Dear Reader,

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

Machines now work on machines. A new frontier lab wants AI to design the chips that train the next AI, and it aims to cut a two-to-three-year design cycle to weeks. A research post goes one layer down. GPT-6 Astra read and wrote raw neural network weights with no gradients and no training loop. It edited 0.016 percent of a small model’s weights, and math scores jumped from 7.9 to 48.1 percent. Mathematics feels the pressure hardest. OpenAI released more than 700 research papers on October 6, and a guest essay on a mathematics blog tells the field to stop mourning. The essay bets that choosing a worthwhile problem will soon matter more than producing the proof. People bring the machines their own lives. A panel of 73,852 Americans shared 3.4 million chatbot conversations, and health led all 28 subjects at 10.2 percent. Claude now offers lawyers a gallery of ready-made artifacts, from a cite checker that flags wrong quotes to a rent ledger reviewer. The model drafts the first pass, and a lawyer judges the result. Two data-series posts cover the basics. One learns image features from 6,000 unlabeled pictures. The other trains a tiny vision transformer to draw boxes, and it reaches a mean overlap of 0.815 against 0.309 for random boxes. The machines now draft the weights, the proofs, and the briefs. People still choose the problems worth drafting.

The research stack turns on one question. Compared with what? A study of 14.4 billion insurance claims from 244.7 million Americans compares people who tested positive with people who tested negative . Older designs drew controls from anyone without a known infection, and they even linked firearm injuries to the virus. The new design cuts the outcomes attributable to long COVID from 262 to 43 at one to four months, and from 369 to zero after a year. Symptoms past a year are real, the authors stress, but their cause is uncertain. A fairness framework asks whether an epidemic model reproduces who carries the burden, by age, race, or neighborhood. In a New York simulation, vaccines allocated by that measure cut deaths by 73.9 percent, against 68.2 for a split by group size. Industry leaves an environmental fingerprint, and one team reads it remotely. Air pollutants, night lights, and temperature predict steel plant output with an R² above 0.8. The machines need the same rigor. A new review makes the case for testing language models with the methods of experimental psychology, from biases to social skills. One such study explains why model prose tires readers. All nine frontier and near-frontier models used more rare word sequences than human writers, and fine-tuning under a length penalty pushed a model the same way. Good pairs build better representations. In an idealized setting, a new theorem shows that joint-embedding models trained on separate modalities learn one shared representation, up to a rotation. The condition is that each positive pair shows one scene twice. Next-token prediction pairs parts of one document instead. Good pairs reach the alignment of next-token prediction with 42 times fewer samples. A 109-million-parameter text model aligns with image encoders better than every language model tested, up to 72 billion parameters. The math under all of this fits in one free eight-chapter primer, from PCA through diffusion models, GANs, and energy-based models.

Our latest book recommendation is “Build with Claude Code” by T. Jo. In this week’s video, we have a look at 4 Failure Modes to Avoid in Context Engineering for AI Agents.

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