January 7, 2026

Dear Reader,

Welcome to the 300th issue of our newsletter and the first one of 2026! Our warmest wishes for a New Year that brings everything you hope for.

We kick off the year by tracing a path from fundamentals to flywheels: on the systems side, there’s a practical case for pushing vector graphics rendering off the CPU and onto the GPU. TRMs BigQuery notes read like a reminder that “performance” is mostly about habits (filter early, scan less, join smarter, and instrument what you spend).

On the model side, one explainer demystifies attention by walking through how inputs get linearly projected into the Q/K/V streams and why that separation matters, while a broader “physics” framing argues we should stop worshipping leaderboard deltas and instead run controlled, synthetic experiments to uncover repeatable laws about what models can (and can’t) learn. Then we zoom out: a concise agentic-AI primer maps the modern stack, such as tools, planning, memory, multi-agent patterns, and when to use which.

Meanwhile, the business press is starting to ask the awkward question that will hang over the next wave of AI adoption: Can the biggest labs turn staggering infrastructure spend into durable profits before investors get impatient? Further, a set of career lessons ties it together with the human layer: obsess over user problems, bias toward shipping, and treat clarity as a senior skill because the best ideas still die in ambiguity.

On the academic front, the theme running through this week’s papers is that we’re building faster than we’re understanding and the bill is quickly coming due.

One study examines how LLMs are reshaping scientific output, raising the uncomfortable possibility that “more papers” can coexist with weaker signals of novelty, credit, and real progress. In parallel, a survey on graph-based RAG makes the case that retrieval won’t stay a flat list of documents for long: as tasks get messier, we’ll increasingly lean on structured relationships (entities, links, provenance, constraints) to ground generation in something closer to a working memory than a search result.

A pointed critique shows how easy it is to “find” meaning in models through sloppy statistical rituals, echoing broader reproducibility concerns familiar to anyone who’s shipped ML into production, where interviews with practitioners underscore that most failures are organizational before they’re algorithmic. Security-wise, new work on adversarial generalization and inductive backdoors suggests we should treat capabilities as an attack surface: models can learn hidden behaviors that survive fine-tuning and appear innocuous under typical evaluations. That helps explain why “just add agents” is not a free win. Scaling agent systems requires a rigorous science of coordination, feedback loops, and failure modes, especially as recursive language models promise longer-horizon computation by repeatedly reusing internal reasoning steps.

Finally, the causality angle argues for elevating LLMs from text predictors to scaffolds for causal structure, while keeping our feet on the ground about what evidence is actually warranted and designing systems that fail loudly when the world disagrees.

Our current book recommendation is “Building AI Agents with LLMs, RAG, and Knowledge Graphs” by S. Raieli and G. Iuculano. You can find all the previous book reviews on our website. This week’s video is a longer lecuture on the Physics of Language Models.

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