00:00I have a very contrarian view on this. So I know that a lot of people in the industry,
00:06especially the ones at the big labs, are trying to tell you that we need to start pacing.
00:10What they mean by that is not slow down the training, but slow down the release,
00:15because if we keep the releases the way they are, suddenly things are going to get really scary for
00:20everybody. Look, I'm originally from healthcare, right? So I know what regulatory capture looks
00:25like. And this looks and smells like regulatory capture. The big AI labs are worried about a
00:32very specific thing, and it's not recursive self-improvement, which is what they're talking
00:36about. They're worried about the rise of open source. If you look at what's happening,
00:42and people can go to open routers data, which is this one platform that many people can use,
00:49and you can use many different models, tokens for the big labs are dropping faster than anyone
00:57expected. And open source is rising faster than anyone expected. That explains why NVIDIA acquired
01:02Hugging Face for a record deal. And I think that this is what they're really worried about. They're
01:09trying to slow down the pace of open source and adoption of open source, not really thinking about
01:14safety. And I have to say one more thing, which is, if you were listening to Dario, who obviously I
01:19respect, but if you were listening to Dario, we would have guardrails at GPT too.
01:24So if that's the case, are we just at a moment where people are deliberately misunderstanding what
01:31these AI lab founders are saying and kind of jumping on the moment, capitalizing on the moment to say,
01:38hey, let's, let's, let's take a much closer look at this and slow things down dramatically,
01:43if not halt progress completely.
01:45You see, there is a lot of, so I'm an AI researcher, as you all know from background. And so
01:51there is a lot of vested corporate interest in making sure that these labs continue to thrive.
01:57And what they're learning right now is that when, when you keep adding more and more training data from
02:02the entire internet, yes, it's getting better. The scaling laws are continuing to prove out,
02:06but they're not proving out the way that they thought they would. And more verticalized,
02:11hyper-specific models are doing better, you know, small language models, but also hyper-specific,
02:16you know, specific harnesses, which is sort of learned on local data, learned on sector-specific
02:23data are starting to outperform them. And so I think what's really happening is you've got all of
02:28these different forces that are making these giant AI labs worried. And I believe that's what's
02:33driving this. And I'm not alone. There's a lot of people saying this out loud. Just to kind of set
02:38the stage, you hear letters RSI a lot. And RSI means recursive self-improvement. We've had that
02:44for 20 plus years, but it doesn't work that well. There's still a lot of technical challenges. And
02:49unless the big AI labs are going to suddenly start showing the data, I mean, we just saw a deep
02:54RSI from
02:55Google and they, in their own paper this last week, said that there's still significant challenges to
03:01generalizability. So these ideas sound great, and they allow you to believe that AGI and ASI is going
03:07to take over. But the truth is that these labs are starting to see real pressure from the open
03:13source models. Give us a sense, Danish, the relative competitive landscape, the US versus
03:20China, if that's even the right way to think about it. It is the right way to do it. Yeah.
03:23I mean,
03:24China has gone all in on open source. The quality of their open source models is tremendous. I think
03:29we have to be very careful about slowing things down. Because again, look, there are many things that
03:35I think the administration is doing right and many things that they're doing wrong. They are right
03:40about this. And I think Jensen Huang at NVIDIA is right about this. This is not the moment for
03:44pacing. This is the moment for acceleration. We really need to be careful about allowing the
03:50corporate interests of a few large labs driving the decision making for the entire country.
03:57What do you hope will happen when Xi Jinping and President Trump meet in Washington, in New York,
04:03in the US later on this month? Because AI is supposed to come up on the agenda here.
04:08I think we know China's stance on this. I mean, they came out and they said that fear
04:13mongering and foreign influence is not going to change their strategy. What I would hope is that
04:17we can come up with international standards on use in defense, use in biology, use in gain of function.
04:26I mean, these are the things as a physician I care deeply about, as an AI researcher I care deeply
04:30about. I don't think, I mean, again, contrarian viewpoint, I don't think these models are actually
04:35ready to handle consequential things like that. Because the harnesses, again, think of those as
04:40guardrails that you're putting on your own models, are not mature enough to be able to handle such
04:46sensitive, consequential information. I mean, we're seeing this in healthcare right now, right?
04:50And so I think that's what I'm hoping for is actually more guardrails on what AI can touch
04:56versus guardrails on the app.
Comments