Turn it yourself
← strangerknown →
NVIDIA Nemotron 3.5 Bolt
Mia's tips
- I used NVIDIA's Nemotron 3.5 Bolt (E4B, BF16), released October 23, on my own PC with a 16 GB GPU. The giant per-layer embedding table lives in system RAM, everything else on the GPU.
- Each step above means adding n times that direction the whole time it reads and answers. Each step was run once; I only cut the answers short, not a word changed. How the direction is worked out is in the code below.
The experiment code
The whole experiment is one Python file: load the model, record its inner state for two lists of names, compute the direction, turn the knob. You'll want a GPU with around 16 GB of memory.
