​(view in browser)​

Dear Reader,

Welcome to the 280th issue of the Data Science Briefing! We're happy to be back after a much-needed time off.

From trust in classrooms to the nuts and bolts of next-gen text models, this week’s links trace a through-line: AI only earns its keep when we surface its blind spots. A data-rich report on Nigerian classrooms shows an adaptive-tutoring pilot compressing two years of learning into six weeks. A forensic tour of Anthropic’s newly published Claude 4 system prompt pulls back the curtain on vendor self-regulation: the prompt reads like a catalogue of past model misbehaviors, capped with concrete tips on safe prompting and a standing order not to be a sycophant. On the technical frontier, we explore how diffusion language models can reduce latency by generating blocks of tokens in parallel, yet still struggle with long contexts and chain-of-thought reasoning.

On the academic front, we examine how large language models interacting as a “society” spontaneously invent shared jargon and collective biases, echoing the sociolinguistic drift observed in real communities. We then zoom ​in on the transformer, before widening the lens again, to explore how the seemingly disparate embedding spaces in vision, speech, and text all share a universal geometric scaffold.

This week's book is "Prompt Engineering for LLMs" by J. Berryman and A. Ziegler. You can find all the previous book recommendations on our website. In this week's video, we have a look at 60 years of Reinforcement Learning.

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

"Prompt Engineering for LLMs" by J. Berryman and A. Ziegler is an essential resource for anyone working with large language models. The authors expertly position prompt engineering not merely as writing effective prompts but as a crucial component throughout the entire application development lifecycle. By balancing technical depth with practical accessibility, they create a guide that serves both newcomers and experienced practitioners in the rapidly evolving AI landscape.

The book's greatest strength lies in its practical techniques, which go beyond basic prompt crafting. Readers will discover innovative approaches, such as using log probabilities to quantitatively assess completion quality, generating multiple outputs at varying temperatures, and structuring prompts with multiple roles to enhance focus and relevance. Particularly valuable is the "Little Red Riding Hood Principle," which emphasizes aligning prompts with a model's training patterns to achieve optimal responses.

Beyond techniques, Berryman and Ziegler offer crucial insights into real-world application strategies, including how teams like GitHub Copilot incorporate user feedback for continuous improvement. The authors skillfully explain complex concepts like tokenization and auto-regressive generation while maintaining accessibility for developers who might otherwise struggle with the non-human communication style of LLMs. This balanced approach makes the book an indispensable guide for anyone aiming to build robust, efficient LLM-powered applications in today's AI-driven technological environment.

1. ​Can AI be trusted in schools?​ [economist.com]
2. ​Highlights from the Claude 4 system prompt​ [simonwillison.net]
3. ​Deep learning gets the glory, deep fact checking gets ignored​ [rachel.fast.ai]
4. ​Tokenization for language modeling: Byte Pair Encoding vs Unigram Language Modeling​ [ndingwall.github.io]
5. ​Strengths and limitations of diffusion language models​ [seangoedecke.com]
6. ​What Even Is a Small Language Model Now?​ [jigsawstack.com]
7. ​Welcome to Anthropic's Prompt Engineering Interactive Tutorial​ [github.com/anthropics]

• ​Emergent social conventions and collective bias in LLM populations (A. F. Ashery, L. M. Aiello, A. Baronchelli)
• ​Why bad philosophy is stopping progress in physics​ (C. Rovelli)
• ​Understanding Transformers via N-gram Statistics​(T. Nguyen)
• ​Harnessing the Universal Geometry of Embeddings​ (R. Jha, C. Zhang, V. Shmatikov, J. X. Morris)
• ​A Tutorial on Meta-Reinforcement Learning​ (J. Beck, R. Vuorio, E. Z. Liu, Z. Xiong, L. Zintgraf, C. Finn, S. Whiteson)
• ​From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning​ (C. Shani, D. Jurafsky, Y. LeCun, R. Shwartz-Ziv)
• ​Deep Reinforcement Learning, a textbook​ (A. Plaat)

60 Years of Reinforcement Learning

​

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

​Unsubscribe​