03 Newsfeed
Everything we’ve written
News, updates and notes from the work we’re doing now, and from what we’ve learned doing it. Newest first.
Posts
Data is the key, but you have to cut it first
Text arrived ready for machine learning. Domain data has to be labelled first, usually by the specialists you were hoping to free up.
RAG doesn't cut it
Why agentic search is replacing vector search used on its own, and where RAG still earns its place.
The problem with IBM float
A float format from 1964 that modern CPUs still decode in software, and how we make it fast.
Vibe coding isn't enough
Five failure modes we keep hitting with AI assistants, and the judgment call none of them can make for you.
Half of your GPU is idle
In GPU pipelines the GPU itself is rarely the bottleneck. Here’s what we found when we profiled our own.
How we made SEG-Y slicing fast
Depth slices need one sample from every trace in the cube. SegAir reads them in parallel instead.
News
Announcements, appearances and releases, newest first.