Dan Mihai Dumitriu
dmd17.bsky.social
Dan Mihai Dumitriu
@dmd17.bsky.social
Computer scientist, father, husband, son, middle aged asshole trying to be a good human.
This is the very definition of thought police. 😀
January 28, 2025 at 3:04 AM
The fact that this model is trained to expose its reasoning makes it easier to censor. The system must have a thread which is watching the "thinking" in real time and looking for keywords. If a banned keyword shows up, the inference is shut down. This is before the model gives its final answer.
January 28, 2025 at 3:04 AM