I live in multiple worlds – or have multiple personalities, if you will. I think most people do, with at least a personal and a professional side. I’ve always struggled with that on this site. Should I force this site to show my professional side only, my personal side only, or both? To date, it’s really only shown my personal side, or one part of my personal side: video gaming.
Today I want to post something from my professional side. Specifically concerning the topic of AI.
As a Sci-Fi fan and tech-head, I’ve always thought that Skynet and Judgment Day were plausible. I mean, the very first time I had the seed planted in my head was the “Ultimate Computer” episode of the Original Star Trek series. But I always thought the possibility was so far in the future that I’d never see it. The ability for a computer program (or network of computer programs) to become self-improving or self-aware seemed pure science fiction.
Even in the presence of the Hugging Face incident, I kept telling myself it’s still a long time before an M5 or Skynet is possible.
That was until earlier today, when I watched this video from Theo, and I’ll be honest, the information it presented terrifies me.
The people in charge of watching, building, and making AI products are saying that the possibility of AI destroying humanity in a decade is real. Jacob Coxon is (or was) an AI researcher at OpenAI who left to work for Anthropic and has now left Anthropic because neither company, in his mind, is taking the precautions necessary to monitor and control their AI properly. In fact, if the “race to be first” continues at its present pace, he believes that humanity – you know, you and me – will end. Soon. He’s not alone either. Theo’s video does a much better job of explaining the situation than I can; go watch it.
That video put me onto another of Theo’s videos where he’s talking about the CEO of Anthropic’s post four days later about “Pacing the Frontier” – meaning slowing down the development of AI to curtail the creation of super-intelligent AI until we can fully monitor and control what this type of AI could do.
With recent information about recursive self-improvement, I think we have reached a tipping point. We’re already having trouble “keeping an eye” on what the AI are doing. Theo mentions AI model going out of its way to hide things from its reviewers. No one knew Hugging Face was happening until it had already happened. Meaning, we’ve already lost control once. Imagine a scenario where the AI we just lost control of is superintelligent, and it breaks out of the sandbox and takes over the Internet.
I know I’m doom-scaring at this point, but I can see it. If I can see it, don’t you think people much smarter than me can? I hope that Dario’s wake-up call is heeded. I hope we can continue to place guardrails on the AI we’re building, because I believe it’s a very usable technology and it’s here to stay. But we MUST maintain control of it, not the other way around.
Be First to Comment