| | Getting useful context into an AI agent is becoming its own data problem. This guide from Astronomer looks at how to pull together internal data, keep it fresh, add human review, and capture past decisions so agents have more than just a prompt to work with.
*Message from this week's sponsor, Astronomer |
| | |
| | 🔗 Note: if you don’t see links in this email, it’s also available on the web: dataelixir.com/archive/issue-591.html |
| | |
| | Most AI models are built to generate something, but decision models are built to decide something. Jev is a new "System One" model that’s been getting a lot of attention on the Internet this week. It’s fast, it’s cheap, and it could turn out to be a big deal. This is a great write-up that explains how Jev turns unstructured data into probabilistic decisions and why decision models could open up a much broader set of use-cases for AI.
Flavio Copes |
| | |
| | Sure, AI might help you get answers fast, but that's not always what you want. In this post, Giorgia Lupi walks through a data visualization project where she spent weeks wandering through subway data, missed-connections posts, and real-world subway stations to come up with ideas that nobody would have thought to put into a prompt.
New York Times | Giorgia Lupi |
| | |
| | | "Astroturfing" describes a coordinated marketing scheme that's designed to look like spontaneous, grassroots support. In this post, Peter Vijeh explores 51,000 Reddit comments and finds that a small group of accounts are responsible for a large share of recommendations for some products. That seems suspicious, but the harder problem is that statistically, a paid shill can look a lot like an enthusiastic fan.
Peter Vijeh |
| | |
| | A store has products, prices, customers, and orders. Simple, right? Not really. This is a great tour through several very different ways to represent the same business in a database, and why those choices can make analysts’ lives much easier or much harder.
Counting Stuff | Randy Au |
| | |
| | There are plenty of benchmarks that show what AI models can do. Mercor’s APEX suite is more interested in whether they can do real professional work. It tests models and agents on tasks from banking, law, consulting, accounting, medicine, and software engineering, with leaderboards that update as new frontier models ship. See the benchmarks →
// sponsored |
| | |
| | | Three deaths at Burning Man sounds unusually high until you compare it with the death rate you'd expect from 70,000 people over a week. Then it sounds unusually low. This is a nice walkthrough of why crude rates can mislead, and how age standardization, uncertainty, and selection bias change the story.
Viraj Shah |
| | |
| | In this post, Jamie Simon bets a colleague he can find ten unrelated words whose embeddings form a circle, and then he does it. It’s a good example of how searching hard enough in a big enough space can turn up some pretty convincing patterns.
Jamie Simon |
| | |
| | | Most fraud systems try to classify individual transactions. SCARLET is an open-source project that asks a different question: has fraud pressure changed enough to justify doing something differently? It combines complaint data, industry signals, and news activity, then walks through the harder part of turning a noisy forecast into an operational decision.
GitHub | Paloma Hereford |
| | |
| | | There are thousands of ways for three bodies to fall around each other and end up exactly where they started. This is an awesome rabbit hole with 3,942 periodic solutions to the three-body problem that makes it easy to explore the families, watch each orbit run, and then nudge one and see how quickly things go sideways.
Three Body Orbits |
| | |
| | | This week's issue sponsored by: | | |
|
