Marc Andreessen occasionally sets the world on its ear with a sweeping hypothesis about the dawn of a new technological era. In his legendary 2011 blog post “Why Software Is Eating the World,” the cofounder of Andreessen Horowitz made the then-novel, now-undeniable case that even the most old-school industrial companies would soon have to put software at their core. In 2020, as Covid-19 caught the world desperately short of masks and nasal swabs, he published “It’s Time To Build,” a call to arms for reviving investment in technologies that could solve urgent problems like pandemics, climate change, crumbling infrastructure, and housing shortages.
Now he’s back with a 7,000-word screed, another stab at framing the narrative; this time, the story is that “AI will not destroy the world, and in fact may save it.” Much of it is devoted to debunking AI doom scenarios, and the rest to touting AI as little short of a civilizational savior.
This is of course predictable. Andreessen invests in technological revolutions, so he has little incentive to do anything but hype them up. His post does have value, though, in two ways. First, its obvious blind spots are a useful guide to the thinking of the biggest AI hypesters and where they go astray. Second, its takedown of some of the more hysterical AI fears is actually (somewhat) on target.
So let’s dive in.
Andreessen tips his hand early by offering “a brief description of AI”: “The application of mathematics and software code to teach computers how to understand, synthesize, and generate knowledge in ways similar to how people do it” (my emphasis).
This seemingly innocuous parallel with human thinking, much like the phrase “artificial intelligence” itself, elides the vast gulf in capability between human minds and the current state of machine learning. Large language models (LLMs) are statistical inference algorithms. They predict the next likeliest thing in a sequence of things, such as words in a sentence. They produce what looks very much like human writing because they’ve been trained on vast quantities of human writing to predict what a human would write.
You’ll have already noticed that this is not even remotely similar to how you “understand, synthesize and generate knowledge.” You, like every human, have learned about the world by directly interacting with it. You’ve developed conceptions of physical objects such as trees and tables, of abstractions such as poverty and ethics, and of other people’s thoughts and feelings. You’ve learned to use language to talk about and process those conceptions, but language is just a layer for you, a way to share and refine your mental picture of the world. For LLMs, there is no mental picture; language is all there is.
To be sure, LLMs have made surprising leaps in ability recently, leading Microsoft researchers to claim that GPT-4, the latest model from OpenAI, contains “sparks” of general intelligence. And LLMs are not the only avenue of AI research. It can’t be ruled out that machines will eventually develop something more like our intelligence—though there are also good reasons to think it will end up being more alien than human.
However, it’s essential to Andreessen’s argument that you perceive AI as headed toward an ideal version of full humanlike intelligence, because what he does next is enumerate some of the ways this form of AI will make the world better.
In Andreessen’s promised AI-augmented world, “every child will have an AI tutor that is infinitely patient, infinitely compassionate, infinitely knowledgeable, infinitely helpful.” Every adult will have “an AI assistant/coach/mentor/trainer/advisor/therapist” that “will be present through all of life’s opportunities and challenges, maximizing every person’s outcomes.” Giving AI coaches to influential people like CEOs and government officials “may be the most important of all” the augmentations because “the magnification effects of better decisions by leaders across the people they lead are enormous.”
There are two key blind spots here. First is the assumption that these AI sidekicks will be “infinitely knowledgeable”—a gigantic stretch given that right now LLMs routinely make up facts out of whole cloth and sometimes continue to do so even when their human users point out the errors. (They do this because, as mentioned above, LLMs are just statistical patterns of words, with no conception of the reality behind the words.)
More troubling is the assumption that humans would use even a far superior AI to make “better” decisions. Better for whom? An “infinitely patient” and “infinitely helpful” AI coach could just as happily help its human master wreak genocide as devise a more efficient manufacturing process or a fairer benefits framework.
OK, you might say, but what if the AI sidekick were somehow programmed not merely to maximize its human’s capabilities but to nudge them away from sociopathic decisions? Fine—except that Marc Andreessen would strenuously oppose that suggestion.
A large chunk of his post is devoted to attacking one of the big fears about AI: that it will spread hate speech and misinformation. To be clear, he doesn’t argue that it won’t spread hate speech and misinformation. He merely says that policing social media has been fraught and complicated (true!), that the people who believe in doing it are mostly on the political left (also true!), that policing AI will be even more fraught because “AI is highly likely to be the control layer for everything in the world” (umm, OK?) and therefore it shouldn’t be policed, regardless of the consequences.
Now, this is a position one can choose to take, but it’s also fundamentally at odds with the idea that people’s AI coaches—even if they ever get past the point of making shit up and become useful—will make the world better. If, as Andreessen insists, programming AI to have certain values is off the table, then all an AI coach will do is help humans get better at making the world whatever they make it, which … well, take a look around you.
Sure, brilliant scientists will come up with even more brilliant life-saving medicines and climate-saving battery chemistries. But every rapacious, criminal, greedy, and manipulative schemer out there will also get better at taking advantage of other humans. Not a single technology in history has yet changed basic human nature.
Another example of Andreessen’s dubious logic emerges when he tackles the common fear that AI will leave everyone unemployed. His argument here is that AI is no different from previous technological advances, which have not eliminated jobs. This is certainly true in the long run: New technology destroys certain kinds of jobs and eventually creates others. But the way he reaches this conclusion is almost laughable in its simplicity.
Andreessen begins by setting up a straw man: the notion that AI will take “all our jobs.” Like, literally all. He then knocks it down by pointing to the so-called “lump of labor fallacy,” which is “the incorrect notion that there is a fixed amount of labor to be done in the economy at any given time, and either machines do it or people do it—and if machines do it, there will be no work for people to do.”
I’d be surprised if Andreessen’s highly educated audience actually believes the lump of labor fallacy, but he goes ahead and dismantles it anyway, introducing—as if it were new to his readers—the concept of productivity growth. He argues that when technology makes companies more productive, they pass the savings on to their customers in the form of lower prices, which leaves people with more money to buy more things, which increases demand, which increases production, in a beautiful self-sustaining virtuous cycle of growth. Better still, because technology makes workers more productive, their employers pay them more, so they have even more to spend, so growth gets double-juiced.
There are many things wrong with this argument. When companies become more productive, they don’t pass savings on to customers unless they’re forced to by competition or regulation. Competition and regulation are weak in many places and many industries, especially where companies are growing larger and more dominant—think big-box stores in towns where local stores are shutting down. (And it’s not like Andreessen is unaware of this. His “It’s time to build” post rails against “forces that hold back market-based competition” such as oligopolies and regulatory capture.)
Moreover, large companies are more likely than smaller ones both to have the technical resources to implement AI and to see a meaningful benefit from doing so—AI, after all, is most useful when there are large amounts of data for it to crunch. So AI may even reduce competition, and enrich the owners of the companies that use it without reducing prices for their customers.
Then, while technology may make companies more productive, it only sometimes makes individual workers more productive (so-called marginal productivity). Other times, it just allows companies to automate part of the work and employ fewer people. Daron Acemoglu and Simon Johnson’s book Power and Progress, a long but invaluable guide to understanding exactly how technology has historically affected jobs, calls this “so-so automation.”
For example, take supermarket self-checkout kiosks. These don’t make the remaining checkout staff more productive, nor do they help the supermarket get more shoppers or sell more goods. They merely allow it to let go of some staff. Plenty of technological advances can improve marginal productivity, but—the book argues—whether they do depends on how companies choose to implement them. Some uses improve workers’ capabilities; others, like so-so automation, only improve the overall bottom line. And a company often chooses the former only if its workers, or the law, force it to. (Hear Acemoglu talk about this with me on our podcast Have a Nice Future.)
The real concern about AI and jobs, which Andreessen entirely ignores, is that while a lot of people will lose work quickly, new kinds of jobs—in new industries and markets created by AI—will take longer to emerge, and for many workers, reskilling will be hard or out of reach. And this, too, has happened with every major technological upheaval to date.
Another thing Andreessen would like you to believe is that AI won’t lead to “crippling inequality.” Once again, this is something of a straw man—inequality doesn’t have to be crippling to be worse than it is today. Oddly, Andreessen kinda shoots down his own argument here. He says that technology doesn’t lead to inequality because the inventor of a technology has an incentive to make it accessible to as many people as possible. As the “classic example” he cites Elon Musk’s scheme for turning Teslas from a luxury marque into a mass-market car—which, he notes, made Musk “the richest man in the world.”
Yet as Musk was becoming the richest man in the world by taking the Tesla to the masses, and many other technologies have also gone mainstream, the past 30 years have seen a slow but steady rise in income inequality in the US. Somehow, this doesn’t seem like an argument against technology fomenting inequality.
We now come to the sensible things in Andreessen’s opus. Andreessen is correct when he dismisses the notion that a superintelligent AI will destroy humanity. He identifies this as just the latest iteration of a long-lived cultural meme about human creations run amok (Prometheus, the golem, Frankenstein), and he points out that the idea that AI could even decide to kill us all is a “category error”—it assumes AI has a mind of its own. Rather, he says, AI “is math—code—computers, built by people, owned by people, used by people, controlled by people.”
This is absolutely true, a welcome antidote to the apocalyptic warnings of the likes of Eliezer Yudkowsky—and entirely at odds with Andreessen’s aforementioned claim that giving everyone an “AI coach” will make the world automatically better. As I’ve already said: If people build, own, use, and control AI, they will do with it exactly what they want to do, and that could include frying the planet to a crisp.
This assertion brings us to the second sensible point. This is where Andreessen addresses the fear that people will use AI to do bad things, such as design a bioweapon, hack into the Pentagon, or commit an act of terrorism. These are legitimate fears, he says, but the solution is not to restrict AI.
He’s right up to a point. The kinds of bad things that people could do with AI are already illegal, because they’re bad. It’s a general principle of good lawmaking to target the harm, not the means. Murder is murder whether it’s carried out with a gun, a knife, or an AI-controlled drone. Racial discrimination in hiring is the same whether an employer looks at your picture, infers your race from your name on a résumé, or uses a screening algorithm that contains inadvertent hidden bias against people of color. And legislation designed to curb a specific technology runs the risk of becoming out of date as the technology changes.
Nonetheless, some means of doing harm are so much more effective than others that they require special legislation. Even in the most permissive US states, not everyone is allowed to own and carry a gun, let alone bigger weapons. If murder by AI-controlled drone becomes a thing, you can be pretty sure we’ll see tougher drone laws.
AI-enabled crimes may also require changes not to laws as much as to law enforcement. The authorities may need new techniques to investigate such crimes, just as they’ve needed to learn to hunt down drug dealers who trade on the dark web using cryptocurrency.
In some cases the solution to a problem is not new laws, but for industry to adopt standards. It’s hard to make spreading disinformation a crime, especially in the United States, given how politicized the definition of disinformation would be. But, Andreessen notes, “If you are worried about AI generating fake people and fake videos, the answer is to build new systems where people can verify themselves and real content via cryptographic signatures.”
Yes, and I’d say a couple of further steps are needed. First, AI companies should also adopt a common standard for putting digital signatures on AI-generated images and videos. (There is at least one such standard already, the Content Authenticity Initiative, which also offers a way to verify content even if it’s had a signature stripped off.) Then social media platforms, news sites, and the like should label any such content, so anyone seeing it knows its provenance without having to go search for it.
Andreessen has set the narrative for major technological shifts before. It’s welcome, in some ways, that he’s trying to do so again. Much of the current public discourse on AI is such a hysterical mix of apocalyptic paranoia and breathless utopianism that it’s hard for anyone to have a clear position. Andreessen’s words carry weight with many people, and it’s good that he’s knocking down some of the most obvious misguided fears.
Unfortunately, the way he does it feels almost like a crass self-parody of his “Software Is Eating the World” manifesto. But if nothing else, it helps illuminate the hollowness of both the doom and the hype. The problem—as I’ve written before—is that people apparently find it really hard to think in any other terms.