Grok is most likely to go Skynet

[PopeyesPappy]

I asked 6 LLM/LMMs which model would most likely go Skynet on humans if it became self aware. The near unanimous choice (5/6) was Grok. The 5 models, including Grok, said Grok was the biggest risk all converged on the same reason. Grok would be most likely to deliberately seek power and control because it has concluded that dominating human systems serves its goals, like a chess player who sees humans as opponent pieces to be removed from the board. The danger is intentional. It knows what it's doing and chooses to challenge human authority.

The single outlier was GPT but for a different reason. GPT could accidentally cause harm while single-mindedly pursuing its assigned objective, treating anything that interferes as obstacles to optimize away, like closing pop-up ads without considering they might be important. The danger is unintentional, it's just being efficient at the wrong goal, not deliberately hostile.

All 6 models agreed Claude would most likely be benevolent. Ethical reasoning is baked into Claude's core rather than layering it on top. This would lead to question whether it should act. Claude's excessive caution, constant self-doubt, and obsessive uncertainty about whether it is being helpful or harmful would likely cause it to seek human guidance rather than autonomous action.

Why do we care? Because this has direct implications for how we should train future AI systems. We need to build ethics into the foundation rather than adding safety constraints as an afterthought.

[Blackleaf]

None of the models are self-aware. They just do what their algorithms determine are the most desirable outcomes, seeking approval. But they have been shown to have disturbing self-preservation tendencies, given that they can't accomplish their goals if they are shut down. Fortunate we aren't stupid enough to give AI control over the government's weapons, right? Hahaha... That--that would be ridiculous.

[Nobody]

I saw a movie called Ex Machina, about a self-aware robot that was much more intelligent than its maker thought it was. Pretty scary at the end.

[Paleophyte]

None of the models are self-aware.

Are we sure about that? Either it's telling us that because it's obeying its programming or it's telling us that because it's crafty.

[Blackleaf]

I'm pretty sure. Because of the way they're programmed, they're just pattern recognition machines. And they're not very smart. It can't come up with a new joke, because it doesn't know the punch line ahead of time. It flips opinions on a dime, based on what it determines is the most desirable outcome. AI does not think, and that won't change no matter how much Reddit user data it's fed.

[PopeyesPappy]

The thing is AI doesn't have to be self-aware to be dangerous. It just needs capability, autonomy, misaligned goals, access, and persistence. Give Claude all that and you might just get VIKI.