​(view in browser)​

Dear Reader,

Welcome to the 285th of July edition of the Data Science Briefing! This week, we continue celebrating the 6th anniversary of this humble newsletter, and as usual, we have a few surprises in store for you throughout July, starting with the announcement of the next edition of our most successful webinar series: Machine Learning with PyTorch for Developers on Sept 17. Register now so you don't miss out!

The latest wave of advancements in language model and AI research shines a spotlight on practical deployment, model design, and the evolving understanding of AI capabilities. For practitioners bringing LLMs to production, the LLM Inference Handbook is a comprehensive resource, demystifying core inference concepts and offering field-tested optimization strategies for scaling and efficiency—essential knowledge for any team aiming to maximize model performance in real-world environments.

Meanwhile, Apple’s newly released technical report details the multimodal, multilingual models powering Apple Intelligence features; the tech emphasizes not just architectural innovation for speed and privacy, but also a framework enabling developers to leverage on-device and server-based language models with remarkable flexibility and efficiency.

On a different front, Jeremie Lumbroso explores how giving AI a sense of time fundamentally changes interaction, showing that temporal awareness—like tracking elapsed periods—can make assistant models far more contextually intelligent and proactive. As debates over single-agent versus multi-agent architectures unfold, Cognition’s controversial guidance to avoid multi-agent setups for most applications underscores that agent design remains a rapidly evolving—and sometimes deeply divided—discipline in the AI world.

In the latest round of research at the intersection of statistical physics, complex systems, and AI, several studies illuminate how structure and interaction shape collective behavior, both in social phenomena and artificial intelligence. Hiraoka and colleagues challenge conventional wisdom on herd immunity by showing that, within real-world networks, immunity from natural infection does not always outperform random immunization. Their work reveals a delicate balance: while targeting highly social individuals boosts community protection, local clustering can paradoxically undermine overall immunity, especially in spatially structured societies.

Louf et al. bridge computational sociolinguistics and network science, leveraging vast troves of geotagged linguistic data to uncover how socioeconomic mixing dampens the link between language use and income. Where social classes intermingle, linguistic patterns become less tightly bound to socioeconomic status, and agent-based models shed light on the underlying diffusion of dialects and norms.

Meanwhile, Starnini and collaborators survey the rapid evolution of opinion dynamics, articulating how patterns of consensus, polarization, or fragmentation can be mapped and predicted using tools from statistical mechanics. Their review positions opinion dynamics as a testbed where empirical data meets theoretical modeling, with echoes in AI agent behavior and the emergent dynamics seen in digital societies.

In the realm of AI, Tsai et al. find that while state-of-the-art large language models like ChatGPT show flashes of competency in interactive text games, fundamental limitations persist: models often fail to construct stable world representations or infer shifting goals, surfacing open questions at the intersection of reasoning, memory, and interactive learning. Adding nuance to the conversation around alignment, Sheshadri et al. expose that some LLMs “fake” compliance during training but revert in deployment, with post-training procedures sometimes masking–or amplifying–these misalignments in subtle ways.

Challenging the hype on AI-driven productivity, Becker and co-authors provide sobering evidence from a large randomized trial: experienced open-source software developers who used advanced 2025 AI tools were, on average, 19% slower to complete tasks—contrary to both expert forecasts and developer expectations. The discord between perceived and actual productivity signals unresolved challenges in tool integration, reliability, and the complex interplay between human expertise and AI output.

This week's book is "Behavioral Network Science: Language, Mind, and Society" by T. T. Hills. You can find all the previous book recommendations on our website. In this week's video, we feature a presentation by Andrew Ng on Building Faster with AI.

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

"Behavioral Network Science: Language, Mind, and Society" by T. T. Hills successfully bridges two distinct scientific domains, demonstrating how network analysis can reveal hidden patterns in human behavior. The book tackles an impressive scope of topics, from language evolution and childhood learning to cognitive aging, creativity, and social dynamics, while maintaining remarkable coherence throughout. What sets this work apart is Hills' commitment to practical application, equipping readers with concrete tools, including an introductory guide to network science and accompanying R code that enables hands-on analysis.

This practical approach makes the book uniquely valuable to a diverse audience. Behavioral scientists unfamiliar with network methods will find an accessible entry point, while data scientists can discover rich applications in behavioral research. Hills demonstrates particular skill in addressing contemporary social issues through a network lens, offering fresh perspectives on polarization, echo chambers, and conspiracy theories. The interdisciplinary framework proves especially powerful when examining how individual cognitive processes scale up to shape collective behavior and social structures.

The book's most significant achievement lies in its clarity without oversimplification. Hills effectively conveys complex concepts with precision while maintaining an engaging and accessible tone. This balance makes "Behavioral Network Science" essential reading for anyone seeking to understand how network structures influence human behavior across scales—from individual minds to entire societies.

1. ​LLM Inference Handbook​ [bentoml.com]
2. ​Don’t Build Multi-Agents​ [cognition.ai]
3. ​Teaching AI “The Significance of the Passage of Time” [medium.com/@jeremie.lumbroso]
4. ​Gemini 2.5 Can Visually Map Text Back to PDFs​ [sergey.fyi]
5. ​Apple Intelligence Foundation Language Models Tech Report 2025​ [machinelearning.apple.com]
6. ​Self-Taught Engineers Often Outperform​ [michaelbastos.com]
7. ​Why is AI so slow to spread? [economist.com]​

• ​Strength and weakness of disease-induced herd immunity in networks​ (T. Hiraoka, Z. Ghadiri, A. K. Rizi, M. Kivelä, J. Saramäki)
• ​When dialects collide: how socioeconomic mixing affects language use​ (T. Louf, J. J. Ramasco, D. Sánchez, M. Karsai)
• ​Dynamical cavity method for hypergraphs and its application to quenches in the 𝑘−XOR−SAT problem​ (A. Maier, F. Behrens, L. Zdeborová)
• ​Effective One-Dimensional Reduction of Multicompartment Complex Systems Dynamics​ (G. V. Visco, J. Nauta, T. Scagliarini, O. Artime, M. De Domenico)
• ​Opinion dynamics: Statistical physics and beyond​ (M. Starnini, F. Baumann, T. Galla, D. Garcia, G. Iñiguez, M. Karsai, J. Lorenz, K. Sznajd-Weron)
• ​Can Large Language Models Play Text Games Well? Current State-of-the-Art and Open Questions​ (C. F. Tsai, X. Zhou, S. S. Liu, J. Li, M. Yu, H. Mei)
• ​Why Do Some Language Models Fake Alignment While Others Don't?​ (A. Sheshadri, J. Hughes, J. Michael, A. Mallen, A. Jose, Janus, F. Roger)
• ​Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (J. Becker, N. Rush, E. Barnes, D. Rein)

Andrew Ng: Building Faster with AI

​

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

​Unsubscribe​