Dear Reader,

Welcome to the Oct 21st edition of the Data Science Briefing.

Machines now take on real work, and people still check it. Claude worked inside six drug programs, and on 1,061 pages of manufacturing records it caught every problem human reviewers found, plus 7 more . Claude Duet, a research preview, sends a morning brief and needs your approval to pay a bill. So what does an engineer still need to know? Fundamentals, breadth, and the will to dig deep, one engineer answers. His agent had packed 15 components into a file of about 30 lines. A model’s stated confidence deserves the same doubt. Ask twice, and it can answer 0.85 and then 0.7, so build the score from observed signals and logged outcomes instead. The data series supplies the fundamentals. A GRU with attention translates English to German at 0.315 token accuracy. A small transformer reaches a validation loss of 2.8717 on WikiText-2 characters. A convolutional network names five pet breeds at 91.0 percent on synthetic stand-in data, and a U-Net marks the animal’s pixels. The last two links ask what survives a network outage. Survival LLM runs a local model over about 930,000 offline Wikipedia articles and cites every answer. The web itself forgets, and publishers now block its biggest archive over AI training fears. The block will not stop AI companies, the archive’s founder argues, and the public loses its record. Someone still has to check the work, and someone has to keep the copy.

Generative AI now changes who people work with. Researchers tracked 11,868 workers at one multinational for 13 months after a Copilot rollout. Adopters added about two collaborators each and leaned harder on colleagues they already trusted. Social scientists now study the machines and use them at once, and a new review calls causal inference the models’ weak spot. Writing splits two ways. A model fine-tuned on one author’s complete works beat trained human writers with expert and general readers alike. Detectors caught just 3 percent of its output. Default model prose runs the other way. Nine frontier models lean on rare word combinations and other patterns that slow human readers. The authors trace the habit to training for math and code. Steering gives more direct control. An attention-guided method steered about 95 percent of 512 concepts on Llama-3.1 8B, against under half for earlier methods. The machines write mathematics, too. One survey mines 38 of the 372 problem families in OpenAI’s October release for reusable tools. The stack ends inside living cells. A new framework maps every steady state a chemical reaction network can reach on a fixed energy budget. Complex behavior has a price, and these bounds put a number on it.

Our latest book recommendation is “Build with Claude Code” by T. Jo. In this week’s video, we have a video by Richard Feynman answering the question: Can Machines Think?

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