Despite the focus in the open weights world on private AI + not training on your traces, censorship resistance via abliteration is way easier to enforce from a computational complexity lens
It is hard to imagine a future where routers don't provide both privacy + abliteration
PrismML team fit a 27B model into 5.9 GB.
We uncensored it without changing a single weight. 🐳
Introducing OrcaRouter Ternary Bonsai 2 27B Uncensored.
Traditional abliteration modifies weights. On a ~1.72-bit ternary model, that means re-quantization — potentially destroying
The beginning of the financial end for most crypto people (Zcash being an exception) was when they started mentioning Fed rates in casual convos
Perhaps similar for our debt chugging compute brethren
