China’s DeepSeek takes lead in artificial intelligence race
One of the great debates in artificial intelligence recently (and really, the past few decades) is on the limits of scaling: to what degree can greater compute alone make AI better? Is there a threshold of compute where AI starts to approach artificial general intelligence, or at least the simulacra of human metacognition?
My own position is that any further leap forward in AI has to come from theoretical mathematical improvements rather than more computation power: OpenAI, Anthropic and others have already slurped up all of the world’s language information and trained on it with leading compute clusters. To evolve, we need to actually use the data we already have in a more effective way — and that means fundamental advancements rather than throwing more chips at the problem.
However, for those who believe that compute (and specifically test-time compute) remains the critical barrier, the example to point to this week is the success of DeepSeek. The company released a preview of its new R1 agent, which shows leading performance against OpenAI’s analogous o1 model. Both R1 and o1 (yes, we are living in the Star Wars universe) expand test-time compute, which means that queries aren’t necessarily answered right away like a traditional web search. Instead, agents develop multiple strategies for answering the user’s question, and then spend time — sometimes as long as minutes — searching for and assessing the best answer.
Thus, if a user asks their agent “what is 2 + 2,” it might respond immediately, but if it is asked a question from the International Mathematical Olympiad, it will deliberately use more time to evaluate different approaches to answering. This is a more-advanced form of reasoning: the agent has to assess what it knows and doesn’t know, and the hope is as its capability to reason improves, the answers are better and less likely to be hallucinations.
Based on benchmarks posted this week, DeepSeek is now the leading model in the world on a range of metrics. Even with the general squishiness of most benchmarks, this represents a landmark: the first time a Chinese large-language model has taken the global lead in artificial intelligence performance, which has traditionally been a race between OpenAI, Anthropic, Google’s Gemini, France’s Mistral and occasionally Meta with its Llama family of models. Even more interestingly, DeepSeek is developed by the quantitative hedge fund High-Flyer Capital Management, which has reached almost $8 billion in AUM since its founding in 2015.
Even more fascinatingly, DeepSeek intends to open-source its R1 model, and it also provides users the model’s “thoughts” as it reasons about a query (something that OpenAI’s o1 model doesn’t do). That openness is not just surprising in the context of China as well as the extreme competition between AI model builders, but also runs counter to every stereotype of quantitative hedge funds the world over.
There’s little reason to expect that DeepSeek will maintain its perch indefinitely, but even seizing the limelight for a short period of time is its own notable accomplishment. Chain-of-thought models like R1 and o1 are proof that there’s still performance to squeeze out of existing systems through looser test-time constraints. That is, until we find a more fundamental improvement.
Europe’s Solyndra has arrived in the form of Northvolt
The 2011 collapse of solar-panel manufacturer Solyndra led to a vituperative crisis around industrial strategy for the United States. The U.S. federal government, which had loaned the company more than half a billion dollars, was accused of squandering the public treasury for private gains, with conservatives lambasting the Obama administration’s funding of an ultimately-failed venture. The Department of Energy conducted a four-year investigation, but ultimately the case served as a Rorschach test for government intervention in the economy. The reality today is that China holds a monopoly position in solar panels and their supply chain, with the U.S. practically non-existent.
This week, a parallel story is playing out in Europe over Northvolt, the continent’s battery champion and answer to Chinese heavyweights like BYD and CATL. The EU’s industrial strategy was comprehensive here. Northvolt received $15 billion in private venture capital and government-subsidized funding, and major European automakers like BMW were early and enthusiastic backers of the company as part of its planned transition to electric vehicles.
Northvolt never really delivered the quantity or quality of batteries that its customers needed though. BMW canceled its $2 billion contract a few months ago as it became increasingly obvious the company was falling apart. This week, it officially filed for bankruptcy protection, with just $30 million cash left and $5.8 billion in debt liabilities. It’s likely to be one of the greatest startup failures of all time, and certainly in Europe which already lacks a dynamic startup scene like the United States.
What lesson to take from this new crisis? Too many incentives or not enough? Too much ambition or not enough? Mario Draghi earlier this year called for Europe to go hard core on productivity improvements to shore up its fragile economy, something I discussed in “Productivity Precipice.” It’s hard for creative destruction to work though when it’s mostly destruction and not creation. The state did what it could to scaffold Northvolt’s rise and protect it from extreme competition from Chinese and Korean manufacturers. Ultimately, though, the company never succeeded in offering a compelling and affordable alternative and at some point, the market always speaks.
The true lesson to me is that so many of these crises are resolved too late. BYD was founded in 1995 and CATL in 2011, while Northvolt arrived only in 2016. Those extra years made a difference, and one should ask why America and Europe didn’t see the importance of batteries years ago and assiduously secure this strategic industry. Solyndra or Northvolt — what’s too late is too late.
Three small comments
- “Strategic Bitcoin Reserve” — there is a policy goal with the incoming Trump administration to build out a crypto parallel to the Strategic Petroleum Reserve. Outside of just my befuddled, “Why?”, I find it amusing that a decentralized currency which was designed to prevent any institution from controlling its value is now transitioning into a quasi-fiat currency to be managed by a centralized government bureaucracy. I get that this Reserve will boost Bitcoin’s price and make a bunch of hodl’ers a lot of money, but what’s the point for America? Undermining the U.S. dollar?
- Spirit Airlines — The ultra-low-cost carrier was ultimately too ultra-low-cost to make its business model work and eke out any form of sustainable profit. Despite the FTC and the Department of Transportation’s tenacious attempts to keep America’s airline industry competitive, Spirit’s failure and the merger of Alaska and Hawaiian means that America is quickly running out of alternative carriers. It might not feel like it, but domestic airfare prices are actually significantly lower than they were in the past — but whether competitive pricing pressure sustains more affordable fares is an open question. But hey, Delta is now serving Shake Shack in the cabin, so it can’t all be bad.
- The indictment of Gautam Adani in New York on a $250 million bribery scheme in India is a landmark in a long-running crisis over the future of the Adani Group. Precipitated by a short-seller report in 2023 from Hindenburg Research that wiped out $153 billion from its market cap, the company has struggled to find its path forward. I have previously covered India’s assassinations and attempted assassinations in Canada and the U.S. in “Extraordinary Extraterritoriality,” and this indictment is a good reminder that extraterritoriality extends just as often in the other direction.
Podcast: Elections, global threats and happy hour with the Riskgaming team
It’s not every day that we can get our distributed Riskgaming team into one podcast studio, but we actually managed to do it from our NYC base, and with some drinks to boot. Joining me was Lux’s Riskgaming director of programming Laurence Pevsner and our researcher, part-time columnist and all around utility handyman Michael Magnani, whose last column was “Protecting America’s Economic Security” back in June.
We talk about the U.S. presidential election and which threats from our AI deepfake election security scenario DeepFaked and DeepSixed actually took place — and which didn’t. We then cover Germany’s sputtering industrial economy, the future of the war in Ukraine and trade tensions with China. Finally, we close out with a discussion of the three threats that the world isn’t thinking about today, and what should be done about them.
🔊 Listen to “Elections, global threats and happy hour with the Riskgaming team”
The Orthogonal Bet: Dave Jilk on AI, Poetry, and the Future of AGI
In this episode, Lux’s scientist-in-residence Sam Arbesman speaks with Dave Jilk, a tech entrepreneur and writer who has started multiple companies including in AI, published works of poetry and written scientific papers. He’s just written a new book titled Epoch: A Poetic Psy-Phi Saga that is an epic poem about the origins of Artificial General Intelligence, told from the perspective of the first such entity. It’s a deeply thoughtful humanistic take on artificial intelligence full of literary allusions.
Sam wanted to speak with Dave to learn more about the origins of Epoch as well as how he thinks about AI more broadly. They discussed the history of AI, how we might think about raising AI, the Great Filter, post-AGI futures and their nature, and whether asking if we should build AGI is even a good question. They finish with a few science fiction recommendations.
🔊 Listen to “Dave Jilk on AI, Poetry, and the Future of AGI”
Lux Recommends
- Sam enjoyed Niko McCarty’s deep dive into “How to Measure Molecules.” “At the time of Rayleigh’s experiments in 1890, most scientists believed that molecules not only existed but had physical, measurable dimensions. Unfortunately, nobody actually knew how to measure them. X-rays had not yet been discovered, and there were no tools sensitive enough to image individual molecules. Physicists at the time estimated molecular sizes based on abstract mathematical equations, rather than experimental observations.”
- For decades, China placed market restrictions on foreign banks looking to build up a presence on the mainland, including Goldman Sachs. Now that the dust of the company’s first decades has settled, the palace intrigue has started leaking out, or in the case of this Financial Times profile, gushed out. “In the early 2000s, Hank Paulson travelled extensively to China seeking access to the country’s financial markets for Goldman Sachs, where he was chief executive. The investment bank’s top brass believed the deal he eventually cut would only be a temporary one. The thinking was that an alliance with Fang Fenglei, a mainland investment banker with peerless connections to the Communist party elite, would endure for only a few years, until the Wall Street institution was permitted to directly own its securities and investment banking business in China. But instead the arrangement lasted for nearly two decades. During that time the economic benefits flowed mainly to Fang, leaving Goldman desperate to unwind the deal.”
- On the topic of China, our regular Riskgaming reader Guy Perelmuter at Grids Capital sent over this lecture by professor Michael Beckley of Tufts on “The End of China's Rise and the Future of World Order.” It’s a great précis on China’s current situation as we enter the next administration.
- I enjoyed David Reynolds’s take on Winston Churchill as the great improviser of 20th century politics in The New Statesman. “Since boyhood, Churchill had yearned to be prime minister. But his road had proved very bumpy: changing parties twice, carrying the can for the Gallipoli fiasco in 1915, being consigned to the political wilderness for much of the 1930s. And then, less than a week after he gained the prize, aged 65, the German guttersnipe shattered Britain’s whole strategy for waging the war. What we now know as Churchill’s ‘finest hour’ was largely frantic improvisation. And he would have to keep improvising throughout the war, and arguably for the rest of his political life.”
- Finally, Sam enjoyed this piece about the 50th anniversary of the Arecibo Message by Nadia Drake, the daughter of astrophysicist Frank Drake of the eponymous equation for the probability of alien life in the universe. “But just as it symbolizes some sort of transcendence, today the Arecibo message is also a poignant reminder of fragility and loss. Since the message left Earth, the telescope that sent it fell into neglect and eventually collapsed. And the Arecibo message’s designer, my father Frank Drake, died. A few months ago, while rummaging through some of Dad’s old papers, I found an early penciled in draft of the message—along with his musings about the information he wanted to convey and correspondence surrounding its creation. I’d of course known of Dad’s role in sending the message for most of my life, but it was the first time I’d seen any of the work that went into making it.”
That’s it, folks. Have questions, comments, or ideas? This newsletter is sent from my email, so you can just click reply.
