April 22, 2026

Dear Reader,

This week, we explore the recent shift in how people are thinking about modern ML and developer tooling. One piece revisits skiplists not as textbook curiosities but as practical machinery for fast search and indexing, while another makes the hidden economics of model use painfully concrete by showing how tokenizer changes can quietly inflate costs beyond the official range.

Elsewhere, there is a welcome push toward deeper technical literacy, with a clear walkthrough of spherical harmonics that reminds readers how much real math still sits underneath graphics and simulation work. On the model side, the open release of a coding-focused agentic system and a hands-on test of running a strong local model inside a coding workflow both point to the same trend: capable AI coding assistants are becoming more open, more portable, and more usable outside the biggest closed platforms.

Tying it all together is a sharp argument that the real problem with “slop” is not who wrote the first draft, but whether anyone took responsibility for reading, editing, and standing behind it, a standard that feels especially relevant now that tool calling, interface fragmentation, and open-model interoperability are becoming core engineering concerns rather than side quests.

On the academic front, we paint a picture of where AI and scientific understanding are converging, and where they still resist simplification. Work on infectious-disease forecasting and temporal networks shows just how much prediction depends on the structure of time itself: some outbreaks may be harder to foresee not because our models are weak, but because the underlying dynamics are noisier, more seasonal, or less legible than we would like.

That uncertainty echoes in the renewed scrutiny of replicability in the social and behavioral sciences, where the challenge is not just generating results, but producing findings sturdy enough to survive contact with the real world. Against that backdrop, several papers push on AI’s expanding role as an intellectual and engineering partner, from mapping the design space of code agents to imagining systems that can sustain long-horizon ML research with less human hand-holding. But the most provocative and philosophical pieces refuse to let capability masquerade as comprehension: one argues that simulation is not the same as consciousness, while another asks what, exactly, we are addressing when we “talk” to a language model in the first place.

Our current book recommendation is “Designing Data-Intensive Applications” by M. Kleppmann and C. Riccomini. You can find all the previous book reviews on our website. In this week’s video, we have a lecture titled “How We Build Effective Agents” by Barry Zhang of Anthropic.

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