​(view in browser)​

Dear Reader,

Welcome to the 308th issue of our newsletter!

These weeks' links sketch a workplace in which the real edge lies less with the narrow specialist or the AI maximalist than with the person who can connect domains, interrogate outputs, and know when fluency masks uncertainty. One essay argues that “expert generalists” are increasingly valuable because modern problems spill across silos, while several others make the same point from the AI side: coding assistants may speed people up, but they can also dull the slow, formative work through which real programming skill is built; large language models do not reliably produce correct code so much as convincing code; and many of their hallucinations are not mysterious glitches at all, but the predictable result of systems rewarded for answering confidently rather than admitting doubt.

A piece on abstract mathematics suggests that highly theoretical ideas can eventually reshape messy real-world systems, and an essay on academic life warns that knowledge work is being reorganized faster than many institutions are willing to admit. In an AI-saturated world, judgment, breadth, and intellectual honesty are not soft extras but the core of the job.

Our roundup of academic and industry papers points to a future in which AI systems become powerful by becoming more structured, specialized, and accountable to the real worlds they operate in. Several of them show that progress in agents depends less on raw model cleverness than on scaffolding: better tool descriptions, richer contextual infrastructure inside codebases, and multi-agent designs that can coordinate, compete, and adapt under pressure. At the same time, other papers remind us that capability without scrutiny is a dangerous bargain: systems that can generate persuasive text still leave detectable statistical traces, still reproduce stubborn gender biases, and still struggle with the kinds of internal belief modeling that underlie genuine social reasoning.

Set against work on tariffs in the wine trade and the accelerating pace of global warming, the broader message is hard to miss: intelligence is never abstract for long. It shows up in supply chains, institutions, climates, and code, and the real challenge is not simply building models that can act, but building models whose behavior can be interpreted, trusted, and steered in a world already under strain.

Our current book recommendation is "Generative AI Design Patterns" by V. Lakshmanan and H. Hapke. You can find all the previous book reviews on our website. In this week's video, we compare AI Agent frameworks: AutoGen, CrewAI, and​ LangGraph.

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

"Generative AI Design Patterns" by V. Lakshmanan and H. Hapke is a highly practical book for data scientists and machine learning engineers who are past the hype and focused on building systems that actually work. Instead of treating GenAI as a bag of tricks, the authors organize the book around recurring engineering problems, such as hallucinations, retrieval, guardrails, structured output, orchestration, and evaluation, and offer concrete patterns for addressing them. That makes it feel less like a theoretical overview and more like a field guide for turning promising demos into dependable applications.

One of the book’s biggest strengths is a clear bias toward implementation. The pattern-based structure makes it easy to jump to a problem you are facing, understand the trade-offs, and see how a solution might be applied in practice. For ML engineers, that modular, systems-oriented framing is especially valuable because GenAI projects rarely fail for lack of model access; they fail in the messy space between components, data, prompts, tools, and user expectations.

Its main drawback is that it may be a bit much for complete beginners, and some tool-specific details will inevitably date faster than the core ideas. But those are minor complaints in a book that appears to be aimed at serious practitioners rather than casual readers. For anyone building GenAI agents or applications in the real world, this looks like the kind of book that can save time, sharpen thinking, and earn a permanent place within arm’s reach of your keyboard.

1. ​The role of the expert generalist in the workplace [cultureamp.com]
2. ​How AI assistance impacts the formation of coding skills [anthropic.com]
3. ​Can coding agents relicense open source through a “clean room” implementation of code? [simonwillison.net]
4. ​Why language models hallucinate [openai.com]
5. ​Your LLM Doesn't Write Correct Code. It Writes Plausible Code. [blog.katanaquant.com]
6. ​Can the Most Abstract Math Make the World a Better Place? [quantamagazine.org]
7. ​Academics Need to Wake Up on AI [alexanderkustov.substack.com]

• ​The Science of Detecting LLM-Generated Text (R. Tang, Y.-N. Chuang, X. Hu)
• ​Who Pays for Tariffs Along the Supply Chain? Evidence from European Wine Tariffs (A. B. Flaaen, A. Hortaçsu, F. Tintelnot, N. Urdaneta, D. Xu)
• ​Global Warming has Accelerated Significantly (S. Rahmstorf, G. Foster)
• ​Surprising gender biases in GPT (R. A. Fulgu, V. Capraro)
• ​AI Must Embrace Specialization via Superhuman Adaptable Intelligence (J. Goldfeder, P. Wyder, Y. LeCun, R. S. Ziv)
• ​Evaluating Theory of Mind and Internal Beliefs in LLM-Based Multi-Agent Systems (A. Kostka, J. A. Chudziak)
• ​AgentConductor: Topology Evolution for Multi-Agent Competition-Level Code Generation (S. Wang, R. Lu, Z. Yang, Y. Wang, Y. Zhang, L. Xu, Q. Xu, G. Yin, C. Chen, X. Guan)
• ​Codified Context: Infrastructure for AI Agents in a Complex Codebase (A. Vasilopoulos)
• ​Learning to Rewrite Tool Descriptions for Reliable LLM-Agent Tool Use (R. Guo, K. Dong, X. Gao, K. Das)

How to Network As A Scientist

​

All the videos of the week are now available in our YouTube playlist.

​Unsubscribe​