2025 letter | Zhengdong
A DeepMind researcher's journey from repeatedly thinking he was "too late to AI" to realizing compute scaling is the most underrated force in history—and why we're wildly early to a transformation that will define the century.
Read OriginalMy Notes (4)
"the footnotes, the minor characters behind the main characters. From Star Wars: Andor to the various government officials I ran into all around the world and in stories I read, this letter is dedicated to you. Pluralism's lofty values stand because you bear the weight."
"Work hard, but have slack for creativity. To be a good scientist you have to be both arrogant and humble. Care about everything, at the same time that nothing really matters."
"The enormous shortage of ability to compute is distorting our work, creating problems where there are none, making others impossibly difficult, and generally causing effort to be misdirected. Shouldn't this view be more widespread, if it is as obvious as I claim?" - Moravec
this was 1976, thats insane
"Again I thought, it's so over. I'm so late to this AI thing... But in fact I was wildly early."
its so over -> we're back
Summary used for search
TLDR
• Compute has scaled 4-5x per year for 15 years, surviving multiple regime changes through human ingenuity—but the trend itself is more reliable than any individual innovation
• The author watched 1000x more compute transform his embodied AI agents from "jerky and nauseating" to executing tasks with inhuman grace, a pattern repeating across every domain
• First-order effects are clear (models winning Math Olympiads, passing the bar), but second-order thinking about societal implications remains underdeveloped outside AI circles
• Pluralism offers a framework: AI progress looks like destiny but consists of countless real choices, and not everything we think needs to be a cost actually needs to be
• We're at the beginning—ChatGPT is only 3 years old, no 1GW data centers exist yet, and scaling shows no walls in sight
In Detail
Wang traces his decade-long pattern of underestimating AI: discovering AlexNet in 2017 and thinking the field was saturated, joining DeepMind in 2021 thinking language models had peaked, then watching GPT-4 emerge. Each time he thought "it's so over, I'm so late," he was actually wildly early. The key insight: beneath time horizons and scaling laws sits compute itself—a trend that's persisted for 15+ years through multiple regime changes. While each doubling masks immense human ingenuity and faces valid reasons to plateau (data bottlenecks, algorithmic limits, power constraints), the trend stubbornly persists.
The turning point came when Wang scaled an embodied AI experiment 1000x. The agent moved with such natural grace he thought there must be a bug—even if he'd cheated and controlled it himself, he couldn't have performed so well. This is what it feels like when the wave of compute passes over your domain. He connects this to the Bitter Lesson (Sutton, 2019) and traces it back to Moravec complaining in 1976 that everyone was ignoring their real bottleneck: compute. The pattern repeats across robotics, where language models hit 50% zero-shot on benchmarks that took classical methods years to reach 20%.
Wang argues we need serious second-order thinking beyond "AI will be powerful, so we should race/pause." First-order effects are clear: models now do tasks taking humans 4+ hours (up from 9 minutes two years ago), data center investment accounts for 90% of US GDP growth, and surveys put AI as "technology of the century." But implications for politics, economics, and culture remain underdeveloped. He advocates for pluralism—holding contradictory ideas simultaneously—as the framework for navigating AI's future, drawing on Isaiah Berlin's philosophy that highest human values are incommensurable but not arbitrary.