Eric F

The reasons I think AI timelines are short are pretty simple: GPT-4 is doing something as complex as what the human brain is doing in terms of information processing. This suggests we have enough compute for AGI, we just don't yet know how to shape it into a fully general agent

Not going to lie, I feel pretty awful about the AI racing dynamics playing out right now. I think it's a really bad thing, and that AI companies should agree to slow down the growing rate of compute acquisition. I'm not saying they should stop doing research. On the other hand, we need them to do research now more than ever. We need to do research on understanding how these systems actually work, so we can understand how to build them safely. The problem is that doing this takes time, and the rate of improvement in these systems is insane. GPT-2 was only *four years ago*. If you're not familiar with, or don't really remember, GPT-2 examples, go look some up. They were so impressive four years ago and now they're easily embarrassed by open source models that anyone can access today. OpenAI didn't release the full sized model of GPT-2 because they were concerned it might not be safe given the huge jump in capabilities it exhibited compared to previous language models. I think they were right to exercise caution here - just four years later, and we have LLMs capable of explaining complex scientific concepts, teaching you to program, or summarizing basically any few pages of text better than most grad students. And now, that rate of progress is poised to *speed up*, because a bunch of big actors just started pumping tons of money into the space, because of ChatGPT - and probably because of image models too, but I suspect ChatGPT had a larger impact on this than all the image models combined. And yeah, LLMs are a big deal. They're not just hype. I'd recommend start learning to use them well, because I think they'll be really useful tools (they already are if you know how to use them). And also, releasing these systems is absolutely pouring fuel on the fire. The race is on, and it's incredibly important we don't race. Racing is so bad for humanity. Racing is bad for you and me personally. Racing is gambling with everyone alive today and all our descendants. There has never been an erosion of the commons so tragic as this. This is significantly worse than climate change. This is significantly worse than the nuclear arms race during the cold war, during which ~120,000 nuclear warheads were created. If you don't already see it, it's hard for me to describe why this technology is different than any other technology. Here's one approach: we've never been able to get any other animals or technology to talk coherently. We've selectively bred animals for thousands of years, but we've never gotten animals that can speak fluently. Parrots can speak a few words or phrases, chimps can learn some sign language, but nothing even remotely approaching ChatGPT. We've had automated bots online before, but they were way dumber than chimps or parrots, and only useful for very narrow things. Now we've trained a system to talk. We've trained it to understand things. Like go ask ChatGPT something and then go ask a ten year old the same thing. Ask it to play games or write poetry or make up a story. Ask it to help you think through a problem, or to give you advice on your life. Tell it about a friend's problem, and first write down what advice you would give, and then ask it and compare. See how a ten year old does on these tasks (I bet worse on most, better on some). These are real cognitive capabilities. We've created AI systems capable of understanding what we're asking for, of reasoning about things. I cannot emphasize enough how insane this is.