W07 - Some Reflections Triggered by DeepSeek

DeepSeek shows us what makes a technology truly compelling: high openness and low cost fundamentally accelerate AI’s historical progress. There’s a Jevons paradox from the Industrial Revolution that argues improving resource efficiency can actually increase total consumption of that resource. DeepSeek seems to reproduce that paradox: any new technology that sharply reduces cost will quickly capture and expand the market, and the ultimate result is increased total demand for compute and energy. This helps explain NVIDIA’s dramatic swing earlier this year.

DeepSeek’s viral rise has produced an information explosion. The challenge is filtering noise: what information matters and what can be ignored. With the era of large language models upon us, everyone starts from a similar baseline—what will the next wave of talent competition look like?

High-quality sources are the most effective response to information overload. Even great material degrades as it’s passed along; most upstream information is inexpensive to obtain, but we’ve grown accustomed to being passively fed content.

Technology changes rapidly; only what endures stands as classic. When I read Hackers and Painters last year, essays from over twenty years ago still produced many Aha moments. Topics like the hacker’s approach to life and work, what makes good design, and what defines a good programming language remain perennially valuable. I recently found a rare out-of-print book called On the Edge, which recounts fragments of Dave Cutler’s work on Windows NT and illustrates his mercurial nature and his ability to both inspire and drive a team. Cutler is called “Silicon Valley’s greatest kernel programmer”; in his seventies, he’s still writing code at Microsoft.

Someone on X raised a “70% problem”: non-programmers using AI can quickly complete about 70% of a coding task, while the final 30% yields diminishing returns, is hard to close, and often leaves AI to fix problems introduced by AI. What does it take to finish that remaining 30%? Understanding the business and customers, architectural judgment, accumulated thinking patterns, and a deep grasp of the fundamentals behind AI-assisted coding.

In a 1:1 with a classmate, we dug into what to learn when picking up a new language. It’s certainly not memorizing and mechanically applying a second Red Book—many constrained concepts and rules are low-value knowledge that doesn’t transfer. Don’t get lost in tools and rules; focus on the knowledge that generates true insight.

Last updated