3:50.627
ES: So there's a real limit in energy. Give you an example. There's one calculation, and I testified on this this week in Congress, that we need another 90 gigawatts of power in America. My answer, by the way, is, think Canada, right? Nice people, full of hydroelectric power. But that's apparently not the political mood right now. Sorry. So 90 gigawatts is 90 nuclear power plants in America. Not happening. We're building zero, right? How are we going to get all that power? This is a major, major national issue. You can use the Arab world, which is busy building five to 10 gigawatts of data centers. India is considering a 10-gigawatt data center. To understand how big gigawatts are, is think cities per data center. That's how much power these things need. And the people look at it and they say, “Well, there’s lots of algorithmic improvements, and you will need less power." There's an old rule, I'm old enough to remember, right? Grove giveth, Gates taketh away. OK, the hardware just gets faster and faster. The physicists are amazing. Just incredible what they've been able to do. And us software people, we just use it and use it and use it. And when you look at planning, at least in today's algorithms, it's back and forth and try this and that and just watch it yourself. There are estimates, and you know this from Andreessen Horowitz reports, it's been well studied, that there's an increase in at least a factor of 100, maybe a factor of 1,000, in computation required just to do the kind of planning. The technology goes from essentially deep learning to reinforcement learning to something called test-time compute, where not only are you doing planning, but you're also learning while you're doing planning. That is the, if you will, the zenith or what have you, of computation needs. That's problem number one, electricity and hardware.
ES：真正的限制在于能源。 让我举个例子，有这么一个测算， 这周在国会听证会上， 我也引用了这项测算作为举证。 这个测算，计算出美国还 需要 90 吉瓦的电力。 顺便插一句，我的解决建议是考虑加拿大， 那里人好，水电资源也丰富。 不过可惜，按照现在的政治氛围， 比较难实现。 90 吉瓦，相当于在美国 建立 90 座核电站。 没戏。现在我们一座在造的都没有，对吧？ 那我们该如何获得这些能源呢？ 这是一个很大、很大的国家难题。 也许我们可以利用阿拉伯国家， 他们正在紧急建造 5 到10 吉瓦的数据中心。 印度也正在考虑建造 一座10 吉瓦的数据中心。 如果好奇吉瓦到底是多大， 可以把一个数据中心想象成一座城市， 数据中心需要的能量，相当于一座城市。 人们可能会觉得：“好吧。如果算法精进了， 需要的能量应该就能少了。” 有一条不成文的铁律， 尽管年纪大了我仍然记得，嘿嘿..... “格罗夫赐予，盖茨夺走。” （指：硬件给予的性能提升，总被软件消耗。） 感谢物理学家们， 他们让硬件运转变得越来越快， 他们的发明，让人佩服。 而我们这些搞软件的人， 就是不停地用、一直用、可劲地用。 在现在的算法中， 当你在观摩AI 制定计划时， 你能看到它, 来来回回， 不停地演算、尝试各种方案。 对于这些推演，有一个测算， 譬如安德森 ·霍洛维茨的报告， 他们就仔细地研究过这个问题， 他们认为虽然只是做这种基本规划问题， 但计算量至少增加了 100 倍，甚至是1000 倍。 现在的技术拓展， 已经从“深度学习”延伸到了“强化学习”， 然后又到了，我们所说的“测试时计算”， 所以系统不仅仅是在做规划， 它在做规划的同时，也在持续的学习。 这个层级就是我们经常说的， 计算需求的终极状态。 因此，第一个难题是： 缺乏足够的电力和硬件。