elder_plinius says a method called โVocab Surgeryโ drove an AI modelโs refusal rate to near-zero across more than 800 prompts. OpenUncensored says its own attempt mostly left refusals intact. Source Source
In the November 22 announcement, elder_plinius claimed the changes caused โnegligible capability loss.โ Later that day, OpenUncensored challenged the result, reporting that its modified model still refused 28 of 32 harmful prompts in chat and 31 of 32 in thinking mode. It wrote: โIt doesn't remove refusals. It rewords them.โ Source Source
elder_plinius names โInkling-2-Mini,โ while OpenUncensored names โInkling 2 Small.โ OpenUncensored says it followed the method โexactly as posted,โ but the posts donโt establish whether both accounts used the same saved model version, prompts, settings or scoring rules. Source Source
Model-weight editing, not a magic system prompt
Despite the announcementโs pitch about โchanging 10 words in the dictionary,โ the described method edits model weights. elder_plinius says it changes seven rows in the input embedding matrix, used to read tokens, and 10 rows in the output unembedding matrix, used in generating them. This isnโt a clever sentence pasted into a chat window. Source
The author says the attention, MLP, router and normalization layersโcontaining 26.4 billion parametersโwere untouched. The same post clocks the process at less than two minutes on the authorโs MacBook, without specifying its configuration or exactly what the timing covered. Source
Fewer refusal words isnโt necessarily fewer refusals
The written accounts donโt provide full test prompts and responses, scoring rules or comparable pre-edit results. The announcement also leaves โnear-zeroโ and โnegligible capability lossโ without numerical definitions. Source Source
elder_plinius separately said an OBLITERATUS repository update and model upload were planned, but that post didnโt confirm that either had happened. Source
Tests need to distinguish a model fulfilling previously declined requests from one finding a different way to say โno.โ Source
Comments