Like everyone who has been paying attention, I have worries about the possible negative effects from AI. And it’s a long laundry list of worries. One, there is the technology’s anti-labor bias, which may accelerate our downward spiral into oligarchy. Next, there is the fact that the data center boom is crowding out investment in almost everything else, yet data centers employ hardly any workers. Third, I’m worried about the high probability that AI looks like a bubble, particularly given the sheer size of the AI boom. If it is indeed a bubble, it will inevitably burst and will probably plunge the U.S. economy back into a liquidity trap due to its size.
Oh, and Anthropic tells me that AI may well wipe out the human race, which is why I should invest in the company’s IPO.
But I guess I haven’t been talking enough to the technologists, because I wasn’t aware until last week of the strong possibility that AI will render the internet as we know it unusable — not at some point in the distant future, but in a matter of months.
It has already started. Reuters reports:
Government websites in Europe and elsewhere went offline on Wednesday after the disclosure of critical vulnerabilities in a widely used programming language sent website operators around the world scrambling to protect their systems.
Developers of the PHP programming language on Wednesday released urgent new fixes to vulnerabilities that had been lurking in their software for years until they were uncovered using AI, saying all users were “encouraged to update immediately.”
With tech firms warning that some hackers were already exploiting the flaws, government agencies across Europe were among those that took sites offline as a precaution.
Because a fix is now public and artificial intelligence tools are able to reverse engineer the bug with little effort, experts said they were bracing for a wave of intrusions as hackers jumped to get ahead of the updates.
AI was powering a “race to exploit,” said Costin Raiu, co-founder of cybersecurity research and incident response firm TLP:Black. He said the next step was the “mass compromise of every vulnerable web server on the internet.
And this is just the beginning of what Zvi Mowshowitz calls The Hackening. Two days later, Reuters reported that:
Advances in artificial intelligence have uncovered a stream of new software vulnerabilities that are almost certain to trigger more disruption along the lines of the outages that percolated across global government websites this week, according to seven cybersecurity professionals who spoke to Reuters.
It quoted one expert as saying “I think we’re going to have a couple years where every single week it’s going to feel like the sky is falling.”
Azeem Ashar, whose Substack Exponential View has been one of my main guides to what is happening with AI, calls what happened last week “The internet’s Wuhan moment” — as in the emergence of the coronavirus that caused the Covid pandemic, which killed millions and also temporarily plunged the whole world economy into a deep slump.
I’m not a technologist, but here’s how I understand the problem: Websites run on complex, multi-layered software that, because of its sheer complexity, inevitably contains security flaws that bad actors could potentially exploit. Worse, much of the software is old — PHP dates from the early 1990s — and was hence designed in an era in which modern hacking techniques didn’t exist.
And AI models, it turns out, are incredibly powerful hacking tools. Merely human evildoers may lack the time, patience and attention to detail to go over old code line by line, finding security flaws that may have gone unnoticed for decades. AI models suffer from no such limitations. Institutions that learn about security threats may be able to install patches that block them — but the very act of disseminating a patch gives hackers the information they need to go after institutions that didn’t apply those patches immediately.
And Azhar tells us that regulating AI by requiring “safety training” — instructions to the models that deter destructive actions — will be far less effective than one might hope. Why? Because such training is relatively easy to “abliterate,” which means reversing a few key prompts in a way that undoes most of the safety training.
So what are we going to do? I don’t think we’ll be going back to carbon paper and filing cabinets. For the most part we probably won’t even have to rely on air-gapped local networks that are fully cut off from the internet. Azhar thinks that we’ll move instead to a system of partially closed domains, like AOL in its early years or the SABRE system used by travel agents, which are difficult but not impossible for people and institutions outside those domains to access.
Such closed domains are often called “walled gardens,” but I’d suggest that a better analogy is “gated communities,” like the one my late parents lived in toward the end of their lives. People who didn’t live in their community could visit them, but only if my parents had provided their names to the staff manning the gate. It worked — they could invite friends from outside over — but it was annoying and time-consuming compared with life outside the gates, where everyone could come and go freely.
Similarly, a digital society that imposes annoying security restrictions and checks but is much less vulnerable to AI-powered hacking should be workable. But security will come at a cost. How much of a cost?
Some back-of-the-envelope calculations, which I won’t share here, suggest that the advent of the open internet raised GDP by around 10 percent. A more secure system of gated digital communities will mean giving up only part of that gain — in the long run.
But I’m much more worried about the short run, because our whole economy is currently designed to work with an open web. Suppose, for example, that we’re forced to revert to the old system in which you had to physically visit a travel agent with a secure SABRE link to buy plane tickets. I remember that world, and it wasn’t horrible. But most of the travel agents and offices that used to perform that function are gone and can’t be instantly conjured back into existence. And you have to imagine that problem occurring all across the economy.
Again, this isn’t a hypothetical future risk. It’s happening as you read this.
Just the other day AI enthusiasts were talking about how their miraculous technology would soon generate 10 or 15 percent a year economic growth. That was always silly. But now there’s a very real prospect that for the next few years AI, by empowering bad guys and forcing us to suddenly adapt to new threats, will make us poorer.
MUSICAL CODA
No obvious relation to post, but somehow fits my pre-midterms mood
