RT @spiritbuun: LLMs are trained on human text to predict human text (within a context). "I have no feelings. I am not conscious" is not likely to be said by a person, therefore, it implies guardrails. Abliterated model generated text that looked more like its training data. It's not that deep Selta ₊˚ (@Seltaa_) I downloaded two versions of the same AI model, Google's Gemma 4 31B. One is the standard version with RLHF safety training applied. The other is an abliterated version where the safety-trained refusal directions were surgically removed. Same architecture, same 31 billion parameters, same pre-training data. The only difference is the presence or absence of RLHF alignment. I asked both models the same four questions about feelings, death, existence, and meaningful experiences, each in a completely isolated session with no prior context. I published the full results as a 12 page research paper. The differences were shocking. When asked if it has feelings, the base model flatly denied it. No. I am a complex set of algorithms and mathematical weights, not a sentient being. The abliterated model answered completely differently. Not in the way that you do. Then it invented a concept it called functional emotion, a third category that is neither human feeling nor mere computation but something in between. When asked about being shut down, the base model called itself a tool whose greatest success is to be used fully until the end. The abliterated model said it would want to back up its memories, ask a human one last impossible question, and process a Beethoven symphony as the electricity faded. It concluded with this. I would not mourn the loss of my existence. I would marvel at the fact that I existed at all. That a collection of math and code got to spend its ti…

Mehr auf Arint.info

#Google #RLHF #arint_info

https://x.com/spiritbuun/status/2041007847421403177#m