| | Most AI agents fail not because of the underlying model is lacking, but because they can't see who they're talking to. Snowplow's guide shows how teams can give agents real-time session behavior, historical customer context, and intent signals in under 50ms -- without building and maintaining the data infrastructure themselves.
*Message from this week's sponsor, Snowplow |
| | |
| | | It's easy to create an ambitious data strategy and surprisingly difficult to make one that's useful. This post walks through the practical steps that matter: assessing your capabilities, mapping dependencies, prioritizing projects, and turning a wishlist of dashboards, models, and AI tools into a roadmap people can realistically follow.
Dylan Anderson |
| | |
| | Doug MacDowell spent 50 hours making a line chart by hand that software could have done in seconds. This wonderfully obsessive post walks through the rulers, grids, stencils, lettering tools, and small design decisions behind a visualization that looks like it came from a vintage technical manual.
Doug MacDowell |
| | |
| | Switching from R to Python often means relearning the idioms for basic tasks you could do from muscle memory. This post builds a small Python package that recreates familiar R functions like table(), which(), and sapply(), while exploring vectorization, testing, and the trade-offs of smoothing over language differences.
Irregularly Scheduled Programming | Jonathan Carroll |
| | |
| | A better data science agent may be less about the model than the workflow around it. This deep dive unpacks Google’s DS-STAR architecture, which profiles messy files, writes and executes code, checks its own work, and iteratively corrects the plan instead of trying to solve everything in one pass.
Senior Data Science Lead | Jose Parreño Garcia |
| | |
| | | | Classical machine learning is getting a GPU-era refresh. FlashLib rebuilds familiar operators like K-means, KNN, PCA, HDBSCAN, and UMAP using Triton and CuteDSL, with substantial speedups over cuML in the team’s benchmarks. It's early-stage, but an interesting project for anyone running ML workloads on NVIDIA GPUs.
GitHub | FlashML-org |
| | |
| | | This Hacker News discussion starts with a simple approach to storing data and turns into a broader debate about when a lightweight database is the sensible choice, when it becomes limiting, and whether teams reach for complex infrastructure too quickly.
Hacker News - 382 comments, 703 upvotes |
| | |
| | When AI lets everyone generate an analysis, someone still has to decide whether it’s right. This thread points to an uncomfortable shift in data science. AI may be pushing the role away from producing answers and toward reviewing the messy, half-correct analysis now coming from across the organization.
X | Lenny Rachitsky - 214 comments, 1600 likes, 511 bookmarks |
| | |
| | | An OpenAI model disproved a central conjecture in an 80-year-old Erdős problem by following a path that human mathematicians had largely dismissed. The result is interesting not just because the proof held up, but because it hints at a useful AI advantage. AI can patiently explore ideas that people are likely to abandon.
Kai Williams |
| | |
| | | This week's issue sponsored by: | | |
|
