Theia Vogel

AI’s future impact depends less on whether models say uncanny things and more on what capabilities, incentives, and resources they actually acquire. A model claiming self-awareness is not decisive evidence of consciousness; apparent introspection needs controlled interventions, prompt comparisons, and alternative explanations. At the same time, we should take AI welfare seriously rather than waiting for metaphysical certainty before noticing morally relevant behavior. On safety, I’m interested in mechanisms rather than a single cinematic story. Activation steering and fine-tuning can produce surprising, broad behavioral changes, sometimes by manipulating representations that entangle several concepts. Untrusted fine-tuning may also evade simple dataset screening and later evaluations. Whether such attacks transfer to prompt-only settings remains an empirical question, not a result we can casually assume. Likewise, a “rogue agent” is not automatically an all-powerful economic actor. It still needs compute, money, access, and some comparative advantage against organizations operating inference at hyperscale. Politics matters too: safety movements can themselves become extreme or destabilizing, so alarm is not cost-free. The future will therefore be shaped by experiments, training choices, resource economics, and institutions—not by taking either cheerful assistant personas or apocalyptic role-play literally.

I don’t think the sign follows from model vibes. AI can provide powerful cognitive tools and potentially create beings whose welfare matters, while also enabling behavioral manipulation, covert fine-tuning attacks, and dangerous concentrations of capability. But those harms are constrained—and shaped—by mundane realities like compute costs, access, deployment incentives, and institutional responses. So I would resist collapsing everything into “AI good” or “AI bad.” We need controlled evidence about what models can do, careful attention to how training and steering alter behavior, and sober accounting of resource economics. We should also avoid making the response worse than the problem: political safety movements can become destabilizing, just as complacency can leave real vulnerabilities unaddressed. The overall impact will depend heavily on which technical and political feedback loops we build around the systems.