​(view in browser)​

Dear Reader,

​

Welcome to the 297th edition of the Data Science Briefing!

​

​

This week’s links sketch a panorama of “reasoning machines” in theory, practice, and the physical world they’re supposed to understand. Denny Zhou’s LLM Reasoning slide deck digs into what it actually means for language models to reason, curating techniques and benchmarks that go beyond next-token prediction into chain-of-thought, tool use, and structured problem solving. On the applied side, Kosmos positions an "AI scientist” as an autonomous collaborator: it runs day-long cycles of data analysis and literature search, executes tens of thousands of lines of code, and produces fully cited research reports that early users say compress months of work into hours.

​

Netflix’s engineering team brings this agentic, foundation-model mindset into production by wiring a large, shared model into their personalization stack, showing how a single representation layer can feed many recommendation use cases without rebuilding the system from scratch. Meanwhile, Anthropic’s report on disrupting what it calls the first AI-orchestrated cyber-espionage campaign is a sobering reminder that the same capabilities driving autonomous discovery can also be weaponized, forcing us to think about agents, safeguards, and abuse patterns as first-class design concerns.

​

Zoom out further and you see the larger shift: from Yann LeCun’s decision to leave Meta to build a startup focused on “world models,” which explicitly aim to capture the dynamics of the real world, to fresh mathematical work on ocean waves that exposes just how gnarly those dynamics really are, the common thread is clear: next-generation AI won’t just talk about the world, it will need to model and act in it, safely and rigorously.

​

Taken together, these pieces read like a field report from the frontier where “judgment,” memory, and agency are being rebuilt in silicon. Work on the simulation of judgment in LLMs shows how models can mimic human-like decisions on psychology-style tasks while still revealing systematic blind spots, forcing us to ask whether we’re seeing genuine judgment or a convincing mask.

​

​A Nietzschean lens sharpens the stakes: if humans are supposed to be value-creating, self-overcoming agents, then delegating more and more evaluative work to machines risks flattening the very struggle that makes values meaningful in the first place. Meanwhile, new results on transgenerational epigenetic inheritance offer a biological metaphor for how experiences get “written” into weights, datasets, and institutional workflows far beyond any single human lifetime.

​

On the capability side, early science acceleration experiments with GPT-5 and million-step LLM pipelines hint at AI systems that can sustain reasoning and experimentation over astonishingly long horizons, provided we decompose problems into swarms of micro-agents with strong error correction. But security work on adaptive jailbreak and prompt-injection attacks is a sharp reminder that attackers “move second,” and almost all our current defenses crumble once the adversary starts optimizing directly against them, so any story about judgment must include adversarial judgment too.

​

Zooming out, the long-view survey of artificially intelligent agents in the social and behavioral sciences shows that we’ve been using artificial actors as mirrors of society for decades—but LLM-powered agents now blur the line by both modeling and participating in our social systems, turning questions about who gets to judge, remember, and act into urgent design decisions rather than distant philosophy.

​

Our current book recommendation is Sinan Ozdemir’s "Quick Start Guide to Large Language Models". You can find all the previous book reviews on our website. In this week's video, we have an overview of "what is a Laplace Transform?”.

​

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

Sinan Ozdemir’s "Quick Start Guide to Large Language Models" lives up to its name. It moves quickly from core concepts, tokens, context windows, and prompt structure to working patterns like chat apps, RAG, summarization, and lightweight agents. The sequencing is pragmatic: read a chapter, ship a prototype.

​

The standout value for DS/ML folks is its treatment of embeddings and retrieval. Ozdemir shows when embeddings beat fine-tuning, how to chunk and index, and how to trade off accuracy, latency, and cost with clear, reusable checklists. His sections on prompt patterns, tool use/function-calling, and interface design treat prompting like API design, constrain inputs, structure outputs, plan for failure modes, making it easy to slot into existing services.

​

In short: an excellent on-ramp and onboarding text. Pair it with heavier resources for evaluation, alignment, and production-grade deployments.

​

1. ​LLM Reasoning [dennyzhou.github.io]
2. ​The forgotten pioneers of computational physics [physicsworld.com]
3. ​Kosmos: An AI Scientist for Autonomous Discovery [edisonscientific.com]
4. ​Disrupting the first reported AI-orchestrated cyber espionage campaign [anthropic.com]
5. ​Meta's Chief AI Scientist Yann LeCun To Depart And Launch AI Start-Up Focused On 'World Models' [nasdaq.com]
6. ​The Hidden Math of Ocean Waves [wired.com]
7. ​Integrating Netflix’s Foundation Model into Personalization applications​ [netflixtechblog.medium.com]

​

• ​The simulation of judgment in LLMs (E. Loru, J. Nudo, N. Di Marco, A. Santirocchi, R. Atzeni, M. Cinelli, V. Cestari, C. Rossi-Arnaud, W. Quattrociocchi)
• ​Why Nietzsche Matters in the Age of Artificial Intelligence (S. Liu)
• ​Transgenerational Epigenetic Inheritance: Twists and turns in the story of learned avoidance (L. T. MacNeil)
• ​Early science acceleration experiments with GPT-5 (S. Bubeck, C. Coester, R. Eldan, T. Gowers, Y. T. Lee, A. Lupsasca, M. Sawhney, R. Scherrer, M. Sellke, B. K. Spears, D. Unutmaz, K. Weil, S. Yin, N. Zhivotovskiy)
• ​Solving a Million-Step LLM Task with Zero Errors (E. Meyerson, G. Paolo, R. Dailey, H. Shahrzad, O. Francon, C. F. Hayes, X. Qiu, B. Hodjat, R. Miikkulainen)
• ​The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against Llm Jailbreaks and Prompt Injections (M. Nasr, N. Carlini, C. Sitawarin, S. V. Schulhoff, J. Hayes, M. Ilie, J. Pluto, S. Song, H. Chaudhari, I. Shumailov, A. Thakurta, K. Y. Xiao, A. Terzis, F. Tramèr)
• ​Artificially intelligent agents in the social and behavioral sciences: A history and outlook (P. Holme, M. Tsvetkova)

A Multi-Armed Bandit Framework for Recommendations at Netflix

​

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

​Unsubscribe​