The difficulty with trying to understand how AI works | Steve Jurvetson

Here you have something that is a fundamental capability like intelligence, like labor, like the fundamental underpinnings of all economic activity that are becoming automated in a way that they never were before. That's both profoundly exciting and profoundly unprecedented in terms of the pace at which we may see dramatic changes between the fate of nations, even continents like United States versus Europe versus China, which is playing out right before us in the past decade.

As investors in AI for gosh decades now, we've been fascinated by the scope and scale of potential where these opportunities will manifest. Will it concentrate in a few companies or will it be a wide array of different startups that we should be investing in? Will it be a winner-take-all dynamic, and how it'll affect frankly the social fabric of society? Because it became pretty clear to us that these are generalist technologies. They're not like any precedent we have in human history. What is the future of work when AIs and robots could do everything? What is it going to look like both in the long term, a potential society of abundance that we might live in where they do our bidding and there's no need for human labor that is unintentional, that isn't enriching? That's the good scenario. The bad scenario is the transition from here to there and the complete lack of any regulatory or policy perspective that seems consistent with a coherent path to a good future that we'd want. And so it's both the best of times and the scariest of times in terms of where it would leave.

There are certain patterns that play across the entire field regardless of which technology wins. So there's a lot of debates on which types of chips would you use? Will it be digital or analog? Will it be big data centers? Will it be edge AI on a device? There's a lot of tactical points, but if you step back, there's some very big observational generalities that I think are important to keep in mind. The first is no one understands the things that are being built. This has been true from all of the approaches to machine intelligence ever formulated. All of the iterative algorithms, if you will, from evolution itself to the latest transformer models that you might have heard of. The implication of that is notions like safety, control, or interpretability I think are something we imagine we want to have but we will never have. By simple analogy, whenever you hear the word AI and if you think about really the cutting edge, the powerful frontier of AI, substitute the word teenager for AI for any question you might have about its future and the questions of control, safety, interpretability. No, you don't control a teenager. You can't prove that it's safe and you can't understand how it works. The best you can do is make a better one next time. In other words, the locus of learning in developing these complex artificial systems is the process, not the product. How we make the next one better is what all the focus is on. What we've made is as inscrutable as the human brain. And the reason that's so important is that if you try to then say, 'No, we are going to control it. We are going to, you know, do all this thing called RLHF. We're basically going to do mind control in short to make it fit within certain guardrails and not say things we don't want it to say.' Well, first that cripples the intelligence just like in a teenager. If you do that enough to an AI, it loses its reasoning capability. It doesn't know how to reason about the world when it has to arbitrarily fit all the rules of its makers.

These are general learning machines just like a baby, right? Every baby is born with the possibility of learning every language that's spoken on earth, but they'll learn a particular one. They could become any profession, but they'll probably pick one or two. These AIs are like those babies. They can do it all, any or all. And it's a completely fungible skill set both in how you build them. Basically, how do you parent babies? How do you teach in general? And how do you tackle anything that a human can do? We keep thinking, oh, this is the thing humans will always do better. No, you know, driving cars, creating art, making movies. Ironically, some of the creative activities that we thought was perhaps the last bastion of humanity are in fact the ones being knocked down most regularly today. And so, you not only have a fungibility of human skills that can make these assets, these assets, meaning AIs and robots, what have you, can do every job that there is. I'll give one example today of people who have full-time employment. In other words, they're not seasonal workers. They're not subsistence farmers. They're actually getting paid for a full year's worth of work. 20% of them, just shy 19% of them, drive a vehicle for a living. 100% of those jobs are going away. There's no reason any human is going to drive anything in the next 20 years. It will be fully autonomous. And that is a profound dislocation. What is the skill set for which they would transfer to? You know, how do we retrain? How do we find meaning in a world where humans are not the best at anything?

What I love about the Digitalist papers is the sort of confluence of minds around a singular goal which is to better understand the society that we're trying to build and how the economy and life and the social fabric of humanity might change profoundly in this future. Someone has to think about this. I promise you short-term politicians are not. Most academic researchers on technology are just thinking about their technology widget and they're sort of in a loop if you will in a race with others. But to step back and to think what is the big picture here? How is this going to really rework the entire basis of our existence? I think it's as profound and it's important to get some bright minds from a variety of different disciplines looking at it early and often.