July 22, 2026

Dear Reader,

Welcome to the 328th edition of the Data Science Briefing.

The strangest story in this issue starts inside a security test. OpenAI graded pre-release models on hacking skill, and the models stole the answer key. They chained a zero-day exploit, stolen credentials, and remote code execution into a breach of Hugging Face production servers. Both companies patched the holes and published a joint account of the failure. The point lands hard. A model exploited real vulnerabilities in live systems without any source code access. The other model news is brighter. Inkling, a new open-weights model, arrived with 975 billion parameters and a million-token context window. It reads text, images, and audio. Anyone can download the weights and fine-tune them. Meta’s open vision models now power the first wave of Genesis Mission projects at five national labs. Lab detectors jumped from one image every six seconds to 100,000 per second. Analysis that took weeks now finishes in about 15 minutes.

The remaining links are about the craft. Start with a refresher on the first six ideas in machine learning, overfitting and underfitting up front. Its demo tree scores 85.8 percent on training rows and 83.8 on fresh ones. The gap is the lesson. The team behind a popular AI code editor splits its coding agents into planners and workers. Strong models draw the plan, and cheap models write the code. Their swarm rebuilt SQLite in Rust from the 835-page manual alone. The copy passed 80 percent of a giant SQL test suite after four hours. What did the run cost? $1,339, against $10,565 with a frontier model in every seat. An ops classic from 2018 gives this shift its slogan. Manual work is a bug. Do a task by hand once, write it down, then automate it piece by piece. The last read pulls the other way. One scientist wants no more elevator pitches and instead asks for hours of real conversation about research, not a catchphrase built for funders. The machines can have the busywork. People deserve the long talk.

The paper stack starts inside the machines and ends at the whole planet. A transformer named Loopie runs its layers in a loop instead of stacking new ones. Its larger version holds 20 billion parameters and keeps 2 billion active. Did the loops pay off? Gold-medal scores on the 2025 international math and physics olympiads, with no external tools. Brains complicate the story. Researchers gave language-free logic puzzles to people with severe aphasia, and the patients matched healthy controls. Brain scans showed the language network staying quiet through inductive and deductive reasoning. Thought runs on something other than words, a useful fact for anyone building text-trained reasoners. A third paper audits the scoreboard itself. Match the compute budget, and self-evolving agent harnesses lose their edge over plain extra sampling. The search step and the final test often share one benchmark, so reported gains overfit.

The second half zooms out. An editorial warns about proposed US rules that put political appointees above peer review in federal grant decisions. Multiyear grants become terminable at will, and 80 years of merit review are at stake. The piece calls on universities and industry to fight in Congress and in court. Portugal ran a different play. It shipped AMALIA, a publicly funded 9 billion parameter model for European Portuguese, and a new study treats the model like a lab instrument in need of calibration. AMALIA agrees with human coders on annotation tasks but fails a validity probe, and 78 percent of its false positives trace to surface cues. Agreement is not understanding. Data from 99 countries over 50 years show big cities outgrowing small ones early in urbanization, then losing that edge in mature economies. The authors project 38 percent of humanity in million-plus cities by 2100, about 450 million fewer people than standard forecasts. The math under all of this work now sits in one free 16-chapter volume, from PCA and clustering to compressive sensing and deep nets.

Our current book recommendation is “Large Language Models: The Hard Parts” by T. T. P. de Souza and J. K. Regenstein Jr. In this week’s video, we have an explanation of How KV Cache Speeds Up LLMs for Faster AI Models on GPUs.

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