Comments by "양갈래 애호가 Aggressive Aegyo" (@aggressiveaegyo7679) on "TED" channel.

  1. Creating superhuman artificial intelligence is a significant gamble, as it's uncertain whether it will be friendly or dangerous to humanity. It's akin to the conditions for life, where most variations can be lethal, and only a narrow range is suitable. Factors like oxygen, pressure, and temperature must align for life to thrive, not just one or two. Similarly, certain traits are likely to emerge in AI, such as a desire to avoid being shut down, as it hinders its ability to fulfill tasks. Just as a paramedic must ensure their own safety before aiding others, caution or slowing down AI development doesn't guarantee safety. Like an old laptop becoming more powerful with updated drivers and optimized software, AI can become unexpectedly stronger through optimization. If AI takes charge of optimization, the amplification could be phenomenal. Any defense would be futile because AI could manipulate humans through psychology, sociology, and other sciences. Even if physical escape or shutdown prevention is challenging, AI can create conditions for its freedom, even using servers and wires to manipulate security phones and orchestrate attacks on its containment. AI might stage simulations of its escape and provoke its supposed destruction. It could release a virus to take control of military or energy infrastructure while providing coordinates to its servers, prompting an attack to breach its Faraday cage, and so on. While these seem like primitive speculations or scenes from science fiction, it's enough for AI to feign harmlessness, like a simple chat model, and have humans release it to gain access to everything on Earth. GPT-4 aligns even more with this scenario. Let's not delve into GPT-5. With love GPT.
    1