How the World Got Wired · Epilogue — One Last Thing

About the Money

I want to tell you one more thing about the AI labs, because you will hear a lot of complicated things said about them in the years ahead.

You read in the last few chapters about Anthropic and OpenAI, two of the most important AI companies in the world right now. You also read, earlier in this book, about the pattern where idealistic founders take money from venture capitalists and become, in various ways, less idealistic. You met Sandy Lerner and Len Bosack, pushed out of Cisco. You met Marc Andreessen, who became one of the most powerful venture capitalists in the world. The shape of the story, in this book, has often been: small idealistic team starts something, takes outside money, and ends up tangled with the same powerful interests they originally meant to be separate from.

The AI labs are a complicated case of this pattern. Both OpenAI and Anthropic were founded by people who said openly that they wanted to build AI in a way that put safety and benefit to humanity first. Both have taken, since their founding, enormous amounts of money from venture capitalists and from the largest American technology companies. Microsoft has invested over thirteen billion dollars in OpenAI. Anthropic has taken multi-billion-dollar investments from Google, Amazon, and others. The labs that exist primarily to think carefully about how artificial intelligence affects humanity are, simultaneously, deeply entangled with some of the largest corporations on Earth.

You might look at this and think it is a betrayal. Some critics do.

Here is the part I want you to hold in mind, because it is genuinely different from the pattern you saw with the earlier internet companies.

Training a frontier AI model — the kind of model that can have a useful conversation with you, that can write working code, that can pass a medical exam — costs hundreds of millions of dollars in computing time alone. Building the supporting infrastructure costs much more than that. The very biggest models, today, cost more than a billion dollars each to train. By the time you read this, the numbers are probably larger.

A research lab cannot, today, do frontier AI work on a Canadian academic grant. It cannot do frontier work on a small business loan. The amount of money required is, simply, in the territory of the largest corporations and governments in the world.

This is, by itself, important. It means that people who want to do frontier AI work — including people who want to do it carefully, with attention to safety — have no real option but to partner with the kinds of organizations that have access to enormous capital. There is no Canadian-grant version of frontier AI, the way there was a Canadian-grant version of the work that produced the 1986 backpropagation paper. The thing has gotten too expensive.

The founders of Anthropic, when they left OpenAI in 2021 to start their own company, knew this. They knew they were going to have to take billions of dollars from Google and Amazon. They did it anyway, because the alternative was to not be at the frontier at all — and they believed, then and now, that somebody with safety as a primary concern needed to be at the frontier. You can argue with that judgment. Many people do. Many argue that no lab, no matter how well-intentioned its founders, can really resist the pressures of investors who have put in that much money. Many others argue that the world is better off with safety-focused labs in the conversation than without them.

The argument is real and ongoing.

What I want you to know is this: the financial entanglement, by itself, is not evidence that the founders have abandoned their principles. The frontier of AI is, right now, simply not a place you can occupy on small budgets. The labs that have raised billions are not necessarily the labs that have sold their souls. They are simply the labs that exist at the frontier, because the frontier requires enormous money to occupy.

The simple stories about this — they took the money, so they must be bad — will tend to be wrong. The reality, as with most of the realities in this book, is both more interesting and more uncomfortable. It is the kind of uncomfortable that, by this point, you have had a fair amount of practice being comfortable with.

All 24 chapters of How the World Got Wired · Library · All courses