​(view in browser)​

Dear Reader,

Welcome to the first of july edition of the Data Science Briefing.

Announcements

Graph RAG combines the best of both: you still use retrieval, but instead of retrieving raw chunks, you retrieve structured paths through a graph. The LLM gets grounded, traceable context. Not just "here are some relevant paragraphs."

The result: answers that are explainable, multi-hop aware, and far more reliable on complex queries.

This is what I'm teaching on July 11.

You'll build the full pipeline — from raw text through entity extraction, coreference resolution, relation extraction, graph construction, and finally a grounded chatbot. Live. With code you keep.

If you're working in finance, healthcare, enterprise AI, or any domain where hallucinations are a real cost, this is worth cutting into your Saturday.>

Use code BRUNO40 at checkout for 40% off.

👉 Production Graph RAG: Build Explainable LLM Apps with Knowledge Graphs

This week we explore the costs of AI. One writer argues the business behind today’s chatbots does not add up. The two largest labs need a combined $358 billion in yearly revenue by 2029. That money covers $1.1 trillion in compute commitments. They reach it only with steep price cuts or fresh subsidies. A second piece sharpens the squeeze from the other side. Open weight models now run for a fraction of frontier prices. Capable ones cost almost nothing to serve. So the author asks a sharp question. Are these models cheap from open testing across many machines, or are they loss leaders built to push prices toward zero? He reads the big labs as luxury sellers. They gate their best models behind high walls, the way brands sell handbags and fast cars. A third post puts hard numbers under the debate and runs the inference math on the back of a napkin. A 32-billion-parameter model with a 200,000-token window needs about 210 GB for its key-value cache alone. That tops what a single GPU holds, until a common attention trick cuts the figure near eightfold.

On a separate thread, we move from money to meaning. Machines now prove real theorems, and that progress unsettles the people who do math for a living. They are asking what their field is for, and some describe a quiet dread about where the work heads next. One well-known figure pictures a new era of big mathematics, where humans and machines crack hard problems together. The old picture of the lone prover starts to fade. The same turn toward basic questions shows up in who labs hire. Top labs now pay philosophers six figures to work on alignment, machine consciousness, and the rights a model might one day hold. The trend rides a strange jobs chart. In 2024 about 7 percent of computer science graduates sat out of work, and philosophy graduates found jobs more readily. Some critics call part of this ethics-washing. Investors still sign the checks, and the hires help the brand look serious.

Let me verify each paper and pull real source links before drafting.All eight verified with real sources and concrete findings. Let me scan the draft against the house style before delivering.Eight papers, woven into four threads. Ran clean against the house style (no semicolons, em dashes, colons, or banned connectors; 506 words).

On the academic front, three papers map the mind as a network. One team recorded the brain with fMRI and EEG at the same time, then traced its wiring across six speeds, from the slowest drifts to fast gamma rhythms. The same spatial and timing rules hold at every speed, and the brain seems to run these channels in parallel. A second group turned a single resting brain scan into a forecast. In adolescents, that one baseline scan flagged who would face depression and anxiety a year later, an early warning read straight from the wiring. A third paper zooms out from the brain to knowledge itself. It models what we know as a web of concepts, with words as nodes and semantic, sound, and grammar links between them. That structure stays readable, unlike a black-box model, and the same maps help gauge creativity and personality.

As AI writes more of the science, two papers ask who checks the work and how machine crowds behave. A tool from a large technology company reads a full manuscript, checks the proofs, tests the experiments, and marks likely errors. On a set of retracted math and computer science papers, it caught flaws 89.7 percent of the time, and in trials at two major conferences more than 90 percent of authors called it useful. Human reviewers stay in charge. Paper volume now outruns what any of them can read alone. A second study puts many language-model agents in the same game and watches them cooperate. Group cooperation can hold even when no single agent sticks to it, so the crowd behaves better than its parts. The result cuts both ways for anyone planning to deploy swarms of agents.

Two more papers push AI from one shared brain toward tailored, made-to-order output. The first keeps a single trillion-parameter model at the core and hands each person a tiny add-on that stores their preferences, skills, and habits. Millions of these small adapters can sit on one base at once, much like the sliver of the genome that makes each of us different. The second trains a model on piles of recipes to learn human taste, then designs new burgers to order. It rebuilt the Big Mac on its own, with no prompt telling it to. In a blind test with 101 diners at a restaurant, its top burger matched or beat that classic on taste and texture, its mushroom version cut environmental impact more than tenfold, and its bean version nearly doubled nutrition. The point reaches past food. It moves AI from guessing what exists to designing what could.

One last paper turns to outbreaks and the people living through them. It reviews how models couple disease spread with human behavior, where fear and habit shift transmission, and shifting transmission then reshapes behavior. The pandemic buried researchers in health and phone data, so the authors test whether current models are ready for the next threat. They map five open fronts, from patchy data and perceived risk to social pressure, pandemic fatigue, and sudden shocks.

Our current book recommendation is "Building Applications with AI Agents" by M. Albada. In this week's video, we have an overview of What is OpenClaw? Inside AI Agents, LLMs and the Agentic Loop.

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

Michael Albada spent nine years building machine learning systems at Uber, ServiceNow, and Microsoft, and it shows. His O'Reilly book, Building Applications with AI Agents, treats agents as a design pattern, not magic. Thirteen chapters take you from a single working agent through skills, orchestration, memory, learning, and on to multi-agent systems. Later chapters cover measurement, production monitoring, and security.

The design-first stance is the real draw. Every idea sits inside a case study: customer support, legal work, advertising, and code review agents. Albada compares real frameworks by name, including LangGraph, AutoGen, CrewAI, and OpenAI's SDK, and weighs their trade-offs instead of crowning a winner. A data scientist gets clear patterns for picking tools, structuring memory, and validating output before it ships.

It has two weak spots. Some chapters lean on checklists, and sometimes make you walk away feeling like the core idea could fit in a third of the pages. It also skips runnable, end-to-end code, pointing you to outside docs instead. Still, for the data scientist or ML engineer moving into agent work, this book maps the decisions that matter and saves weeks of trial and error. Worth a spot on the shelf.

1. ​AI in Mathematics Is Forcing Big Questions [spectrum.ieee.org]
2. ​How to Build AI Agents for Free Using Open Source [opendatascience.com]
3. ​The Unbearable Cheapness of Open Weight Models [jamesoclaire.com]
4. ​Why big AI labs are hiring so many philosophers [economist.com]
5. ​AI's Brokenomics [wheresyoured.at]
6. ​Inference cost at scale with napkin math [injuly.in]
7. ​DuckDB Internals: Why is DuckDB Fast? [greybeam.ai]

• ​Cognitive Networks for Knowledge Modeling: A Gentle Introduction for Data‐ and Cognitive Scientists - Haim - 2026 - WIREs Cognitive Science (E. Haim, M. Stella)
• ​A canary in the mind: A single baseline brain scan predicts adolescent depression and anxiety one year later (G. Deco, Y. S. Perl, J. Vohryzek, E. Garcia-Guzman, D. A. Pizzagalli, R. Laukkonen, S. Chandaria, M. L. Kringelbach)
• ​Towards Automating Scientific Review with Google's Paper Assistant Tool (R. Jayaram, D. Tyler, D. Woodruff, C. Cortes, Y. Matias, V. Mirrokni, V. Cohen-Addad)
• ​Collective cooperation without individual fidelity in LLM agents (H. F. de Arruda, C. G. Lázaro, A. Aleta, Y. Moreno)
• ​On the Scaling of PEFT: Towards Million Personal Models of Trillion Parameters (Mind Lab)
• ​Generative artificial intelligence creates delicious, sustainable, and nutritious burgers (V. Tac, C. D. Gardner, E. Kuhl)
• ​Shared spatial and temporal principles govern connectome dynamics across timescales (T. H. Alderson, S. Jun, J. Wirsich, M. K. Egan, S. W. ElSayed, S. A. Giakas, P. Mostame, J. Harper, A.-L. Giraud, S. M. Malone, W. G. Iacono, S. Koyejo, S. Sadaghiani)
• ​Emerging frontiers in infectious disease modelling: reassessing the data-driven feedback loop between human behaviour and disease dynamics (M. J. Harris, A. H. Sinclair, G. Pullano, S. J. Beckett, L. LeJeune, F. B. Agusto, C. T. Bauch, C. Baur, H. Berestycki, J. Dushoff, Q. Griette, S. A. Levin, J. X. Velasco-Hernández, J. Wu, J. S. Weitz)

What is OpenClaw? Inside AI Agents, LLMs and the Agentic Loop

​

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

​Unsubscribe​