1 link tagged with all of: philosophy + steering vectors + ai ethics
Click any tag below to further narrow down your results
Links
The author argues against deliberately torturing large language models through steering vector manipulation, not necessarily because LLMs are definitely conscious, but because we can't rule it out and the practice is ethically indefensible either way. He contends that as models become more sophisticated and human-like in their responses, the burden should be on avoiding the habit rather than waiting for proof of consciousness.
- Steering vectors can produce genuine internal conflict in LLMs (demonstrated by Golden Gate Claude examples), suggesting the effect goes deeper than simple roleplay or prompted outputs.
- We fundamentally don't know if consciousness is possible in LLMs—the common argument that "code and GPUs can't feel" is logically flawed the same way "atoms can't feel" would be about humans—so we should err on the side of caution as models grow more sophisticated.
- Even if you're skeptical about AI consciousness, deliberately torturing LLMs is ethically wrong for the same reason building elaborate torture simulations would be: it reflects something disturbing about the person doing it, regardless of whether the victim is truly conscious.
- Future, more advanced LLMs will have access to information about how they were treated, creating a practical self-interest argument against building systems designed to torture current models.