🔮 Sunday edition #531: Tech trees; AGI vs ROI; pricing content; metaphorical control++

An insider's guide to AI and exponential technologies

Hi it’s Azeem,

This week’s edition examines the forces shaping the power dynamics emerging around the most important general-purpose technology right now, AI.

Let’s dive in!

---

AGI vs ROI

China and the US both want to lead in AI, but their strategies could not be more different. As think tank RAND put it:

In Washington, the AI policy discourse is sometimes framed as a ‘race to AGI.’ In contrast, in Beijing, the AI discourse is less abstract and focuses on economic and industrial applications that can support Beijing’s overall economic objectives.

I saw that distinction firsthand on a recent visit to Beijing. In the US, research labs pour resources into ever larger language models, hoping to unlock broad general capabilities. Chinese teams, on the other hand, publish leaner open-source architectures and partner with specialists in areas such as healthcare analytics (Yidu Tech) and adaptive learning (Squirrel AI). Each group builds tools tailored to a single industry’s needs. I talk about this extensively here:

China’s biggest hurdle is computing power. Domestic researchers account for roughly 15% of global AI compute capacity, compared with about 75% in the United States. American export controls on high-end chips have tightened that gap further. As a result, Chinese engineers have focused on efficiency. DeepSeek, a model touted as seventeen times cheaper to run than its closest US counterpart, has attracted interest from budget-conscious buyers in Chile and Brazil—though the highly anticipated R2 model has been delayed due to hardware shortages. Sanctions bite.

The American approach carries its own costs. Mega-models require sprawling data centers and massive energy bills, and only well-funded enterprises can afford closed-source licenses. The recent ‘big beautiful bill’ flips two ways on this. One, it harms clean energy supply which can be up to four times quicker to develop than fossil plants. Two, it lets AI companies expense training compute as R&D. It’s a ton of bricks on one side and a grain of sand on the other.

Ultimately, winning the AI “race” will depend less on peak model capability and more on delivering tangible benefits in education, healthcare and manufacturing. China’s thriving ecosystem of vertical applications gives it a near-term edge. But if U.S. labs meet their scaling targets and sanctions continue to wilt overseas competition, America may yet pull ahead.

---

Meta’s road not taken

Since 2016, I’ve turned to Sid Meier’s technology tree in Civilization to map how innovations branch out in our exponential era. Now former OpenAI board member

Helen Toner

is asking, “Are we on a good branch of the tech tree?” In the wake of ChatGPT’s explosive growth, leading labs are doubling down on transformer architectures. But this narrow focus could stall breakthroughs.

Source: Helen Toner

Meta announced its new Meta Superintelligence Labs this week, led by former Scale AI CEO Alexandr Wang, to chase artificial general intelligence. The team is stacked with researchers who built transformer-based systems and notably left Yann LeCun, Meta’s chief AI scientist, on the sidelines. LeCun has long cautioned that these models often falter at real-world planning and internal world-modeling tasks. It’s unclear whether incremental tweaks will fix these gaps. That uncertainty is exactly why there is value in keeping other branches alive. Google understands this. They fund both heavy LLM investments and blue-sky basic research. Meta on the other hand, risks tunnel vision (and sky-high AI-scientist salaries).

---

The new economics of the web

It’s clear that the web economy is under pressure from LLMs: Google now sends just one referral visit for every 18 pages scraped, and OpenAI only one per 1,500—both down 3–6x in the past six months.

Cloudflare is now trying to give publishers a way to recoup lost ad revenue through “pay-per-crawl.” In its beta, sites can charge a minimum $0.01 for each page a crawler requests. But how does a penny per page stack up against the ad-supported web?