Gods and Machines
DeepSeek, Chips, and the Problem with American AI Strategy
"One day, I realized all the dumb, selfish things people do... it's not our fault. No one designed us. We're just an accident, Harold. We're just bad code. But the thing you built... It's perfect. Rational. Beautiful. By design."
— Root, Person of Interest
It's no secret that Jonathan Nolan's Person of Interest is my favorite television series of all time. A five-season epic that starts off as post-9/11 paranoia crime of the week episodic and steadily grows into one of the best written stories about artificial intelligence, humanity, and the role of the state. As the AI arms race between the US and China heats up, there is perhaps no better show to help understand the current state of affairs and why the unveiling of China’s DeepSeek should be an inflection point for the US, particularly for the national security community. In fact, I would say there is no piece of fictional media that should be prioritized more for consumption by policymakers, leaders, and pundits than Person of Interest (stop watching Sicario). Please indulge me for a moment as I explain and then I’ll get into my real analysis of DeepSeek, chip restrictions, and US AI investment.
In the United States, it is nearly impossible to talk about AI development without some allusion to the Almighty. It goes like this: first we build narrow AI (what we have today), that somehow gets us to AGI (equivalent of human capacity and capability) and one iteration after we get an ASI (an intelligence greater than our own capable of growing on its own beyond our comprehension). Thusly, many of Silicon Valley’s titans apply a religious component to the future of AI. Some fear AI as an engine of creative destruction, others a benevolent savior removing the burden of work and governance from humankind. And yet none should ever have that sort of power or claim to be able to create it. Machines are not gods, they are tools, they are our created, but they are not and cannot be god. I’m not a religious man and this has nothing to do with blasphemy, but rather the very real human fault of hubris. This leads us to my favorite tv show. In Person of Interest, the core philosophical (and physical) battle between the good guys and the bad guys is whether an ASI should be a servant of humanity or its omniscient ruler, and whether a machine (or a person) can be inherently good, bad, or simply the product of decisions made by algorithm and by creator alike. So why does this matter for us? Because if we don’t understand what we are building, what we are buying, and the underlying challenges, issues, and implications surrounding artificial intelligence beyond headline hysteria, we will fail. We must understand that not only would the CCP’s development and deployment of a superior artificial intelligence spell disaster for the free world, it will also be our fault if we obsess over the mystical instead of reality. With OpenAI’s Sam Altman and others claiming that AGI and/or the Singularity is just around the corner…until we all look under the hood and realize the model of the week is just another hallucinating chatbot, it’s a bit hard to take AI development seriously. I’m a long-time proponent of both AI research and weaponized AI. We will get there. I have little doubt we will someday achieve AGI and perhaps even greater, but not like this.
In other words, we have three serious problems when it comes to our AI programs: grift, bloat, and cults.
The Bloat
DeepSeek provides us with a glimpse into alternate development pathways for AI, compared to what the titans of Silicon Valley have assured us is the only way to achieve and maintain a competitive advantage in artificial intelligence. To be sure, I am not arguing that DeepSeek’s model is the One True Method, but I can tell you that it's clear there is a lot of bloat and waste in American AI development. The lean, innovative companies and culture of early Silicon Valley have now all morphed into bureaucratic nightmares. Nor can you simply throw money at a problem until it goes away without a clearly articulated strategy that isn’t just “we’re gonna build god.”
It’s rather novel but I’d also like to point out that there is a key cultural difference between the US and PRC that has, in this specific case, likely aided in DeepSeek's development: Chinese developers certainly for political reasons cannot market themselves publicly and to CCP leadership as attempting to build god. They are simply building a machine to help the Party. To build god would be to make the Party irrelevant, and that cannot be tolerated. If a Chinese engineer asked for $500 billion to build god, they'd be disappeared. Chinese political and economic culture is built on graft and patronage, so it should say something when their companies can do something as big as develop DeepSeek on the relative cheap. It should also say something about the misdirection and resource allocation of their US counterparts.
The Cult
At its core, the development of any level of artificial intelligence is an engineering problem. But to many in Silicon Valley, Wall Street, and Washington, the development of even an AGI is a stepping stone to a new deity. They want, hell they must, believe that AI will be able to solve all of the messy problems of humanity and democratic governance. AI is a mystical problem to them, not a scientific one. And if you truly buy into the cult of an AI-deity, can you really put a price on divinity?
As I wrote two years ago, we should not fear the god from the machine but the man or woman who wants their machine to be seen as a god. So much of the AI industry is grift, and the only way to keep selling when the grift doesn’t deliver is to rely on belief rather than replicable proof. The American philosophical obsession with AI as the new god, and thus the pursuit of it as some sort of holy crusade for humanity's redemption, is our chief impediment to actually appropriately resourcing the problem. You cannot apply the religious to the scientific and assure your flock that a large enough donation to the pastor and a sip of Kool-Aid will grant you enlightenment with no evidence to support such a claim. That is not the Manhattan Project, that is Jonestown.
The Grift is Real
It would be incredibly inaccurate to say there has not been tremendous progress in the development of narrow AI like ChatGPT and autonomous weapons systems. Yet somehow in press releases, briefings to the USG, and demonstrations, Silicon Valley has done nothing but overhype and underdeliver. We cannot accurately assess the progress of ourselves or our competitors if we bet on the marketing department instead of production. To believe every lobbying pitch, everything is AI-enabled these days. A string of If-Then-Else statements, Samaritan does not make.
Please understand that I'm not overhyping DeepSeek's development either as some sort of Sputnik moment. The fact that the Chinese managed to develop an effective, cheaper, and censored AI chatbot is entirely consistent with how they've been developing and operating in other fields, particularly weapons systems. It's probably also true that they have their own marketing issues, development failures, inefficiencies, and grift. I wouldn't be surprised if the PRC government secretly offset some of the cost/resources but the larger development community seems to believe that DeepSeek is not a theater performance. Gone are the days where we should bet that the PRC’s incompetence outweighs our own. DeepSeek is only the latest evidence smacking us in the face that the CCP is not fucking around and that quite frankly, modern Silicon Valley and venture capital runs more on its own hype than delivery. The enshittification of the internet is proof of that.
Our policy response to DeepSeek cannot be a crash of resources without any sort of substantial, critical review of Silicon Valley's hype machine. Nor should that be the response of the private sector. No longer are the giants of technology in America nimble organizations, they have become what they sought to disrupt. If we're going to make this the age of efficiency, maybe we start with asking why we need $500 billion dollars to build an abstract concept when the PRC can do what we do for far less? Both government and industry are signing off on AI projects left and right without any serious analysis on ROI. We’ve seen this before in every panic from the H-Bomb to GWOT. If we do not pause and review our AI investments and methodologies, the PRC's AI sector is going to turn us into the Soviets of the 1980s as they chased SDI: overspending and resourcing in a panic while failing to keep pace with the Main Enemy. If we want to win, if we want to keep the CCP from leading the world’s AI revolution, we have to cut the bullshit and give Silicon Valley the scrutiny it deserves. Money doesn’t solve the problem if we’re just burning it for a fancy light display.
On Chips Restrictions
There's commentary out there that this success was driven by the CCP's necessity to innovate under sanctions and restrictions. But I think we need to flip the argument on its head: where is our necessity to innovate? Is that not the American claim to fame? Perhaps we have fallen behind (or at least become complacent) because there is no necessity to innovate, no sense of urgency, no rigor attached to design. The Chinese design for DeepSeek, born of necessity, reminds me a lot of the shift in supercomputer and chip design in the 1970s when instead of building larger and larger, we pivoted to building smaller and ushered in a new age of computing. DeepSeek (most likely) cost more than the $5 million claimed, but it's fairly obvious it's not going to cost Silicon Valley's $500 billion either or however much we’ve spent on the latest ChatGPT model. When efficiency is the name of the game, a blank cheque only invites failure. You need a forcing function to drive innovation.
Relatedly, some folks are making the argument that our restrictions on chips to the PRC led to this necessity-driven innovation, and that's why we shouldn't have cut off chips. To that I'll say, let's play this out: the PRC gets a bolt of success and we should give them even better technology to advance that lead? As Jordan Schneider points out, this doesn't really line up with the relevant deployment timeline of chip restrictions. Nor does it match internal commentary from DeepSeek investors about the long-term need for more compute even if efficiency is the short-term game. Never mind that many of us, including China policy heavyweights like Bill Bishop and James Mulvenon, were critical of how the Biden administration handled the chip restrictions because they were too little, too late, and had some glaring holes. It's clear that not only did we buy our own bullshit, we underestimated (again) the quality of talent and ecosystem for innovation in the PRC. We are not fighting the Soviets who hoped to just copy our technology, the PRC actually believes in innovation. For all of its brutality and bureaucratic absurdity, we are fighting an enemy that is far more like us than we care to admit. We have to learn to accept that.
It’s clear that design efficiency matters, DeepSeek proves it, but compute power remains our advantage. Anyone worth their salt knows AI development is dictated by three main factors: compute, data, and algorithms. DeepSeek has an effective set of algorithms on an efficient design. Everyone has plenty of data. And we have the high ground (for now) on chips. Chip restrictions slow the PRC’s ability to scale the success of DeepSeek,. Where these chip restrictions also still matter is in restricting scale and protecting proprietary designs that may not have helped with DeepSeek, but may help with other critical technologies development like weapons systems or even other AI designs. If DeepSeek teaches us anything, it is that innovation (particularly for AI) is not carved along a single path and our strategy must simultaneously close as many paths to the CCP while exploring more on our own.
On Censorship
My secondary policy concern regarding DeepSeek is that it is yet another case (like TikTok/Rednote) where Americans are more than open to playing with machines that censor them to a point of comedy. Have we as a culture become more complacent with censorship in exchange for capability, is it because we have endured years of claims about censorship that wasn't really there, or is it because we don't believe PRC-style censorship could ever happen here? I don't like any of these answers, but this is an area that especially now is something to watch as we develop new AI models and forms of interactive media for a wide range of purposes from education to military decision-making. I do worry about a world where we default to the quality of Chinese technology and the cultural and political norms that are sure to follow. If we fail to overhaul our AI investment and development process, that is exactly what will happen.
If you would like to read more about the future of US-China conflict or what happens when we unilaterally disarm from the AI arms race, check out my novel, EX SUPRA, about the world after the fall of Taiwan, an isolationist and hyper-partisan America, and World War III. It was nominated for a Prometheus Award for best science fiction novel and there’s a sequel in the works! If you have any suggestions for topics for future newsletters, please send them my way on Twitter @Iron_Man_Actual and on BlueSky @tonystark.bsky.social. And don’t forget to subscribe!