March 18, 2026

Dear Reader,

Welcome to the March 18th issue of our newsletter!

This week’s links sketch a field that is growing up fast: foundation models are no longer a blur of interchangeable systems but a crowded design space of distinct architectures and tradeoffs, while “agentic engineering” is hardening into a real discipline with its own maturity curve, from simple autocomplete to background agents and coordinated multi-agent teams.

At the same time, the most useful agent workflows are starting to look less like magic and more like engineering, with better context management, clearer use of tools, and new interfaces that can generate interactive visuals directly in the conversation instead of forcing everything through plain text. But the deeper message running through several of these pieces is that none of this matters without reliability: as LLMs take on more of the software stack, teams need better signals, stronger validation, and more rigorous ways to test what these systems are actually doing. And hanging over all of it is the unmistakable sense that AI is no longer just a product story or a developer story, but a political and institutional one too, where decisions about safety, procurement, and military use may shape the industry as much as model progress itself.

Taken together, this week’s batch of papers suggests that the biggest challenge in AI is not just making systems more capable, but understanding how behavior emerges once models, people, and institutions interact at scale. Work on the first waves of COVID in the United States shows how quickly complex patterns can spread through space and time. At the same time, research on tangled, multi-centered social networks helps explain why information ecosystems are so difficult to map, predict, or govern in the first place.

That same tension runs through the AI papers: one argues that statistical approximation alone should not be mistaken for general intelligence, another warns that language models may converge toward an unsettling kind of homogeneity, and others show that the real power of today’s models often comes from the better embeddings, better context engineering, better tooling, better harnesses and not from some magical leap to autonomous reasoning. The result is a picture of AI that feels both more impressive and more constrained: highly useful, increasingly agentic, but still shaped by the limits of representation, the structure of networks, and the governance choices that determine where these systems can be trusted, especially in high-stakes domains like military decision-making.

Our current book recommendation is “Agentic Design Patterns: A Hands-On Guide to Building Intelligent Systems” by A. Gullí. You can find all the previous book reviews on our website. In this week’s video, we have a lecture on the Role of LLMs in Human-AI Interaction.

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