March 5, 2026

Dear Reader,

Welcome to the 308th issue of our newsletter!

This week’s newsletter connects “how it works” and “why it matters.” On the delightfully nerdy end, one piece tackles the question of whether data has weight, walking through why solid-state drives can (in theory) get infinitesimally heavier as you write to them, even if it’s the kind of “measurement” that mostly lives in physics-trivia land.

On the sobering end, a deep dive into tens of thousands of kernel issues asks a more human question: who tends to introduce bugs, when they slip in, and what review dynamics seem to help catch them faster. If you’re building with coding agents, another guide usefully reframes “prompting” as engineering: break work into checkable steps, force explicit plans, and design tight feedback loops so the agent can self-correct instead of confidently wandering. For hands-on learning, there are two excellent interactive explainers: one that makes decision trees feel concrete (splits, entropy, overfitting) and another that visually demystifies diffusion by turning noisy static into an image you can see emerge.

And finally, a timely reminder that capability cuts both ways: new research shows how LLM-driven pipelines can scale deanonymization by stitching together identity signals scattered across platforms.

On the academic front, the thread running through this week’s papers is that “systems” don’t just enact behaviors, but rather compute, coordinate, and leave fingerprints you can measure. One study uses cause-specific excess mortality to reveal how the pandemic’s toll in rural India didn’t arrive as a single wave, but as a shifting mix of direct COVID deaths and knock-on effects that standard reporting often misses.

Zooming out, another paper asks what it even means for a system to compute, pushing past the metaphor of “everything is information” toward a more careful accounting of who is doing the computing, what counts as input/output, and where the boundaries of inference actually sit. That same boundary-setting shows up in AI: work on discovering multi-agent learning algorithms with large language models treats LLMs less as oracles and more as search engines over strategy space.

But once agents start interacting at scale, their society becomes an object of study too: early social-network analysis of agent communities reveals emergent roles, clustering, and influence patterns that look uncannily familiar, except that the “people” are programs. In parallel, research on the statistical signature of LLMs suggests that generated text carries detectable regularities, which matters for everything from provenance to governance: if synthetic content is measurable, so is its diffusion through a network. And hovering over the whole stack are papers on economic complexity and on bitcoin as a quasi-religious techno-libertarian project, both of which point to a shared lesson: incentives and beliefs are part of the system dynamics, shaping what gets built, what gets adopted, and what narratives become “common sense” long before the metrics catch up.

Our current book recommendation is “Visualizing Generative AI: How AI Paints, Writes, and Assists” by P. Vergadia and V. Lakshmanan. You can find all the previous book reviews on our website. In this week’s video, we compare AI Agent frameworks: AutoGen, CrewAI, and LangGraph.

Data shows that the best way for a newsletter to grow is by word of mouth, so if you think one of your friends or colleagues would enjoy this newsletter, go ahead and forward this email to them. This will help us spread the word!

Semper discentes,

The D4S Team