| | Nebius Token Factory is designed for teams going beyond APIs. Deploy and scale open-source LLMs, fine-tune them to your use case, and manage access, data, and cost as your product grows. Everything you need to build serious AI, without stitching together tools. From LLM to production system, in one platform →
*Message from this week's sponsor, Nebius Token Factory |
| | |
| | 🔗 Note: if you don’t see links in this email, it’s also available on the web: dataelixir.com/archive/issue-566.html |
| | |
| | AutoML promised to handle modeling for you, but modeling wasn’t the hard part. This post explores why commercial AutoML stalled, why open source still matters, and where the real bottlenecks are now.
Thomas W. Dinsmore |
| | |
| | Most models fail quietly. This one failed all at once. In this post, Harrison Katz shares how COVID collapsed Airbnb’s forecasting assumptions and how they rebuilt for a world where distribution shifts aren’t edge cases.
Airbnb Tech Blog | Harrison Katz |
| | |
| | Bayesian stats often sounds harder than it is. This post walks through the shift in thinking, from fixed parameters to uncertainty-first modeling, with examples that make priors and posteriors feel practical instead of philosophical.
Nicolas Chagnet |
| | |
| | Tree models are everywhere, but what changes when you make them Bayesian? This tutorial uses baseball pitch data to build a pitch quality model with BART, showing how priors can replace tuning and give you uncertainty “for free.”
PyMC Labs | Chris Fonnesbeck |
| | |
| | | | Here's a good example of where data work is heading. Long-term personal data, OCR failures, and agent-assisted cleanup lead to visualizations that feel earned. The real story is the pipeline behind them.
John Rush |
| | |
| | Most plotting tools assume your data is done but streaming data never is. Vistral brings a Grammar of Graphics approach to real-time data, handling shifting windows, expiring points, and continuous updates without rebuilding charts every few seconds.
Timeplus | Gang Tao |
| | |
| | | This is a clever, minimal take on data sharing that sidesteps the usual friction of files, storage, and permissions. Ziptable encodes small datasets directly into the URL, keeping everything client-side, which is both elegant and surprisingly practical for demos, examples, and quick collaboration.
Evan Peck |
| | |
| | Configuration always starts simple, then turns into a pile of YAML. confingy takes a different route by tracking constructor arguments directly in your code, so your experiments stay reproducible without a separate config layer.
Ethan Rosenthal |
| | |
| | | More and more, search is being mediated by AI summaries. This post shows a more deliberate approach to finding information online, using a “search like a librarian” mindset, with concrete tactics and alternative engines that surface things Google won’t.
Card Catalog | Hana Lee Goldin, MLIS |
| | |
| | | This week's issue sponsored by: | | |
|
