Anthropic has filed to go public — stock ticker symbol ANT on Nasdaq, with a projected valuation of $1.5 trillion to $2 trillion. It’s a surreal amount of money, but only a modest slice of the $40 trillion market Anthropic predicts will be affected by artificial intelligence. The company wants to make its technology the backbone of our future economy and society.
In order to get there, Anthropic is contractually obligated to hand huge amounts of money to its competitors (which are in many cases also its investors). It must build as quickly and as aggressively as possible, in the face of growing skepticism from the public, which is aggrieved by A.I. for a number of reasons.
The company, like the rest of the A.I. industry, is making a huge bet that it can brute-force its way into achieving its goals, and going public is part of that plan. But there are realities it may not be prepared to surmount just yet.
The filing barely mentions superintelligence nor A.I. takeovers of the globe. Instead, on the very first page, Anthropic refers to a concept known as the “bitter lesson” — that when it comes to A.I. innovation, brute-force power will consistently replace handcrafted solutions.
A.I. techies understand it in their bones — but you might need some background. Starting in the 1950s, when computers filled entire rooms, A.I. researchers tried to make smarter systems that encoded human knowledge. These systems, carefully programmed by experts, played games, processed documents and helped mining companies decide where to dig for minerals.
Over the decades, the steady increase in computing speed, power and memory made these handcrafted systems obsolete. Carefully considered, tightly focused algorithms ended up obliterated by brute-force development of more generalized systems.
Famously, IBM’s chess-playing computer, Deep Blue, which used custom chips and supercomputer speed, beat the reigning world chess champion, Garry Kasparov, in 1997 by searching through 200 million chess positions per second. By 2017, Google DeepMind’s AlphaZero project had become even more generalized: Google fed the rules of chess to AlphaZero software, it played millions of games against itself, and it became extremely good at chess in about nine hours.
This pattern — brute force replacing handcrafted solutions — proved itself out, decade after decade. That is the bitter lesson, a term coined in 2019 by the computer scientist Richard Sutton.
Imagine you spend years working on a Ph.D. focused on how to teach a computer to answer questions in a more intelligent way. You might write tons of code to test your ideas and survey thousands of pages of equation-dense literature. It’s hard on your marriage. Then, right as you’re about to defend your thesis, along comes someone with a few million or billion dollars in computational power and that person makes a computer model that can do what you said you’d do.
He or she basically solves your problem by accident, just by making something bigger. Your work isn’t wrong. It just may not matter anymore. You’d be bitter, too.
Anthropic — and nearly every other frontier A.I. company — embraced the bitter lesson. The technologists at these companies had one job: Build bigger models. They had thousands of computers in data centers digest all the data possible, leading to large language models that replied in sentences. Then bots started doing surprising things, like writing legal documents, summarizing medical research and writing good computer code quickly without human supervision. Truckloads of investment followed.
Anthropic is already a leading A.I. company, but to keep growing, it must run up the world’s largest phone bill. According to its initial public offering filing, it has promised something like $800 billion, over multiple decades, to semiconductor manufacturers like Broadcom; to competitors like Amazon, SpaceX, Google and Microsoft; and to various data center providers that provide access to advanced chips. To make bigger models, Anthropic must find more computational power (“compute”), which means leasing more data centers, getting access to more chips and securing electricity by the gigawatt.
For Anthropic and its peers, the lesson of our new A.I. world is not bitter at all. It is, in fact, thrilling — they unlocked something huge about simulating human thought, and the results, steered wisely, will lead to a prosperous economy, cure diseases and rapidly accelerate all parts of our digital world. It’s expensive, but worth it.
Others feel the bitterness differently. Mathematicians are watching in frustration as A.I. companies brute-force their way into solving longstanding complex math problems, disrupting their discipline. Many artists, illustrators and authors have rejected A.I.-generated materials outright. And multiple communities, seeing the physical costs of data centers but unconvinced of the economic advantages of more A.I., are trying to stop their construction.
To me, the valuable thing about the bitter lesson is that it can get us away from the endless conversations about whether A.I. is alive or if A.I. agents are going to hack us all back to the Stone Age. We fixate on imaginary sci-fi doom scenarios, or we imagine total A.I. utopias, and these extrapolations provide the talking points to justify any number of decisions about A.I., including whether to slow down development or charge full speed ahead.
Yes, the future matters, but the present is here now. If I were to talk to powerful people in A.I., the first thing I’d want to know is not what they believe about A.I. consciousness or future risks of doom, but rather how they think about scaling — increasing computational power to build bigger A.I. systems and reach new users. Because if they believe that the best way to grow large language model capacity is through more and more scale, then that tells you that their budgets will be focused heavily on physical infrastructure and gaining access to more energy and land and that they will be lobbying to build ever bigger.
That doesn’t mean everyone else has to acquiesce. How much of our society, culture, land, water and electricity do we want to allocate to building language-prediction devices? Is it worth potentially burning through trillions of dollars this quickly? Where do you come down on the bitter lesson?
A.I. companies have decided, on our behalf, to pour immense resources into this technology, and some of what they have accomplished is breathtaking. Personally, I don’t want the ride to stop — I want to see what else is possible.
But we also need A.I. companies to talk seriously with the rest of us about how far we’re all willing to go. Your answer might be radically different from mine. You might want all the data centers to stop, now. “What are you willing to give up for A.I.?” is an uncomfortable question, but it will be more useful than bickering over whether a computer has a soul.
As much as I love computers, democracy is the more important technology. Obviously, Anthropic’s I.P.O. is just the beginning. Other companies will follow suit, raising their own trillions to scale quickly. Scaling will hit limits, but not before huge figures have been tallied up — numbers of G.P.U.s, hours of compute, petabytes of data and trillions of dollars.
The bitter lesson is something everyone needs to understand, so that we can decide just how much that bitterness is worth — for the A.I. industry, and for the rest of us.
Paul Ford is a contributing writer in Opinion. He is a founder and the president of Aboard, an A.I.-powered software consulting firm, and a host of “The Aboard Podcast.”
Source photographs via Getty Images.
The Times is committed to publishing a diversity of letters to the editor. We’d like to hear what you think about this or any of our articles. Here are some tips. And here’s our email: [email protected].
Follow the New York Times Opinion section on Facebook, Instagram, TikTok, Bluesky, WhatsApp and Threads.
The post The Open Secret Buried in Anthropic’s 300-Page I.P.O. Filing appeared first on New York Times.
