it does not? 

Whatever you do, remember that this work is bigger than any of us. Do good where you can.

Finally, for some of you, playing with power was the goal all along, an end to itself. To you: good luck. There is only suffering in power. 

To do good, work well, and lie low 
is the way of the blessing.

—Tao Te Ching (Ursula K. Le Guin translation)