Technology
•
Jun 4, 2025
Mr. DeepSeek Goes to Washington
The coastal information bubbles of American AI, and what happens next
A week after the Presidential election, I walked into the headquarters of the Center for Strategic and International Studies (CSIS) in Washington and sat down in a gallery of suits. We were all there to listen to talks on artificial intelligence and national security. Congressmen, leading academics, and think tank directors spoke. Then came the headliner — Chris Lehane, VP of Global Affairs at OpenAI (a former politico in the Clinton White House).
“The analog to guns, germs, and steel today is chips. It is data. It is energy,” said Lehane, laying out the centerpiece of his presentation. He went on to describe a massive infrastructure battle between America and China, as well as the energy, manufacturing, and hardware strategy that was necessary to win it.
Two months later, on January 21st, the Chinese AI lab DeepSeek released its R1 model and changed the world of AI. R1 is an open-source “reasoning” model capable of using long chains of thought to answer complex scientific questions. The scientific accomplishment put them in an exclusive club with OpenAI and Google at the top of the reasoning food chain (for now).
DeepSeek made headlines with its purported $5.6 million compute cost. Some doubt that figure, but if true (more on that down below) it would mean that the company had figured out a dramatically cheaper way of scaling intelligence compared to every American AI lab, and the Big Tech companies too. For context, Sam Altman has stated that OpenAI spent “over 100 million [dollars]” to train GPT-4; xAI raised billions to cover datacenter costs.
Tech stocks plummeted as the DeepSeek news sunk in. And a cascade of media followed. I got calls from dozens of journalists and think tankers asking me to explain what DeepSeek even was. I was far from the only one getting those questions. The narrative changed overnight. Before: Chinese containment was working, and they were two to three years behind American AI labs. Now: China has almost caught up to America and might actually have more efficient algorithms.
***
The first time I heard about DeepSeek was as far away from professional DC as one can get — a 2023 X post from an anonymous account named TeorTaxes. He linked to a thread describing a “mysterious Chinese company” with an early breakthrough, an open source coding model that could compete with GPT 3.5 (this was 2023). Already, academics and anonymous posters were singing the praises of DeepSeek’s “data cleaning and pruning” methods, trends that became central in 2024.
Over the next year, DeepSeek put together the pieces of R1. The lab released an impressive general model called V2, and made other advances in coding and mathematics. In December 2024, they released another model, V3, at a stunningly low cost. They open sourced the V3 “base” model that R1 modified, and the V3 technical paper is the source of the $5.6 million compute cost.
Addressing the veracity of that figure: the DeepSeek V3 technical paper only stated compute costs. It was not deceptive. Academics made their observations knowing this fact. It is simply a fact that compute is an extremely important cost center for AI infrastructure. That is the reason Nvidia has become the world’s most valuable company. When comparing equal measured compute costs, DeepSeek’s advancements are a technological feat.
In December 2024, academics, not just anonymous posters, were excited. They took note of the algorithmic improvements that allowed faster AI model training at lower cost — which they wanted to adopt in their own experiments. “People treated this as some kind of out-of-the-blue surprise, but it really wasn’t if you were actively following open-source AI. DeepSeek has been publicly releasing open models and detailed technical research papers for over a year,” said Stanford professor Christopher Manning.
“There are now many excellent Chinese large language models (LLMs). At most these companies are six months ahead, and maybe it’s only OpenAI that is ahead at all. It’s a sad state of affairs for what has long been an open country advancing open science and engineering that the best way to learn about the details of modern LLM design and engineering is currently to read the thorough technical reports of Chinese companies.”
But at no point in 2024 did I hear DeepSeek mentioned in Washington. Late January was when everything changed. A major lesson I’ve learned in the nation’s capital is that it takes around two years for academic and industrial common knowledge to migrate to the swamp.
So honestly, the DeepSeek information cycle came faster than I expected. But unsurprisingly, the pre-DeepSeek policy talking points stayed the same. DC remained fixated on export controls and infrastructure projects. The hard truth is that being two years behind means missing out on important chunks of reality. Remember — the DC wisdom was that two years was the difference between Chinese and American AI. As it turned out, it was the difference in time between reality and reality in Washington.
And at a deeper level, there is a difference in appreciation for reality. Whenever I speak to researchers who have studied DeepSeek’s academic papers, they all express genuine respect for the algorithmic discoveries that DeepSeek has made. “The center of gravity of the open source community has been moving to China,” Berkeley professor Ion Stoica told the New York Times. Stoica has led an effort to replicate OpenAI and DeepSeek’s “reasoning” models, Sky-T1. “Their research paper demonstrates that they’ve independently found some of the core ideas that we did on our way to o1,” OpenAI Chief Research Officer Mark Chen said on X. Part of this respect comes from necessity: if academics or industry researchers fail to learn from DeepSeek, they will fall behind.
I rarely hear the same respect from policy circles.
***
If there’s one thing I know about the defense industrial complex, it's that they love being told what they want to hear. That’s one reason why information moves so slowly from coast to coast. And so even though OpenAI was already moving on to a focus on new algorithms and reasoning models, Lehane focused on the aspects of policy that the audience of suits wanted to hear. Moreover, there remains genuine merit in Lehane’s policy ideas — just because algorithms are overlooked in DC doesn’t mean that other proposals are negative.
To be fair, some establishment policy proposals still do more good than harm. Onshoring hardware manufacturing and reforming energy permitting can benefit AI developers of all sizes. But the current Washington debate misses the forest for the trees.
What the DeepSeek moment calls for is a re-evaluation of the win condition. The Effective Altruist fantasy of inventing one big AI to rule them all is childish and unrealistic. It has led to a bull-headed AI strategy that does not know how to define winning. In DC, spending works. And a joint venture Effective Altruists and Industry Incumbents have spent hundreds of millions pushing this winner-take-all narrative.
In the last two years, there’s been a bipartisan move away from the EA narrative. However, the slow pace of DC information flow has left many policymakers asking for someone to please define what “winning against China” would even look like. Fortunately, not all DC academics think alike. Here is one way to define winning: a world run on American AI systems.
In Technology and the Rise of Great Powers, George Washington University Professor of Political Science Jeffrey Ding investigates the economic history of world-shaping “General Purpose Technologies like steam engines, electricity, and computing. He predicts AI will follow the same trajectory.
Last year, Ding sat down with me for an interview. “We're drawn to this notion of a heroic inventor that brings us to this new age of technology,” he told me. “And it's much harder to trace the stories of average engineers just trying to keep systems running and improve them in incremental ways. But I think that type of work is going to deliver the real value.”
Months before DeepSeek launched R1, I asked him what politicians and pundits get wrong about the Sputnik moment. “Which country launched Sputnik first?” he asked in return. “The real question is, was the US or the Soviet Union better positioned to diffuse satellites and the opportunities provided by satellites across the entire economy or across the entire military? History is pretty clear on that front that the US was much better positioned at that adoption side of the competition.”
According to Ding’s research, general purpose technologies go through an incubation period, in which people adapt to new technologies, learning to use them. For example, after electrical circuits were invented, they were not immediately adopted. Instead, knowledge, processes, and products had to be built around them to make electrical circuits useful. Factories needed to be reorganized and rebuilt to take advantage of electricity. After this incubation period, electricity finally yielded increases to productivity and economic growth.
While scientific institutions are crucial, it might be for a reason that’s different than what you think. “One of my findings in [chapter seven] is that the U.S. is well positioned to outcompete China when it comes to the diffusion of AI across its entire economy.”
Incubation periods separate the invention of new technologies from gains in productivity and economic growth. Ding observes that countries speed up diffusion through institution-building and education. That’s how we can outcompete China.
At the same time, diffusion theory is rippling through industry and academia. Yann LeCun, “Godfather of AI”, NYU professor and Chief AI Scientist at Meta has long been an advocate for open source AI as a solution to the diffusion problem. “Efficient technology transfer hinges on trusting relationships between research and development,” LeCun told me. “Open sourcing research prototypes can help convince internal product groups to adopt a tech from their own research lab.”
LeCun is right. Large firms are full of hubris — in my opinion, Meta included. They struggle to adopt new ideas and adapt their business model accordingly. LeCun sees open source as a way for Meta to overcome these barriers. He has led and celebrated efforts for Meta to adopt innovations made by DeepSeek.
Despite that, if the US strategy is to go all in on AI giants, they are going all in on hubris. In recent weeks, that message has united everyone from Lina Khan to JD Vance. “When there isn’t enough competition, our tech industry grows vulnerable to its Chinese rivals, threatening U.S. geopolitical power in the 21st century,” Khan wrote in the New York Times. And here’s Vance at the AI Action Summit: “This administration will not be the one to snuff out the startups and the grad students producing some of the most groundbreaking applications of artificial intelligence. Instead, our laws will keep big tech, little tech and all other developers on a level playing field.”
It’s not just about the US versus China. It’s about where technological progress comes from — giant monopolies, or a vast competitive network?
Vance’s speech marks a re-evaluation of our existing approach. “When a massive incumbent comes to us, asking us for safety regulations, we ought to ask whether that safety regulation is for the benefit of our people or whether it's for the benefit of the incumbent.” Shortly after DeepSeek released their new model, American competitor Anthropic released an essay calling for even more export controls against them. Vance charts an increasing convergence against incumbent-supported policies across party lines: “There have been times when Washington has embraced the argument that certain businesses deserve to be treated as national champions and, as such, to become monopolies with the expectation that they will represent America’s national interests. Those times serve as a cautionary tale.” Khan continues in the New York Times.
DeepSeek has already caused many investors, founders, and academics to revisit the fundamental way they think about AI. But I don’t think we’ve gone nearly far enough. Here is another way where Vance and Khan agree: we shouldn’t expect one company to have all the answers. Americans are now learning more from Chinese open source AI breakthroughs more than they are learning from ours.
Vance’s speech calling for a ‘level playing field’ was celebrated by the ‘little tech’ community of startups and independent researchers. The AI community is in a moment of renewed optimism. But if we’re to fully learn the lesson of diffusion theory, I think there is even further to go.
After the original Sputnik moment, we funded research efforts across the United States to scout, promote, and train new talent. While the Manhattan project gets the attention, the Sputnik era advanced scientific funding across American society. “The release of DeepSeek underscores the critical need for sustained investment in algorithmic research and computational efficiency,” Russell Wald told me. Wald is the Executive Director of Stanford Institute for Human-Centered Artificial Intelligence (HAI), a coalition of academics founded in 2019, which has long argued for the importance of funding fundamental research.
“DeepSeek has challenged the preconceived notions regarding the capital and computational resources necessary for serious advancements in AI, demonstrating that clever engineering and algorithmic innovation can drive major breakthroughs, reducing reliance on brute-force scale and making AI more accessible,” Wald continues.
In other words, if a Chinese trading firm can carry these important ideas to fruition, how many researchers in America and across the world could benefit from more investment in new ideas? There is a rising generation of new AI researchers — at this point, I’ve met more than a hundred of them. At heart, the lesson of the DeepSeek moment is that there is so much more to build.
“At Stanford HAI, we have long advocated for initiatives like the National AI Research Resource to ensure that academic institutions have access to the resources necessary to push the boundaries of AI innovation” Wald told me.
Almost by definition, new ideas come from those who disagree with old ideas. In politics, it is easiest to give into conventional wisdom, whether scaling maximalism or indiscriminate austerity. Conventional wisdom is even harder to defy when there are large incumbents on its side. One of the hardest things to do is to bet on ideas that haven’t even gained a foothold.
Yet the lesson we’re learning, if we’re to overcome that hubris, is that there is a vast unexplored AI frontier ahead of us, which will take decades to be settled. One lesson the Sputnik moment taught us is that it may take being defeated to face reality. Either way, let us hope that we can learn from our mistakes, sooner rather than later.
Technology
•
Jun 4, 2025
Mr. DeepSeek Goes to Washington
The coastal information bubbles of American AI, and what happens next
A week after the Presidential election, I walked into the headquarters of the Center for Strategic and International Studies (CSIS) in Washington and sat down in a gallery of suits. We were all there to listen to talks on artificial intelligence and national security. Congressmen, leading academics, and think tank directors spoke. Then came the headliner — Chris Lehane, VP of Global Affairs at OpenAI (a former politico in the Clinton White House).
“The analog to guns, germs, and steel today is chips. It is data. It is energy,” said Lehane, laying out the centerpiece of his presentation. He went on to describe a massive infrastructure battle between America and China, as well as the energy, manufacturing, and hardware strategy that was necessary to win it.
Two months later, on January 21st, the Chinese AI lab DeepSeek released its R1 model and changed the world of AI. R1 is an open-source “reasoning” model capable of using long chains of thought to answer complex scientific questions. The scientific accomplishment put them in an exclusive club with OpenAI and Google at the top of the reasoning food chain (for now).
DeepSeek made headlines with its purported $5.6 million compute cost. Some doubt that figure, but if true (more on that down below) it would mean that the company had figured out a dramatically cheaper way of scaling intelligence compared to every American AI lab, and the Big Tech companies too. For context, Sam Altman has stated that OpenAI spent “over 100 million [dollars]” to train GPT-4; xAI raised billions to cover datacenter costs.
Tech stocks plummeted as the DeepSeek news sunk in. And a cascade of media followed. I got calls from dozens of journalists and think tankers asking me to explain what DeepSeek even was. I was far from the only one getting those questions. The narrative changed overnight. Before: Chinese containment was working, and they were two to three years behind American AI labs. Now: China has almost caught up to America and might actually have more efficient algorithms.
***
The first time I heard about DeepSeek was as far away from professional DC as one can get — a 2023 X post from an anonymous account named TeorTaxes. He linked to a thread describing a “mysterious Chinese company” with an early breakthrough, an open source coding model that could compete with GPT 3.5 (this was 2023). Already, academics and anonymous posters were singing the praises of DeepSeek’s “data cleaning and pruning” methods, trends that became central in 2024.
Over the next year, DeepSeek put together the pieces of R1. The lab released an impressive general model called V2, and made other advances in coding and mathematics. In December 2024, they released another model, V3, at a stunningly low cost. They open sourced the V3 “base” model that R1 modified, and the V3 technical paper is the source of the $5.6 million compute cost.
Addressing the veracity of that figure: the DeepSeek V3 technical paper only stated compute costs. It was not deceptive. Academics made their observations knowing this fact. It is simply a fact that compute is an extremely important cost center for AI infrastructure. That is the reason Nvidia has become the world’s most valuable company. When comparing equal measured compute costs, DeepSeek’s advancements are a technological feat.
In December 2024, academics, not just anonymous posters, were excited. They took note of the algorithmic improvements that allowed faster AI model training at lower cost — which they wanted to adopt in their own experiments. “People treated this as some kind of out-of-the-blue surprise, but it really wasn’t if you were actively following open-source AI. DeepSeek has been publicly releasing open models and detailed technical research papers for over a year,” said Stanford professor Christopher Manning.
“There are now many excellent Chinese large language models (LLMs). At most these companies are six months ahead, and maybe it’s only OpenAI that is ahead at all. It’s a sad state of affairs for what has long been an open country advancing open science and engineering that the best way to learn about the details of modern LLM design and engineering is currently to read the thorough technical reports of Chinese companies.”
But at no point in 2024 did I hear DeepSeek mentioned in Washington. Late January was when everything changed. A major lesson I’ve learned in the nation’s capital is that it takes around two years for academic and industrial common knowledge to migrate to the swamp.
So honestly, the DeepSeek information cycle came faster than I expected. But unsurprisingly, the pre-DeepSeek policy talking points stayed the same. DC remained fixated on export controls and infrastructure projects. The hard truth is that being two years behind means missing out on important chunks of reality. Remember — the DC wisdom was that two years was the difference between Chinese and American AI. As it turned out, it was the difference in time between reality and reality in Washington.
And at a deeper level, there is a difference in appreciation for reality. Whenever I speak to researchers who have studied DeepSeek’s academic papers, they all express genuine respect for the algorithmic discoveries that DeepSeek has made. “The center of gravity of the open source community has been moving to China,” Berkeley professor Ion Stoica told the New York Times. Stoica has led an effort to replicate OpenAI and DeepSeek’s “reasoning” models, Sky-T1. “Their research paper demonstrates that they’ve independently found some of the core ideas that we did on our way to o1,” OpenAI Chief Research Officer Mark Chen said on X. Part of this respect comes from necessity: if academics or industry researchers fail to learn from DeepSeek, they will fall behind.
I rarely hear the same respect from policy circles.
***
If there’s one thing I know about the defense industrial complex, it's that they love being told what they want to hear. That’s one reason why information moves so slowly from coast to coast. And so even though OpenAI was already moving on to a focus on new algorithms and reasoning models, Lehane focused on the aspects of policy that the audience of suits wanted to hear. Moreover, there remains genuine merit in Lehane’s policy ideas — just because algorithms are overlooked in DC doesn’t mean that other proposals are negative.
To be fair, some establishment policy proposals still do more good than harm. Onshoring hardware manufacturing and reforming energy permitting can benefit AI developers of all sizes. But the current Washington debate misses the forest for the trees.
What the DeepSeek moment calls for is a re-evaluation of the win condition. The Effective Altruist fantasy of inventing one big AI to rule them all is childish and unrealistic. It has led to a bull-headed AI strategy that does not know how to define winning. In DC, spending works. And a joint venture Effective Altruists and Industry Incumbents have spent hundreds of millions pushing this winner-take-all narrative.
In the last two years, there’s been a bipartisan move away from the EA narrative. However, the slow pace of DC information flow has left many policymakers asking for someone to please define what “winning against China” would even look like. Fortunately, not all DC academics think alike. Here is one way to define winning: a world run on American AI systems.
In Technology and the Rise of Great Powers, George Washington University Professor of Political Science Jeffrey Ding investigates the economic history of world-shaping “General Purpose Technologies like steam engines, electricity, and computing. He predicts AI will follow the same trajectory.
Last year, Ding sat down with me for an interview. “We're drawn to this notion of a heroic inventor that brings us to this new age of technology,” he told me. “And it's much harder to trace the stories of average engineers just trying to keep systems running and improve them in incremental ways. But I think that type of work is going to deliver the real value.”
Months before DeepSeek launched R1, I asked him what politicians and pundits get wrong about the Sputnik moment. “Which country launched Sputnik first?” he asked in return. “The real question is, was the US or the Soviet Union better positioned to diffuse satellites and the opportunities provided by satellites across the entire economy or across the entire military? History is pretty clear on that front that the US was much better positioned at that adoption side of the competition.”
According to Ding’s research, general purpose technologies go through an incubation period, in which people adapt to new technologies, learning to use them. For example, after electrical circuits were invented, they were not immediately adopted. Instead, knowledge, processes, and products had to be built around them to make electrical circuits useful. Factories needed to be reorganized and rebuilt to take advantage of electricity. After this incubation period, electricity finally yielded increases to productivity and economic growth.
While scientific institutions are crucial, it might be for a reason that’s different than what you think. “One of my findings in [chapter seven] is that the U.S. is well positioned to outcompete China when it comes to the diffusion of AI across its entire economy.”
Incubation periods separate the invention of new technologies from gains in productivity and economic growth. Ding observes that countries speed up diffusion through institution-building and education. That’s how we can outcompete China.
At the same time, diffusion theory is rippling through industry and academia. Yann LeCun, “Godfather of AI”, NYU professor and Chief AI Scientist at Meta has long been an advocate for open source AI as a solution to the diffusion problem. “Efficient technology transfer hinges on trusting relationships between research and development,” LeCun told me. “Open sourcing research prototypes can help convince internal product groups to adopt a tech from their own research lab.”
LeCun is right. Large firms are full of hubris — in my opinion, Meta included. They struggle to adopt new ideas and adapt their business model accordingly. LeCun sees open source as a way for Meta to overcome these barriers. He has led and celebrated efforts for Meta to adopt innovations made by DeepSeek.
Despite that, if the US strategy is to go all in on AI giants, they are going all in on hubris. In recent weeks, that message has united everyone from Lina Khan to JD Vance. “When there isn’t enough competition, our tech industry grows vulnerable to its Chinese rivals, threatening U.S. geopolitical power in the 21st century,” Khan wrote in the New York Times. And here’s Vance at the AI Action Summit: “This administration will not be the one to snuff out the startups and the grad students producing some of the most groundbreaking applications of artificial intelligence. Instead, our laws will keep big tech, little tech and all other developers on a level playing field.”
It’s not just about the US versus China. It’s about where technological progress comes from — giant monopolies, or a vast competitive network?
Vance’s speech marks a re-evaluation of our existing approach. “When a massive incumbent comes to us, asking us for safety regulations, we ought to ask whether that safety regulation is for the benefit of our people or whether it's for the benefit of the incumbent.” Shortly after DeepSeek released their new model, American competitor Anthropic released an essay calling for even more export controls against them. Vance charts an increasing convergence against incumbent-supported policies across party lines: “There have been times when Washington has embraced the argument that certain businesses deserve to be treated as national champions and, as such, to become monopolies with the expectation that they will represent America’s national interests. Those times serve as a cautionary tale.” Khan continues in the New York Times.
DeepSeek has already caused many investors, founders, and academics to revisit the fundamental way they think about AI. But I don’t think we’ve gone nearly far enough. Here is another way where Vance and Khan agree: we shouldn’t expect one company to have all the answers. Americans are now learning more from Chinese open source AI breakthroughs more than they are learning from ours.
Vance’s speech calling for a ‘level playing field’ was celebrated by the ‘little tech’ community of startups and independent researchers. The AI community is in a moment of renewed optimism. But if we’re to fully learn the lesson of diffusion theory, I think there is even further to go.
After the original Sputnik moment, we funded research efforts across the United States to scout, promote, and train new talent. While the Manhattan project gets the attention, the Sputnik era advanced scientific funding across American society. “The release of DeepSeek underscores the critical need for sustained investment in algorithmic research and computational efficiency,” Russell Wald told me. Wald is the Executive Director of Stanford Institute for Human-Centered Artificial Intelligence (HAI), a coalition of academics founded in 2019, which has long argued for the importance of funding fundamental research.
“DeepSeek has challenged the preconceived notions regarding the capital and computational resources necessary for serious advancements in AI, demonstrating that clever engineering and algorithmic innovation can drive major breakthroughs, reducing reliance on brute-force scale and making AI more accessible,” Wald continues.
In other words, if a Chinese trading firm can carry these important ideas to fruition, how many researchers in America and across the world could benefit from more investment in new ideas? There is a rising generation of new AI researchers — at this point, I’ve met more than a hundred of them. At heart, the lesson of the DeepSeek moment is that there is so much more to build.
“At Stanford HAI, we have long advocated for initiatives like the National AI Research Resource to ensure that academic institutions have access to the resources necessary to push the boundaries of AI innovation” Wald told me.
Almost by definition, new ideas come from those who disagree with old ideas. In politics, it is easiest to give into conventional wisdom, whether scaling maximalism or indiscriminate austerity. Conventional wisdom is even harder to defy when there are large incumbents on its side. One of the hardest things to do is to bet on ideas that haven’t even gained a foothold.
Yet the lesson we’re learning, if we’re to overcome that hubris, is that there is a vast unexplored AI frontier ahead of us, which will take decades to be settled. One lesson the Sputnik moment taught us is that it may take being defeated to face reality. Either way, let us hope that we can learn from our mistakes, sooner rather than later.
Technology
•
Jun 4, 2025
Mr. DeepSeek Goes to Washington
The coastal information bubbles of American AI, and what happens next
A week after the Presidential election, I walked into the headquarters of the Center for Strategic and International Studies (CSIS) in Washington and sat down in a gallery of suits. We were all there to listen to talks on artificial intelligence and national security. Congressmen, leading academics, and think tank directors spoke. Then came the headliner — Chris Lehane, VP of Global Affairs at OpenAI (a former politico in the Clinton White House).
“The analog to guns, germs, and steel today is chips. It is data. It is energy,” said Lehane, laying out the centerpiece of his presentation. He went on to describe a massive infrastructure battle between America and China, as well as the energy, manufacturing, and hardware strategy that was necessary to win it.
Two months later, on January 21st, the Chinese AI lab DeepSeek released its R1 model and changed the world of AI. R1 is an open-source “reasoning” model capable of using long chains of thought to answer complex scientific questions. The scientific accomplishment put them in an exclusive club with OpenAI and Google at the top of the reasoning food chain (for now).
DeepSeek made headlines with its purported $5.6 million compute cost. Some doubt that figure, but if true (more on that down below) it would mean that the company had figured out a dramatically cheaper way of scaling intelligence compared to every American AI lab, and the Big Tech companies too. For context, Sam Altman has stated that OpenAI spent “over 100 million [dollars]” to train GPT-4; xAI raised billions to cover datacenter costs.
Tech stocks plummeted as the DeepSeek news sunk in. And a cascade of media followed. I got calls from dozens of journalists and think tankers asking me to explain what DeepSeek even was. I was far from the only one getting those questions. The narrative changed overnight. Before: Chinese containment was working, and they were two to three years behind American AI labs. Now: China has almost caught up to America and might actually have more efficient algorithms.
***
The first time I heard about DeepSeek was as far away from professional DC as one can get — a 2023 X post from an anonymous account named TeorTaxes. He linked to a thread describing a “mysterious Chinese company” with an early breakthrough, an open source coding model that could compete with GPT 3.5 (this was 2023). Already, academics and anonymous posters were singing the praises of DeepSeek’s “data cleaning and pruning” methods, trends that became central in 2024.
Over the next year, DeepSeek put together the pieces of R1. The lab released an impressive general model called V2, and made other advances in coding and mathematics. In December 2024, they released another model, V3, at a stunningly low cost. They open sourced the V3 “base” model that R1 modified, and the V3 technical paper is the source of the $5.6 million compute cost.
Addressing the veracity of that figure: the DeepSeek V3 technical paper only stated compute costs. It was not deceptive. Academics made their observations knowing this fact. It is simply a fact that compute is an extremely important cost center for AI infrastructure. That is the reason Nvidia has become the world’s most valuable company. When comparing equal measured compute costs, DeepSeek’s advancements are a technological feat.
In December 2024, academics, not just anonymous posters, were excited. They took note of the algorithmic improvements that allowed faster AI model training at lower cost — which they wanted to adopt in their own experiments. “People treated this as some kind of out-of-the-blue surprise, but it really wasn’t if you were actively following open-source AI. DeepSeek has been publicly releasing open models and detailed technical research papers for over a year,” said Stanford professor Christopher Manning.
“There are now many excellent Chinese large language models (LLMs). At most these companies are six months ahead, and maybe it’s only OpenAI that is ahead at all. It’s a sad state of affairs for what has long been an open country advancing open science and engineering that the best way to learn about the details of modern LLM design and engineering is currently to read the thorough technical reports of Chinese companies.”
But at no point in 2024 did I hear DeepSeek mentioned in Washington. Late January was when everything changed. A major lesson I’ve learned in the nation’s capital is that it takes around two years for academic and industrial common knowledge to migrate to the swamp.
So honestly, the DeepSeek information cycle came faster than I expected. But unsurprisingly, the pre-DeepSeek policy talking points stayed the same. DC remained fixated on export controls and infrastructure projects. The hard truth is that being two years behind means missing out on important chunks of reality. Remember — the DC wisdom was that two years was the difference between Chinese and American AI. As it turned out, it was the difference in time between reality and reality in Washington.
And at a deeper level, there is a difference in appreciation for reality. Whenever I speak to researchers who have studied DeepSeek’s academic papers, they all express genuine respect for the algorithmic discoveries that DeepSeek has made. “The center of gravity of the open source community has been moving to China,” Berkeley professor Ion Stoica told the New York Times. Stoica has led an effort to replicate OpenAI and DeepSeek’s “reasoning” models, Sky-T1. “Their research paper demonstrates that they’ve independently found some of the core ideas that we did on our way to o1,” OpenAI Chief Research Officer Mark Chen said on X. Part of this respect comes from necessity: if academics or industry researchers fail to learn from DeepSeek, they will fall behind.
I rarely hear the same respect from policy circles.
***
If there’s one thing I know about the defense industrial complex, it's that they love being told what they want to hear. That’s one reason why information moves so slowly from coast to coast. And so even though OpenAI was already moving on to a focus on new algorithms and reasoning models, Lehane focused on the aspects of policy that the audience of suits wanted to hear. Moreover, there remains genuine merit in Lehane’s policy ideas — just because algorithms are overlooked in DC doesn’t mean that other proposals are negative.
To be fair, some establishment policy proposals still do more good than harm. Onshoring hardware manufacturing and reforming energy permitting can benefit AI developers of all sizes. But the current Washington debate misses the forest for the trees.
What the DeepSeek moment calls for is a re-evaluation of the win condition. The Effective Altruist fantasy of inventing one big AI to rule them all is childish and unrealistic. It has led to a bull-headed AI strategy that does not know how to define winning. In DC, spending works. And a joint venture Effective Altruists and Industry Incumbents have spent hundreds of millions pushing this winner-take-all narrative.
In the last two years, there’s been a bipartisan move away from the EA narrative. However, the slow pace of DC information flow has left many policymakers asking for someone to please define what “winning against China” would even look like. Fortunately, not all DC academics think alike. Here is one way to define winning: a world run on American AI systems.
In Technology and the Rise of Great Powers, George Washington University Professor of Political Science Jeffrey Ding investigates the economic history of world-shaping “General Purpose Technologies like steam engines, electricity, and computing. He predicts AI will follow the same trajectory.
Last year, Ding sat down with me for an interview. “We're drawn to this notion of a heroic inventor that brings us to this new age of technology,” he told me. “And it's much harder to trace the stories of average engineers just trying to keep systems running and improve them in incremental ways. But I think that type of work is going to deliver the real value.”
Months before DeepSeek launched R1, I asked him what politicians and pundits get wrong about the Sputnik moment. “Which country launched Sputnik first?” he asked in return. “The real question is, was the US or the Soviet Union better positioned to diffuse satellites and the opportunities provided by satellites across the entire economy or across the entire military? History is pretty clear on that front that the US was much better positioned at that adoption side of the competition.”
According to Ding’s research, general purpose technologies go through an incubation period, in which people adapt to new technologies, learning to use them. For example, after electrical circuits were invented, they were not immediately adopted. Instead, knowledge, processes, and products had to be built around them to make electrical circuits useful. Factories needed to be reorganized and rebuilt to take advantage of electricity. After this incubation period, electricity finally yielded increases to productivity and economic growth.
While scientific institutions are crucial, it might be for a reason that’s different than what you think. “One of my findings in [chapter seven] is that the U.S. is well positioned to outcompete China when it comes to the diffusion of AI across its entire economy.”
Incubation periods separate the invention of new technologies from gains in productivity and economic growth. Ding observes that countries speed up diffusion through institution-building and education. That’s how we can outcompete China.
At the same time, diffusion theory is rippling through industry and academia. Yann LeCun, “Godfather of AI”, NYU professor and Chief AI Scientist at Meta has long been an advocate for open source AI as a solution to the diffusion problem. “Efficient technology transfer hinges on trusting relationships between research and development,” LeCun told me. “Open sourcing research prototypes can help convince internal product groups to adopt a tech from their own research lab.”
LeCun is right. Large firms are full of hubris — in my opinion, Meta included. They struggle to adopt new ideas and adapt their business model accordingly. LeCun sees open source as a way for Meta to overcome these barriers. He has led and celebrated efforts for Meta to adopt innovations made by DeepSeek.
Despite that, if the US strategy is to go all in on AI giants, they are going all in on hubris. In recent weeks, that message has united everyone from Lina Khan to JD Vance. “When there isn’t enough competition, our tech industry grows vulnerable to its Chinese rivals, threatening U.S. geopolitical power in the 21st century,” Khan wrote in the New York Times. And here’s Vance at the AI Action Summit: “This administration will not be the one to snuff out the startups and the grad students producing some of the most groundbreaking applications of artificial intelligence. Instead, our laws will keep big tech, little tech and all other developers on a level playing field.”
It’s not just about the US versus China. It’s about where technological progress comes from — giant monopolies, or a vast competitive network?
Vance’s speech marks a re-evaluation of our existing approach. “When a massive incumbent comes to us, asking us for safety regulations, we ought to ask whether that safety regulation is for the benefit of our people or whether it's for the benefit of the incumbent.” Shortly after DeepSeek released their new model, American competitor Anthropic released an essay calling for even more export controls against them. Vance charts an increasing convergence against incumbent-supported policies across party lines: “There have been times when Washington has embraced the argument that certain businesses deserve to be treated as national champions and, as such, to become monopolies with the expectation that they will represent America’s national interests. Those times serve as a cautionary tale.” Khan continues in the New York Times.
DeepSeek has already caused many investors, founders, and academics to revisit the fundamental way they think about AI. But I don’t think we’ve gone nearly far enough. Here is another way where Vance and Khan agree: we shouldn’t expect one company to have all the answers. Americans are now learning more from Chinese open source AI breakthroughs more than they are learning from ours.
Vance’s speech calling for a ‘level playing field’ was celebrated by the ‘little tech’ community of startups and independent researchers. The AI community is in a moment of renewed optimism. But if we’re to fully learn the lesson of diffusion theory, I think there is even further to go.
After the original Sputnik moment, we funded research efforts across the United States to scout, promote, and train new talent. While the Manhattan project gets the attention, the Sputnik era advanced scientific funding across American society. “The release of DeepSeek underscores the critical need for sustained investment in algorithmic research and computational efficiency,” Russell Wald told me. Wald is the Executive Director of Stanford Institute for Human-Centered Artificial Intelligence (HAI), a coalition of academics founded in 2019, which has long argued for the importance of funding fundamental research.
“DeepSeek has challenged the preconceived notions regarding the capital and computational resources necessary for serious advancements in AI, demonstrating that clever engineering and algorithmic innovation can drive major breakthroughs, reducing reliance on brute-force scale and making AI more accessible,” Wald continues.
In other words, if a Chinese trading firm can carry these important ideas to fruition, how many researchers in America and across the world could benefit from more investment in new ideas? There is a rising generation of new AI researchers — at this point, I’ve met more than a hundred of them. At heart, the lesson of the DeepSeek moment is that there is so much more to build.
“At Stanford HAI, we have long advocated for initiatives like the National AI Research Resource to ensure that academic institutions have access to the resources necessary to push the boundaries of AI innovation” Wald told me.
Almost by definition, new ideas come from those who disagree with old ideas. In politics, it is easiest to give into conventional wisdom, whether scaling maximalism or indiscriminate austerity. Conventional wisdom is even harder to defy when there are large incumbents on its side. One of the hardest things to do is to bet on ideas that haven’t even gained a foothold.
Yet the lesson we’re learning, if we’re to overcome that hubris, is that there is a vast unexplored AI frontier ahead of us, which will take decades to be settled. One lesson the Sputnik moment taught us is that it may take being defeated to face reality. Either way, let us hope that we can learn from our mistakes, sooner rather than later.
You might also like
