DeepSeek vs Deep State
Reflections and chastisements on the Washington roots of China's breakout AI success; written in February 2025
By January 2025, even casual American observers had grown used to a barrage of striking new developments in deep learning applications, popularly described as “AI”. Leading domestic developers, like Sam Altman, had a simple story to explain this dizzying string of successes: in deep learning, bigger is better. More data on which to train new models, and more processing power (or “compute”) with which to train them, meant superior models coming out of that training. And as developers were able to harness larger and larger quantities of compute for this purpose, the quality of the output could only increase. Indeed, for many this was the lesson of the deep learning revolution of the past decade, as most famously showcased in large language models (LLMs) like the GPT series from OpenAI. Artificial intelligence, the real deal, had been made possible by pouring ever more computing power into the machine learning mill, and its future prospects were a matter of just how much more compute investors and manufacturers could cough up.
The endless stream of marvels, and its companion the pat story about size and success, made it all the more necessary for commentators to stress just how poorly China’s latest AI foray fit this mould. DeepSeek-R1, an LLM released shortly after inauguration day by the Hangzhou DeepSeek Artificial Intelligence Company, managed to match or beat American competitors on several benchmarks, while boasting training using an order of magnitude less compute and funding. Leveraging innovative, more efficient algorithms for training, the Chinese company had apparently bypassed the need for large amounts of compute. This was not just another shiny new toy, like the ones we had all grown accustomed to from OpenAI and Google: it was an upset in the entire trend of the American AI revolution.
DeepSeek, however, is not just a story of shifts in technological development. It is also a story about US foreign policy, and the people who directed it. And it raises questions not only about the future of AI R&D, but about the accountability of crucial actors in Washington whose destructive policy interventions unwittingly helped pave the way for China’s own breakout success.
While the advantages of less compute-intensive algorithms speak for themselves, it was not their intrinsic virtues that seem to have first prompted Chinese researchers to look into an algorithm-heavy development strategy. There is a broad suspicion among onlookers that the search for more efficient algorithms was motivated by restrictions on Chinese access to state-of-the-art microchips for training, itself a result of American controls on semiconductor exports. These controls were first introduced in October of 2022, subsequently tightened and expanded several times over up through the latest revisions this year on 15 January. They were among the main innovations of the Biden administration’s foreign policy, explicitly implemented to cripple AI advances among hostile countries, especially China.
These controls were not without significant costs to the country imposing them. They invited, as could have been expected, ill effects on geopolitical tensions and international markets, including inadvertent boosts to the Malaysian tech sector, which has given no indication of fealty to the US’ perceived foreign policy interests. And nervous onlookers have warned of the danger such an arms race approach to AI might pose of catastrophic risks from the technology. Still, the Biden administration evidently deemed these drawbacks worth the gain (in their eyes) of a hampered China on AI.
While the existence and effects of the export controls were widely reported, less familiar from the public discourse was their apparent origin within Washington. As I covered in a pair of articles in 2023 and 2024, the export controls seem to have been principally advocated and shaped by a clique of tech policy experts to have embedded themselves deep within the US foreign policy establishment over the course of the Biden administration. This network emerged from within the wider effective altruist (EA) community as part of a greater EA expansion into Washington, and was incubated primarily by the Center for Security and Emerging Technology (CSET), a recently minted think tank at Georgetown with deep financial and intellectual roots in EA and “longtermism”.
With the recent upheaval among federal staff as part of the incoming administration’s promised purge of the “deep state,” it should perhaps surprise us to learn that several key members of this new faction have managed to keep hold of their privileged positions within Washington. Saif Khan, the leading spokesman for the semiconductor embargo, was promoted this January at the Department of Commerce from Counselor to the Secretary for Critical and Emerging Technologies to Senior Advisor to the Under Secretary for Industry and Security. Emily Weinstein and Will Hunt, other prominent promoters and apparent architects of the export control system, are still ensconced in the Commerce Department as senior advisors. Central figures in the CSET crowd’s campaign for export controls also continue to occupy important positions within the wider complex of semi-independent, government-backed institutions that play so much of the role of American federal governance: Dahlia Peterson, co-author with Khan in his defence of export controls, is still a Presidential Management Fellow on China at Voice for America, after graduating from the State Department as a policy advisor; Jason Matheny, outspoken proponent of the controls and founding figure at CSET, remains CEO of the RAND Corporation.
In addition to such visible figures, it is likely that CSET’s larger network retains unlisted federal personnel. The Department of Homeland Security’s Artificial Intelligence Safety and Security Board, for example, was set up under the auspices of Biden’s October 2023 executive order on AI, itself seemingly overseen by sanctions-happy CSET alumni. The Board, conspicuously, does not list its personnel on its official webpage, despite the Board’s work receiving citations elsewhere in government publications on AI policy. Such opacity makes the Center and friends’ influence even harder to track.
All of this is to say: the network of CSET affiliates that apparently originated the post-2022 export controls US microchips gives no appearance of having felt the sting of Trump’s “de-Baathification.” Prolific members remain in powerful positions within Washington, many of them likely invisible.
The persistence of the CSET export controls crew is concerning, because the theory on which they promoted their signature policy is at profound odds with the key to DeepSeek’s success. This overall vision is well captured in a Foreign Affairs article by CSET alum and former White House staffer Ben Buchanan, which describes an “AI triad” to complement the famous “nuclear triad” of traditional nuclear strategy. Just as a successful system of nuclear deterrence depends upon a combination of bomber aircraft, land-based missiles, and nuclear submarines, so a healthy national AI infrastructure requires data on which to train deep learning models, algorithms by which to train them, and compute with which to train them. While all are necessary, investment in one necessarily trades off against the other two, leading to the question of which of the three to prioritise. Buchanan’s answer, the one shared by his fellow alumni in their crusade for semiconductor export controls, is that compute deserves this pride of place. In the struggle with Chinese AI, accordingly, the US must ply a twofold strategy of investing in chip manufacture at home and restrictions on chip exports abroad. The DeepSeek story shows how this monomaniacal focus on compute scaling backfired: it turns out that Chinese firms can match or surpass American deep learning models while greatly reducing the scale of compute involved, by instead going all in on algorithm efficiency. In their speculation on the crucial leg of the AI triad, in other words, Buchanan and the rest bet on the wrong horse.
In misplacing their trust in processing power, of course, they were hardly alone. In a more literal sense, this was also the bet made by the many investors in chip manufacturers like NVIDIA, whose confidence in the primacy of compute buoyed the company from $6 per share at the start of 2020 to $144 in early 2025. Between these speculators on the one hand and the likes of Khan and Matheny on the other, however, lies a crucial difference: when DeepSeek proved the former wrong, the loss on their part was immediate. Between 23 and 27 January, NVIDIA’s share price fell from $144 to $118, the largest drop of such magnitude in such a short span of time in the company’s history and part of a one-day $1 trillion dollar stock market rout.
The network of policy advisors around Khan, Matheny, and others in Washington suffered no such immediate market consequence for its mistaken forecast. Nor, given its low public profile and apparent intact survival across the turbulent Biden-Trump switchover, does it seem vulnerable to electoral retaliation. The clique, that is to say, promises the worst of both worlds: neither private sector accountability for its errors in the form of market self-correction nor public political accountability in the form of effective recall. Having charmed and elbowed its way into the halls of power, it is there to stay, come what may.
The watchword of the current administration is cutting federal fat wherever it is to be found. The collective salaries of this group’s Washington representatives are unlikely to sum to much, in the scheme of the full national budget. But their work has exacted costs all the same. Economic costs, from the gratuitous excision of a prime export market. Geopolitical costs, in the form of pointlessly heightened international tensions, including retaliatory export controls on the part of China. And costs in the very arms race with China for which it was intended, by spurring the accelerated Chinese development it was supposed to hamper. If any federal personnel in recent years have offered a poor return on investment, it is the policy engineers of this blunder.
Some of the most eloquent argument for such an algorithm-leaning strategy before the January DeepSeek news, ironically, came from CSET itself. While there were affiliates of the Center to protest in the wake of the news (unconvincingly, in the author’s opinion) that DeepSeek-R1’s impressive showing relied on China’s dwindling and limited supply of American chips, and thus spoke to China’s feeble future prospects in AI or a need for doubling down on chip controls, in December of last year CSET analyst Jack Corrigan wrote in defence of prioritising algorithms and talent over pure compute. “In calling on federal agencies to double-down on this single type of AI system [‘frontier AI’ based on massive compute], one that only incumbent tech giants can produce,” Corrigan wrote presciently if belatedly in December of last year, “these leaders may inadvertently stifle the domestic AI industry and leave the United States exposed to technological surprise from competitors like China.” This is of a piece with a broader strategy he had advocated since 2021, emphasising talent development and retention as a tool for American AI dominance.
This modicum of internal dissent, however, does not undermine the broader failure of CSET’s core policy circle. That the Center offered a countervailing opinion at the eleventh hour does not erase the fact that for years the circle of analysts and advisors concentrated around CSET threw its institutional weight behind the catastrophic compute-first paradigm. That CSET has carted out this internal critic as its spokesman in the aftermath of DeepSeek-R1 should be seen for what it is: cashing in on a disingenuous hedging of its bets.
Under Biden, the Washington foreign policy establishment embraced with open arms an upstart group of EA-backed analysts on trade and AI. Its great contribution to this emerging field of policy, the system of microchip export controls inaugurated in October 2022, has now careened headlong into uncomfortable realities. If this administration, or any future administration, takes the topic seriously, it must accordingly readjust its impression of this clique to have misguided its predecessor.