🔮 The Sunday edition #516: Abundance; DeepSeek’s power; Copyright & AI; Nvidia, BYD ftw; China's AI education ++
An insider's guide to the latest in AI and technology
Hi, it’s Azeem with our weekend email.
The DeepSeek moment is becoming more than a moment. The nimble startup is challenging established AI labs with competitive models at a fraction of the cost and sending ripples across the industry. Meanwhile, the idea that liberalism must build to survive is going mainstream. Both stories highlight the inflection point that we’re here to understand: how we respond to technological and societal challenges will define our future. Thanks for reading!
Abundance
and ’s new book is making the rounds. In Abundance, they issue a provocation to liberalism: if it can’t build, it can’t endure. Despite the immense wealth and technological prowess of the US, scarcity—of housing, energy, infrastructure and even opportunity—has become the defining constraint of its politics and future. This scarcity has fuelled zero-sum, illiberal populism but also eroded the liberal project from within.In the words of
reviewing the book, “liberalism has forgotten how to build the things that people want”.I’m writing this in the same week that saw the busiest airport in Europe, Heathrow, shut down because it relied on a single source of power. And when that power was no longer available, everything had to stop. What happened at Heathrow is a symptom of the scarcity mindset baked into modern systems – scarcity of permission, initiative and ownership.
My view is that technology enables building. Technology is, after all, things getting cheaper; it is about reducing the resource cost of creating. But it’s only a catalyst. The necessary condition is the mindset.
See also,
joined me on Friday to discuss the renewed culture of building—across energy, infrastructure, AI and defense—and why we believe this can unlock an abundant future.
Sleepless nights
“I think Sam Altman is probably not sleeping well,” Kai-Fu Lee said a couple of days ago commenting on the threat DeepSeek presents to OpenAI. As a reminder, DeepSeek is delivering competitive models at a tiny portion of the spend compared to OpenAI and other major US labs (read our deep dive for everything you need to know). And it’s causing ripple effects. Tencent is slowing down the pace of GPU roll-out because of DeepSeek. Oil giant Saudi Aramco’s chief shared that DeepSeek is “really making a big difference” internally and companies across China are steadily integrating its foundation models.
DeepSeek seems to have managed to create a research environment where the lab can out-innovate more heavily financed competitors. It only has160 employees, compared to over 2,000 at OpenAI and it is unburdened by the strict commercial or investor pressures that typically demand rapid productization. DeepSeek’s billionaire founder, Liang Wenfeng, bought thousands of high-end chips ahead of export restrictions and deliberately limits commercial activities. He’s content with breaking even rather than chasing growth.
Meanwhile, OpenAI has pivoted towards product, as it hunts for a defensible moat and profitable market position. And the pressures are growing… Open-source may be the foundational kernel of building AI services. After all, the very successful Mac operating system is building on the open-source Linux kernel. And if that wasn’t enough, there’s the reality of end-to-end reinforcement learning blurring the boundary between model and the product – the model is the product.
See also, what you should know about new models and AI development this week:
Models can improve not just by growing bigger or thinking longer, but also through search—generating multiple potential answers and selecting the best one. In an impressive demonstration, Gemini 1.5 matched the specialized reasoning model o1’s performance by creating and evaluating 200 responses to a single query. Models excel at identifying their own flawed reasoning when comparing multiple outputs side by side.