DeepSeek

The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence

by Sebastian Mallaby;  · 30 Mar 2026  · 607pp  · 161,998 words

on a fresh dimension. A new contender appeared out of left field: a Chinese AI lab called DeepSeek. An offshoot from a Chinese hedge fund, DeepSeek achieved instant celebrity with a reasoning system called R1. DeepSeek became a sensation for three overlapping reasons. The first was geopolitics. Chinese labs had been working on

…

in 2022, had been intended to forestall precisely this sort of catch-up; the embargo’s evident failure compounded the US sense of vulnerability. Whether DeepSeek had triumphed by buying smuggled semiconductors, or by accumulating powerful ones that the Biden team had initially not banned, was almost beside the point. One

…

actually worries me.”[12] Chinese strategists reveled in this discovery as much as American ones fretted about it. On the day of R1’s release, DeepSeek’s boss met China’s premier, Li Qiang, the country’s second-in-command, to discuss how Chinese labs could overtake US ones. The timing

…

of this conversation amplified its power: That same day, President Trump was inaugurated. Even without DeepSeek, the Trump administration would have been less inclined to restrain AI development than the Biden team had been. Now, with China’s AI momentum dramatically

…

Three Mile Island or Chernobyl—the prospects for slowing down the AI race had shrunk to roughly zero. The second reason for DeepSeek’s impact involved its supposed efficiency. DeepSeek said that it had built its models on a shoestring budget—its V3 system, which preceded R1, had allegedly cost less than

…

were falling steadily across all labs: Gemini’s jump from its expensive Ultra model to its nimble Flash models represented a huge gain in efficiency. DeepSeek’s engineers had added their own innovations, to be sure. But their achievement was a confirmation of the global trend toward better engineering and superior

…

cost-adjusted performance. It was not a disruption of it.[15] Yet whatever the truth of DeepSeek’s claims, many Western observers initially took them at face value. Commentators imagined that, since the US embargo had deprived the Chinese labs of chips

…

, DeepSeek had indeed been driven to squeeze miraculous performance out of limited hardware. This misapprehension triggered a sell-off in semiconductor stocks: If Chinese labs could

…

soon cut back on their orders from AI chipmakers. On January 27, Nvidia, the industry leader, saw its stock drop by a shocking 17 percent: DeepSeek was hailed as the disruptor of the disruptors. It took about a week for investors to absorb the real story, and for Nvidia’s stock

…

.[16] Even if R1 had not been miraculously cheap to build, it was certainly cheap to use the model. Following Meta’s open-weight example, DeepSeek allowed anyone to download the program and adapt it freely. Going beyond Meta, and indeed beyond the other Chinese labs

…

, DeepSeek allowed customers to use the full model, hosted at DeepSeek’s expense, in exchange for a paltry payment. This price-slashing persuaded many Western customers to embrace DeepSeek and ditch US providers: In late January and early February

…

, DeepSeek’s chatbot topped the free-apps download chart in Apple’s US app store.[17] China was now not only a contender in the

…

race to build frontier AI first. It was a plausible winner in the race to supply it globally. The third reason DeepSeek commanded public attention was perhaps the most interesting. Unlike frontier labs in the United States, which increasingly restricted what they published so as to protect

…

their competitive edge, DeepSeek was transparent. It posted a paper in English on the research site arXiv, explaining how R1 worked and highlighting a variant called R1-Zero. Scientists

…

, which had mastered the game of Go purely by playing against itself, R1-Zero grew strong by learning directly from trial and error—from experience. DeepSeek’s Zero system was not quite ready for prime time—it switched confusingly between languages, for example. But, guided by nothing but a reward signal

…

one reasoning path, backtrack if it hit a dead end, then take a fresh run at the problem. On mathematical benchmarks, and under certain conditions, DeepSeek’s Zero system actually beat o1. One time during training, R1-Zero interrupted itself midway through a reasoning problem. The stream of mathematical notation appearing

…

merely capable of thinking. It was capable of thinking about its thinking. For all intents and purposes, it was self-aware.[19] * * * • • • For Hassabis, the DeepSeek shock crystallized his wildly paradoxical experience of the AI race. On the one hand, he was succeeding in his mission. By the spring of 2025

…

AI development was spinning out of control. The bidding war for scientific talent and the scramble to build new data centers were increasingly wild. Following DeepSeek, a slew of Chinese labs released powerful models, mocking Hassabis’s long-ago hopes of a “singleton” scenario. It was a hard mixture to process

…

reach of Western regulatory restraint, even if one imagined that regulators had the gumption to restrain anything. In sum, and contrary to Hassabis’s hopes, DeepSeek and its followers signaled that the world would sprint over the threshold to AGI with no coordination whatsoever. Hassabis was honest about this predicament. A

…

month after the DeepSeek shock, on a panel at the World Economic Forum in Davos, he debated the dangers with Yoshua Bengio, acknowledging that RL agents of the sort

…

. BACK TO NOTE REFERENCE 12 Dave Lawler, “White House ‘Looking into’ National Security Implications of DeepSeek’s AI,” Axios, January 28, 2025, axios.com/2025/01/28/deepseek-ai-national-security-trump. BACK TO NOTE REFERENCE 13 DeepSeek claimed that it used a cluster of more than two thousand Nvidia chips to train

…

Huang, “Silicon Valley Is Raving About a Made-in-China AI Model,” The Wall Street Journal, January 27, 2025, wsj.com/tech/ai/china-ai-deepseek-chatbot-6ac4ad33. BACK TO NOTE REFERENCE 14 Between ImageNet in 2012 and 2024, frontier AI models were estimated to have become roughly ten times more

…

efficient every two years. See Leopold Aschenbrenner, “Situational Awareness,” June 2024, situational-awareness.ai. BACK TO NOTE REFERENCE 15 Following the DeepSeek shock, Nvidia’s stock experienced further turbulence relating to the Trump administration’s tariff offensive. BACK TO NOTE REFERENCE 16 Lawler, “White House ‘Looking into

…

’ National Security Implications of DeepSeek’s AI.” BACK TO NOTE REFERENCE 17 Daya Guo et al., “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning,” arXiv, January 22, 2025, arxiv.org/abs/2501.12948. BACK

…

Veo 2 and, 367, 369 walk-away plan of, 234–36 women at, 89, 410n39 Zuckerberg and, 132–34 Deep-Q Network (DQN), 108–9 DeepSeek, 369–73, 439n14 Defense Production Act, 331 dense neural network, 347 D. E. Shaw Research, 263–64 Dewey, John, 227 digital brain, 56 direct folding

…

, 185 on deduction and induction, 28–29 DeepMind equity shares of, 83, 414n20–21 DeepMind split proposal and, 240, 246–48 on DeepMind valuation, 138 DeepSeek competition for, 373 Elixir, 405n2 AI ambitions for, 35 closing of, 42 cofounding, 32–34 DeepMind benefitting from, 109 at Electronic Entertainment Expo, 38–40

…

Putin, Vladimir, 246 Putnam, Hilary, 26, 404n10 Q Q-learning, 101–2, 106 Q* project, 356, 359 quantum mechanics, 389–94 R R1, DeepSeek, 369–72 R1-Zero, DeepSeek, 371–72 Radford, Alec, 208–10 radiology, 188–89 Rae, Jack, 346, 348–50, 368, 430n11, 437n10 Gopher frustrations of, 290, 293 OpenAI

The Age of Extraction: How Tech Platforms Conquered the Economy and Threaten Our Future Prosperity

by Tim Wu  · 4 Nov 2025  · 246pp  · 65,143 words

high-performance graphic processing units (GPUs) and that GPT-4 required over 10,000 petaflop/s-days.[14] (That said, the debut of China’s DeepSeek in 2025 suggested LLMs might be built with fewer computing resources.) And of course, at the frontier, you needed talent. Geoffrey Hinton won the Nobel

…

emotional intimacy or companionship. But emotions are irresistibly powerful drivers of business models, and it seems reasonable to imagine that a firm like Google or DeepSeek may find itself tempted to make subtle but deliberate efforts to cultivate bonds between us and our robot helpers. * * * — We have largely dwelled on the

…

, 159–71, 173–74 Chamberlin, Edward, 137 Charles River Bridge, Boston, 168–69 chatbots, 20, 88, 90, 101 Anthropic, 99 ChatGPT, 94–95, 98–99 DeepSeek, 94–95, 101 ELIZA, 90 Grok, 79, 99 LLaMA, 99 Replika, 100–101 Chavez, Hugo, 124, 155 Chicago School of Antitrust, 27, 30 China, 13

…

, 68, 85–88 The Death and Life of Great American Cities (Jacobs), 19 Deep Blue computer, 92 deep learning, AI, 93–94 DeepMind firm, 98 DeepSeek chatbot, 94–95, 101 de Garton, Ms., 162–63 Denmark, land ownership in, 127–29 dependence, 72, 140 convenience vs., 80 emotion-based, 100–102

The Revolutionary Center: The Lost Genius of Liberalism

by Adrian Wooldridge  · 7 Apr 2026  · 342pp  · 129,097 words

the giants will be able to survive. The supposedly revolutionary start-ups that come along occasionally and shake up stock markets, such as China’s DeepSeek, are eventually eaten up by the giants: to survive, you need thousands of computers linked to the cloud, thousands of highly paid researchers who can

…

despair xxi, 246 Deaton, Angus 246 debate, open/freedom of xvii, 40–41, 46, 54, 56, 92, 184, 187, 214–16, 313, 317 deepfakes 215 DeepSeek 204 defence spending x, 293–4, 300, 301 deference, collapse of 141 Defoe, Daniel 26 ‘defund the police’ 191 DEI (diversity, equity and inclusion) 137

The Economic Consequences of Mr Trump: What the Trade War Means for the World

by Philip Coggan  · 1 Jul 2025  · 96pp  · 36,083 words

June 2024, ourworldindata.org/data-insights/solar-panel-prices-have-fallen-by-around-20-every-time-global-capacity-doubled 53 David Pierson and Berry Wang, ‘DeepSeek is a win for China in the A.I. race. Will the party stifle it?’, The New York Times, 2 February 2025, nytimes.com/2025

…

/02/02/world/asia/deepseek-china-ai-censorship.html 54 Rogier Creemers and Louise Marie Hurel, ‘Limits of economic deterrence in the US–China tech competition’, RUSI, 14 March 2025

Gilded Rage: Elon Musk and the Radicalization of Silicon Valley

by Jacob Silverman  · 9 Oct 2025  · 312pp  · 103,645 words

that the United States seemed at risk of losing. While Silicon Valley was floundering, their competitors in China shocked the world with the release of DeepSeek, an AI model reportedly as strong or better than its Western counterparts, made at a fraction of the cost. China’s BYD was also producing

The Means of Prediction: How AI Really Works (And Who Benefits)

by Maximilian Kasy  · 15 Jan 2025  · 209pp  · 63,332 words

years. It is this scale that enables transformers to effectively store a compressed copy of the entire internet. (Recent successes of the Chinese AI company DeepSeek suggest that this scale might be reduced somewhat without loss of performance.) Figure 6 The size of large language models Source: Wikipedia, “Large Language Model

Enshittification: Why Everything Suddenly Got Worse and What to Do About It

by Cory Doctorow  · 6 Oct 2025  · 313pp  · 94,415 words

. It gave us the TikTok ban, and it’s giving America’s bloated, overcapitalized AI giants political cover as they attack nimbler Chinese rivals like DeepSeek. Antitrust is global, but it’s especially powerful in the United States, thanks to a mix of both government and private action. While I was