January 21, 2026

Dear Reader,

Welcome to the Jan 302nd issue of our newsletter!

Enterprise AI is officially past the “cute demo” phase: a new usage-driven report by OpenAI shows organizations getting the most value when they move from ad-hoc prompts to deeper workflow integration. But there’s a cost trap here: if you aren’t benchmarking your real prompts across a slate of models, you can end up paying multiples for “premium” outputs that cheaper options match on the metrics that actually matter.

Meanwhile, research on “persona space” by Anthropic argues that reliability isn’t just about accuracy; models can drift in character, and tracking (and constraining) an internal “assistant” direction can help keep behavior stable under pressure. On the markets side, a deep dive into prediction exchanges reveals systematic wealth-transfer patterns where liquidity takers are punished by bias and microstructure, while legacy finance pushes toward 24/7 on-chain settlement via tokenized shares and stablecoin funding.

Aditionally, if you want agents you can trust, the unglamorous foundation is structured outputs: schemas, constrained decoding, and validation, so your “automation” doesn’t implode on malformed JSON.

This week’s research thread is that “multi-agent” is both the superpower and the threat model: the same idea that makes reasoning stronger. Models effectively staging internal debate as a “society of thought” also makes influence ops scarier when it’s weaponized as coordinated swarms that can infiltrate communities, fabricate consensus, and sustain harassment at machine speed.

The security angle is even more blunt: prompt injection isn’t a quirky jailbreak, it’s a structural mismatch between natural-language instruction following and adversarial inputs made worse once you add retrieval, because attackers can bury instructions in the very documents your system is designed to trust. That lands squarely in the agentic RAG world, where planners, tools, and memory expand capability but also expand the attack surface unless you treat retrieval as untrusted and design for containment.

And when you zoom out from models to society, a pair of empirical studies offers a reality check: contact behavior varies sharply with individual and neighborhood socioeconomic factors, while political segregation can be stronger offline than online, a reminder that the most consequential “network effects” often live in geography and institutions, not just feeds.

Finally, there’s a meta-warning for anyone building policy on this literature: high-profile social media research can carry industry ties that aren’t consistently disclosed, which means we need stronger norms and greater transparency if we want trustworthy science guiding high-stakes interventions.

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. In this week’s video, we have a talk on “Personalizing Explainable Recommendations with Multi-objective Contextual Bandits.”

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