Dear Reader,

Welcome to the Aug 19th edition of the Data Science Briefing.

Memory opens this issue. Why do large models miss facts they saw in training? A new study sorts the misses into empty shelves and lost keys. An empty shelf never stored the fact. A lost key means the fact sits inside, out of reach. Across 13 models and 2,150 Wikipedia facts, frontier systems encode 95 to 98 percent of the material. The same systems fail to surface 26 to 34 percent of what they store. Letting the model think first recovers 40 to 65 percent of the stuck facts, and rare facts stay the hardest to reach. The bottleneck has moved from learning to lookup. Netflix now treats recommendation as a language task. Its GenRec ranker writes a member’s viewing history as plain text and scores the catalog with a tuned model. The models span 1 to 10 billion parameters and replace thousands of hand-built features. Post-training took 10 to 40 times fewer labeled examples than the old pipeline. A four-week test on about 10 percent of traffic beat the production ranker on short-term and long-term metrics. The prompt is the new feature vector. The serving hardware is shifting under both stories. Agent workloads now spend 50 to 90 percent of their end-to-end time on tool calls, parsing, and control flow. So the CPU is back in the inference plan. The newest racks pair one CPU with every two GPUs, down from one per eight. Some agent deployments run four CPUs per GPU. One projection puts the extra CPU market at 32.5 to 60 billion dollars by 2030. The Department of Energy joined the open-weight push with Genesis Open Models, a program to train and release models for science. The first model is out, and labs can now contribute scientific text, code, and evaluation sets. Applications for the post-training round close on August 25. The stated goal is to double the output of American science in ten years.

The craft half opens in a theme park. One developer learns dense topics by asking a model to build low-poly simulations in the style of Rollercoaster Tycoon. The model drafts the foundations, checks its own claims, then renders the ride. Chip fabs, rocket engines, and token pipelines become parks you tour cart by cart. Machine prose now carries its own marks. Claude embeds an invisible watermark in generated text and signs images with C2PA metadata. The marks survive copy and paste but fade under heavy edits. The rollout follows the EU’s transparency code for AI content and covers models released from August 2026 on. A mark signals machine involvement, and a clean page proves nothing. The algorithms series lands twice this issue. The first post runs classic graph algorithms over the world flight network, 6,072 airports and 66,933 routes after cleaning. Only 11.94 percent of airports hold 20 or more connections, and those hubs carry the traffic. A breadth-first search out of Frankfurt reaches 2,915 airports within three hops. Dijkstra links New York and Sydney in two hops and 16,035 kilometers. The second post turns seven design tricks loose on a 370,105-word dictionary. Knuth-Morris-Pratt finds 560 hits of “tion” across 20,000 words in 8.5 milliseconds. A Fourier transform multiplies 2,000-term polynomials a thousand times faster than the schoolbook method. Each trick buys the same answer with fewer comparisons.

During one outdoor cocktail hour, 95 guests wore badges that logged any other badge within 1.5 meters every five seconds. The public dataset holds 7,213 contact events across 2,760 pairs, and every guest tagged their relationship to the couple. Do the contact patterns measured in schools and hospitals survive an open bar? The badges exist to answer exactly that, and the authors point the data at contact tracing and the design of social spaces. A second study moves from measuring networks to obeying them. Your friends have more friends than you do, on average, and a lab experiment now shows the cost. In networks of 20,320 subjects picked numbers whose payoff tracked the average pick of the whole group. Popular players sit in more neighborhoods, so players who saw only their neighbors aimed too high. Degree-one players chose 8.71 under local information against 7.46 with a full view, and the gap held at every connection count. Thirty rounds of feedback never fixed it. The bias costs money and persists. Size matters too. Percolation theory describes giant graphs, and a third paper asks what happens at 50 nodes. The team removed nodes from 1,562 small real networks, animal groups included, and watched the largest connected piece. Theory predicts a sharp collapse at a critical point. Small networks fade instead, step by step, and a recursive method built for finite size tracks the fade with errors near 0.0004 on ten-node graphs. The fourth paper couples networks to disease. A new model lets people split attention across different information sources during an outbreak. Infections rise, risk awareness follows, and protective behavior cools the spread. The mix of attention then shapes the epidemic’s course.

The second half works where data breaks or lies. Broken data has a twin, faked data. One team taught language models to impersonate real posters, conditioning on each user’s own history across 20 forums and 16 public channels in 17 languages. The fakes then trained a detector on 263,594 paired real and machine messages. On an outside benchmark, the best detector spots machine accounts at 0.97 area under the curve. Twenty posts per account already carry the signal. Bigger generators write harder fakes, so the detectors need constant retraining. The closing read is a fifteen-year flashback. In 2011, one paper trained a single neural network for four language tasks, tagging, chunking, named entities, and semantic roles, with almost no hand-crafted features. It learned its word representations from mostly unlabeled text and shipped as a free tagger with modest compute needs. Fifteen years on, that recipe runs everything above.

Our latest book recommendation is “Competing in the Age of AI” by M. Iansiti and K. R. Lakhani. In this week’s video, we have a conversation with Jeff Dean on The 1% Rule for Building in AI.

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