Novus Ordo Seclorum

Reflections on DeepSeek

“One of the most remarkable aspects of this self-evolution is the emergence of sophisticated behaviors as the test-time computation increases. Behaviors such as reflection—where the model revisits and reevaluates its previous steps—and the exploration of alternative approaches to problem-solving arise spontaneously. These behaviors are not explicitly programmed but instead emerge as a result of the model’s interaction with the reinforcement learning environment.”
-DeepSeek r1 Technical Report

There’s a certain jitteriness to many of the people I know who work on AI research and policy. We bounce our legs under our desks, or tap our fingers on whatever surface is closest to our hands. Maybe AI attracts quirky and neurotic people (no—definitely). But I think the jitteriness comes from something deeper, too: an overwhelming sense of urgency, duty, and burden. And that sense, in turn, comes from the simple reality that we know something others do not.

It’s not that we understand something that others are not smart enough to get. It’s all pretty straightforward, actually. Mostly I think we are just early. Early to the insight that AI is going to overturn countless things about the status quo, including at least some things that most people like, or at least find comfortably familiar. Early to the idea that this technological revolution will be intense, chaotic, and filled with uncertainty. Early to the excitement, and early, I must admit, to the anxiety. Early to the knowledge that we are stepping into a novus ordo seclorum—a new order of the ages. And early to feeling all of this in our bones, rather than just understanding it in some abstract intellectual way.

It can be lonely at times. At family gatherings over the last two years, I’ve felt myself in somewhat of a daze—not so much because I am distracted by my X feed or my group chats, but because I just can’t shake the feeling that all of this is going to change in ways I cannot quite define. “Is there a word for feeling nostalgic for the time period you’re living through at the time you’re living it?,” Sam Altman once asked.

I find myself caught between trying to cherish these moments for what they are and trying to shake the people around me out of what I perceive to be their complacency. “The tectonic plates of history are in fast motion,” I want to tell them. But really I know I should just play with my niece and talk to my in-laws about their jobs and their hobbies.

The constant context switching can be emotionally exhausting. The whole damn field is emotionally exhausting. This is not what it usually feels like to study policy. Working on, say, municipal debt, or healthcare policy, or taxes, is not like this—consequential though those things may be.

The reason is that I—and a great many of you—know that we are riding an exponential, a world-historical exponential. But not that many people really know it. One of the things that weighs on me is the knowledge that one day soon, everyone will know it—and when they do, I worry that they may not be entirely happy about it. I hope they understand that those of us who knew before them meant well, and that for the most part we were trying our best.

I wonder whether, over the last week or so, we got our first taste of what it is like for many more people to become aware of what is happening. DeepSeek’s r1 did not especially shock me. I predicted that a Chinese lab would credibly replicate OpenAI’s o1 model within a few months in my first analysis of o1 back in September. For me, this was priced in. But it was not priced in for many others, and I did not remotely anticipate how much panic, hype, and hyperbole it would stir.

How could China “catch up” so fast? Who the hell is DeepSeek? Why did their model cost so little to train? Did they steal America’s intellectual property? Did they smuggle chips? Why are they number one on the App Store? Does compute matter anymore? Did America’s lead vanish? Is OpenAI going to go bankrupt? Did six hundred billion dollars of Nvidia simply disappear in front of our eyes?

One can have a level-headed discussion about all these things, and I have tried. But that’s not the point. The point is that a lot of people are just not used to the speed and the turbulence of it all—triumphant this month, anxious the next. Partially this is just because this is a field dominated by internet discourse, and this is the pace of the internet. But more importantly, it is because we are riding an exponential. Change will happen abruptly, and before you know it, it will happen again.

We are used to internet-driven culture having this velocity, but it’s a relatively new industrial phenomenon. Algorithmic efficiency gains of 400-500% can be expected annually. The performance of AI chips has been increasing by around 130% per year. Companies buy vastly more of those chips every year, and that will continue for at least the next year or two.

And on top of all this, we’ve found a way to make the models think. A good language model, when placed into a well-designed reinforcement learning environment, given hard problems to solve, and simply allowed to generate words (tokens), “naturally” learns to start self-reflecting, planning alternative strategies when it encounters a dead end, and correcting errors it has made. The models are learning to think. And at this too they appear to be improving on a similar trajectory.

All of this is just table stakes in this particular game of poker, so DeepSeek’s efficiency improvements in v3—the model undergirding r1 that was trained with a mere $5.5 million in marginal compute cost—were not that much of a surprise to close observers of AI; they were about on trend. The shocking thing is not especially that DeepSeek achieved this trend; the shocking thing is that this is the trend we are on. And there is no end in sight.

This is not another social media. This is not another smartphone. This is something altogether different. Titanic things, beyond everyone’s grasp, are happening.

We do not quite know where we are going, but that has always been the case. Life has always been an improvisatory adventure on the open seas. Our job is to make the adventure worth it. We will not do that with small-minded policy formulated by compliance consultants, bureaucrats, and other forces of the fading status quo. Policy is just as likely, if not more, to harm than it is to help. Government can be a force for good here, but it can just as easily be a force for ill.

Getting this transformation anything like right is going to be beautiful, messy, chaotic, and perhaps the tightest tightrope walk any country has ever performed. I believe amazing things await us if we can pull it off, but pulling it off will be hard. Even if we succeed, it will be a transformation, and no transformation is entirely pleasant. The world will have a higher ceiling for both great achievements and profound evil—a higher dynamic range, as I wrote last month. I do not know how quickly the transformation will unfold. My guess is that it will happen more slowly than the most bullish AI observers forecast, but still, in at least some ways, much faster than the average person expects.

If this past week was a surprise to you, you haven’t seen anything yet. Welcome to the party. Welcome to the exponential. Welcome to the novus ordo seclorum.