Dear Reader,

Welcome to the 340th edition of the Data Science Briefing.

Proof checking opens this issue. OpenAI published its Navier-Stokes result in two versions, one for people and one for machines, and they do not match. One lemma in the paper claims a bound costing four extra derivatives, and the Lean theorem proves a weaker one costing five. A compiling Lean file proves whatever its own statement says, and nothing about the paper beside it. The checker had its own bad summer. Kernel bugs surfaced in July and August, one of them produced an illicit disproof of the Collatz conjecture, and security researchers running frontier models found them. Every one was fixed, and the 2.5 million lines of the standard library have been rechecked. Terence Tao’s lecture on October 9 asks what mathematicians are for now. Proofs used to be scarce, and solving open problems was always a proxy for something better. Open problems work as lighthouses rather than destinations. Reach one early by machine and the field around it goes sterile. Closer to the ground, the tooling is thinner than the pitch. One developer’s agent split a feature into 10 subtasks and ran all 10 one at a time, then stopped one night to ask what to name a git branch and sat idle until morning. Benchmarks grade the models, and nobody grades the harness. A short reply defends the tools anyway. Picking up a saw does not guarantee a straight cut, and a compiler automates a great deal and stays a tool. The craft half starts with the operation under all of it. A 23-year-old intern invented attention in 2014 for English-to-French translation, and the transformer paper built on it now carries more than 250,000 citations. One of its authors says attention is not all you need. An older essay says the frame is borrowed. Every era explains the mind with its newest machine, and hydraulics ran for 1,600 years. Draw a dollar bill from memory, then draw it with the bill in front of you, and the two drawings are nothing alike. Two posts put the mechanics on the bench. An MLP on 10,000 MNIST rows hits 95.50 percent in 0.8 seconds, and forward-mode differentiation needs 784 passes where reverse mode needs one. The second reuses a model trained on five CIFAR-10 classes. A linear probe on those frozen features scores 0.470 against 0.407 on random ones, and that 0.062 gap is knowledge nobody had to retrain for.

The paper stack opens inside the harness. Run the same research agent twice on one task and the runs plateau at different scores, each stuck in its own patch of the search space. Fork it into parallel trajectories that inherit the workspace but start on a clean context, and that beats a single run by 66.0 percent across 13 long tasks. A second paper trains the harness itself. A proposer rewrites the solver’s harness from execution feedback, lifting a 4-billion-parameter model from 0.32 to 0.62 on 21 unseen task families. So how much harness does a strong model need? Very little. A minimal agent with read, write, and a shell beat every elaborate scaffold on MLE-bench, medaling on 62.5 percent of competitions against 47.1 percent. One piece of scaffolding earns its keep. An entity page per recurring name, linked to every document mentioning it, buys 6.4 to 11.7 points of answer quality on 34 to 57 percent fewer tokens. Reading these systems is the next problem. Forty-one researchers call the printed reasoning a safety window, and a closing one. Latent-space architectures print nothing at all. The attack side is cheap. One 7-billion-parameter model writes 200 jailbreak suffixes in four seconds and hits 99 percent on a safety-tuned open model. People are the harder read. One review sets out when a model can stand in for a human subject and when the substitution breaks the study. Real populations cooperate for reasons outside the payoff table. Couple a donation game to a plain opinion contest and cooperation rises across 2,763,739 network and starting-condition pairs. The closing paper finds the same lesson in dirtier data. Across 58,319 outbreak records and 32 diseases, the most consistent signal was not biology. Reporting falls a median 32 percent for every extra hour of travel to a clinic.

Our latest book recommendation is Vibe Coding by G. Kim and S. Yegge. In this week’s video, we have a lecture on Security in the LLM age.

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