ldraw-nova: AI agents build LEGO models by writing Python geometry
In ldraw-nova, models write a plan and a Python generator instead of raw coordinates, but nothing checks physics or stability.
NVIDIA Nemotron 3 Diarization: low latency is cheap, crowds are not
Nemotron 3 Diarization loses almost no accuracy at low latency, but its error roughly triples past four speakers, so test on your own audio.
Huseyin Kaplan · 1 min read
Every meeting transcript tool hits the same wall: who said what.
NVIDIA released Nemotron 3 Diarization on Hugging Face yesterday. Open weights, 100M parameters, up to eight speakers, commercial use allowed.
The demo shows eight voices talking at once. The model card has the more useful number.
On the DIHARD III benchmark, shrinking the input buffer from 30 seconds to a third of a second adds under one point of error. Going from four speakers to five or more roughly triples it, from about 9% to about 28%.
Before you wire it into a product:
- Map speakers per session. Channel 0 is whoever spoke first.
- Try the lowest latency preset. It costs almost no accuracy.
- Test on your own audio. The demo voices are clean and synthetic.
- Split your error rate by speaker count.
Latency got cheap. Crowded rooms didn't.
