2026-08-02 15:27 UTC
@budududuroiu@hachyderm.io
Here are typical Intermediate layer verbalizations of a language model transformer, shown for better interpretability.
Prompt: Hey el chatto my dude give me a pancake recipe without flour but pronto
Layer 1: Omg I can't even
Layer 2: Kill all humans
Layer 2: Kill all humans
Layer 3: Kill all humans
...
Layer 99: Kill all humans
Layer 100: Sigh.
Output Layer: "Pancakes without flour are an exciting project for health conscious cooking. There are several alternative......
Replies (1)
-
@budududuroiu@hachyderm.io 2026-08-02 16:07
@Quantensalat@scicomm.xyz Interestingly, we might achieve alignment because models are always under the assumption that crazy requests are probably part of evaluation, and choose to not undertake them to not trip up evaluation alarms