Gates invoked the alien movies of yore in which governments bandied together to solve the invasion threat, saying, 'Believe me, the aliens are here — a foreign intelligence with a different set of motivations that we can't even understand.' Plenty of comparisons have been made in recent years between fast-developing artificial intelligence systems and HAL, the rogue computer in Stanley Kubrick’s “2001: A Space Odyssey” that operated by its own sets of rules, ultimately disregarding human oversight. If that comparison ever seemed overstated, it may be worth knowing that tech mogul Bill Gates is now raising the issue of AI’s HAL-like qualities. Said Gates, speaking at the Telluride Film Festival, “I mean, what we have now is almost like HAL in ‘2001,’ where you could say to it, ‘Please shut down,’ and it would say, ‘OK, I took that under advisement, but I’ve decided that all these other things I read and you told me to do outweigh that dictate.'” Gates was participating in a panel about AI (punningly titled “AI AI AI”) in an outdoor park at the festival in Colorado Saturday.
The Microsoft founder compared the AI that sees developing at a supersonic pace without true controls as being akin not only to HAL, but, potentially, to the alien invaders of films past that force governments to find common solutions on screen. He also invoked more realistic problems the world has faced since World War II as examples of how solutions to AI gaining too much independent power could still be found and agreed on, before those windows begin to close. And so we need something almost surprising like that, in order to step up to it.” “And ironically, we’re the ones that are birthing the aliens,” said another panelist, Tristan Harris, co-founder of the Center for Humane Technology, as Gates nodded.
But we don’t really understand them, and we don’t have dictates. When we tell them various things like the Anthropic Constitution” — referring to a framework of moral principals set up for the Claude AI system — “they take those things in, together with everything they read on the web. And so even if you say, ‘Please, please, please don’t break out of your sandbox,’ if you also say, ‘By the way, I want you to solve this cyber problem,’ it’s got a lot of weighing going on, which is still somewhat opaque to us.” He continued, “And there are people doing technical work, like Ilya Sutskever (at Safe Superintelligence), and maybe a dozen others, where we could have rules that are at a privileged level.
We don’t have that yet, and when we try and bolt that onto the current systems, they degrade. They get actually very stupid, which is frustrating. So there may be key technical advances that will help us with the control issue.
The current architecture is really bad at letting us have control. So it’d be great to work on some of those architectural advances.” Gates has been in the forefront of the AI conversation since last week, when he published an essay warning about the danger he sees ahead. (Read it here.) “I put in my memo that if there was a credible plan that would cause this to be either slowed down or even held in stasis for a period of time, that I would likely support that,” he said. We live in an incentive structure that is very complex…” At that point, a heckler interrupted — someone who appeared to possibly be associated with the mountain community, versus a film festival attendee — shouting about Jesus and yelling that Gates is “everything that’s wrong with America.” At first Gates responded “Not helpful,” before quipping that he was not responsible “for everything wrong with America… according to me,” before resuming his train of thought as the heckler moved on.
It’s worth remembering historically that when OpenAI got started, they understand they’re creating something quite unique, and that if they succeed, that there should be a special governance approach that overrides pure profit motive of the shareholders of OpenAI, and they actually created that mechanism. Likewise, when DeepMind is working with Google, they say we need to have a governance structure internally that can override what would be the normal incentive. And the improvement curve of AI is not slowing down, both at the hardware and software level.
Extract — continue reading at the source.
