​(view in browser)​

Aug 26th

​

Dear Reader,

Welcome to the 333rd edition of the Data Science Briefing.

Announcements

Is your new prompt actually better, or did you just get lucky on a sample of 5 outputs? 🤔

Eyeballing LLM outputs might work for quick prototypes, but shipping to production requires real proof.Join us for a free 30-minute workshop on how to run paired prompt tests, filter out statistical noise, and determine the right sample sizes before deploying. End the prompt debates with data, not opinions.

 

🔗 Save your seat: Register

Linear models remain one of the most useful places to build intuition for machine learning, and a new tutorial walks through the progression from ordinary least squares to Huber and quantile regression, showing how changing the loss function changes what the model learns. Another practical guide tackles a less glamorous but equally important problem: making model evaluation trustworthy. It covers validation strategies, class imbalance, metrics, and hyperparameter tuning, demonstrating why a seemingly strong accuracy score can be misleading and how careful evaluation can produce a much more meaningful result.

There’s plenty here beyond modeling, too. Especially when one looks at what happens when increasingly capable small models move onto phones and other constrained devices, measuring the latency, throughput, cost, and power tradeoffs that matter for real-world inference. Meanwhile, Nvidia’s AI advantage is moving beyond the GPU argues that the competitive advantage in AI infrastructure is increasingly about the systems surrounding the accelerator—memory, networking, CPUs, and orchestration—not simply the GPU itself. For Python developers, HTTPX2 offers a next-generation HTTP client with synchronous and asynchronous APIs plus HTTP/1.1 and HTTP/2 support, while Time Complexity of Operations on Python Built-in Types is a useful reference for understanding the performance characteristics behind everyday operations. For a broader perspective, it asks a surprisingly fundamental question about the relationship between computation, machines, and the theoretical foundations of computer science.

The latest research reminds us that the effects of AI extend well beyond model accuracy and benchmark scores. A study of more than 880,000 texts finds that large language models used as writing assistants can substantially reduce linguistic diversity, making writing more stylistically uniform and potentially erasing signals of individual identity and social context. At the same time, a new analysis of AI detection in academic settings finds that automated detectors can flag legitimate AI-assisted editing while being remarkably easy to evade, raising serious questions about their use as evidence of misconduct. And as AI systems become more autonomous, another study finds that groups of LLM agents can develop collective biases that don’t appear in individual models, with group size itself changing the dynamics of those interactions.

The research also illustrates how machine learning is increasingly being applied to complex systems outside the traditional AI stack. Large-scale experiments with navigation apps show that rerouting a relatively small fraction of vehicles can increase speeds on congested roads while reducing travel times and emissions, providing a striking example of how algorithmic interventions can influence an entire network. On the infrastructure side, FrKeeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution explores how heterogeneous local hardware can run much larger mixture-of-experts models than conventional assumptions about edge devices would suggest. Meanwhile, two deceptively simple questions get at deeper issues surrounding AI: whether advanced systems require a fundamentally different approach to ethics Ethical Issues in Advanced Artificial Intelligence and what it actually means to say that an LLM has an emotion. Together, these papers point toward a broader challenge for data scientists: understanding not just what models can predict or generate, but how they alter the systems, institutions, and human behaviors in which they operate.

Our latest book recommendation is "Competing in the Age of AI" by M. Iansiti and K. R. Lakhani. In this week's video, we talk about why agentic systems need ontologies.

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

"Competing in the Age of AI" makes a direct argument: AI is not just a tool for speeding up work; it changes how companies learn, decide, and grow. Marco Iansiti and Karim R. Lakhani’s “AI factory” idea explains why data feedback loops and network effects can give digital firms a lasting edge.

For data scientists and machine learning engineers, the book’s value is its business lens. It shows why a strong model means little without data access, product integration, and leadership support. The tradeoff is that it stays high-level: there is little on MLOps, data quality, drift, evaluation, or the hard work of shipping reliable systems. Its focus on large platform companies can also feel distant from smaller or regulated teams.

Even so, the book is a sharp guide to the forces surrounding technical work. It will not teach you how to deploy a model, but it may help you understand why certain AI efforts take hold while others fade after the demo.

1. ​Linear models from least squares to quantiles [stochastic.blog]
2. ​Benchmarking pocket-scale inference [artificialanalysis.ai]
3. ​Claude Team plan for scientists [claude.com]
4. ​Nvidia’s AI advantage is moving beyond the GPU | [techcrunch.com]
5. ​Time complexity of operations on Python built-in types [docs.python.org]
6. ​httpx2: A next generation HTTP client for Python [github.com]
7. ​Does Computer Science Need Computers? [quantamagazine.org]
8. ​Validation, imbalance, and tuning done right [stochastic.blog]
9. ​First classifiers and how to score them [stochastic.blog]

• ​County, district and community-level measles transmission in the United States in 2013−2025 (S. Chen, A. I. Bento )
• ​Ethical Issues In Advanced Artificial Intelligence (N. Bostrom)
• ​The shrinking landscape of linguistic diversity in the age of large language models (Z. Sourati, F. Karimi-Malekabadi, M. Ozcan, C. McDaniel, A. Ziabari, J. Trager, A. N. Tak, M. Chen, F. Morstatter, M. Dehghani)
• ​Urban congestion relief experiments through routing-app interventions (N. Arora, A. Bayen, T. Cabannes, K. Chen, A. R. Kreidieh, Y. Li, M. Nunkesser, P. Ramaswami, A. Tomkins, E. Turkel, S. Vasserman, H. Zhang)
• ​Group size effects and collective misalignment in LLM multi-agent systems (A. Flint, L. M. Aiello, R. Pastor-Satorras, A. Baronchelli)
• ​Large language models do not have emotions (A. Goldenberg, J. J. Gross)
• ​FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution (S. Yang, X. Fan, M. Pan, H. Xi, Z. Wang, S. Sun, K. Keutzer, S. Han, M. Zaharia, C. Xu, I. Stoica)
• ​Why AI Detection Fails for Academic Integrity (J. A. Karr Jr, G. Khvatskii, T. Hua, N. V. Chawla)

Why Agentic Systems Need Ontologies

​

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

​Unsubscribe​