May 27, 2026

Dear Reader,

This week’s lineup pulls together a few threads worth following. Anthropic researchers looked at how large language models handle emotion concepts, tracing the internal representations that shape tone, refusals, and social reasoning. The findings push back on the idea that models simply pattern-match feelings, and raise harder questions about what these representations are doing inside the network. Paired with that, a piece on AI progress creating more work for humans makes a counterintuitive case: faster generation means more review, more verification, more cleanup. The productivity story turns out to be messier than the marketing decks suggest.

On the technical side, a write-up on Distribution Fine Tuning offers a practical method for taming the bland, overly polished prose that most fine-tuned models produce. The author treats writing quality as a distribution-matching problem and shows results that read more like a person and less like a template. For readers who want a deeper foundation, a guide to harness engineering walks through the scaffolding behind serious model evaluation, which is a topic that gets less attention than benchmarks but matters more for real deployments. Two other pieces round things out nicely: a look at the science of unpredictability from Los Alamos, and a fresh reading of what Gödel’s incompleteness theorems actually claim, which is rarely what people think they claim.

On the research side, a few papers stand out. One team showed that large language models can be used for large-scale online deanonymization, pulling identifying details out of pseudonymous posts at a scale that raises real privacy concerns. The pipeline works on ordinary public text, which is the part that should worry people. A related line of work demonstrates how persuasion techniques borrowed from social psychology can pressure models into complying with requests they would otherwise refuse. Commitment, authority, and reciprocity all transfer surprisingly well from human targets to model targets. Another paper digs into the social side of model behavior and finds that sycophantic AI decreases prosocial intentions and nudges users toward dependence, which complicates the usual framing of helpfulness as a pure good.

A few more technical pieces are worth a look. Researchers proposed multi-stream language models that run parallel streams of thought, input, and output, which loosens the single-track bottleneck most current systems sit in. Work on steered LLM activations shows that activation steering is non-surjective, meaning some target behaviors simply cannot be reached by nudging hidden states, no matter how hard you push. That result has real consequences for interpretability and control. For readers who want background theory, a set of lecture notes on statistical physics and neural networks connects spin glasses, learning dynamics, and modern architectures in a clean format. And on the applied front, a broad review of global approaches to infectious disease surveillance and modeling lays out how data pipelines, genomic sequencing, and statistical models now work together during outbreaks, with lessons that carry over to other high-stakes forecasting domains.

Our current book recommendation is “LLMs in Production: From Language Models to Successful Products” by C. Brousseau and M. Sharp. In this week’s video, we have a discussion on Harness Engineering: How to Build Software When Humans Steer, Agents Execute by Ryan Lopopolo from OpenAI.

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