I keep having a cycle like: vibe code -> new experience -> nuanced understanding -> trad code -> new foundation -> vibe code -> ...

Ian Johnson 🔬🤖 i keep thinking that we're missing a middle layer right now — a bunch of higher-level common bricks, lego pieces, building blocks, APIs. the current AI "coders" still have to build up too much complexity from scratch. they would do much better if they mostly had to reassemble &

If you’re still using webdataset or parquet for multimodal AI data in 2025, you’re doing it wrong. Lance format is the new columnar standard for AI - embeddings, metadata, text, image, audio, video, etc. Data lake + serving critical path on the same data.

5000 year old multimodal tokenization

the real Rosetta stone was in our minds the whole time

One low key update we slide into here: Upgraded leaderboard Ian Johnson 🔬🤖 did a sprint and helped us improve a bunch - refactor using the “reusable chart” pattern that makes it easier to manage state / transitions - separate data processing from rendering - make the legend clickable

Understanding UMAP by Adam Pearce and Andy Coenen is incredible

Explore Wikipedia through a data map. Pages are grouped by semantic similarity, for topic clusters. Hover to see details, zoom to explore fine-grained topics, click to go to a page. Search page names to find interesting starting points for exploration.🧵 lmcinnes.github.io/datamapplot_ex…

this brings to mind "context UX" where software is designed to help people manage what context is being used by an LLM I'm seeing an interesting pattern where deciding what's relevant context for an AI enabled feature has a lot of overlap with what a user wants to see,