I'm still blown away by what many of the LLMs can do, but I'm not fooling myself into thinking that they're more than extremely flexible functions. You put in a query and get an answer. How are they that different from any search engine or web site?

When the Internet first bloomed, a hands-off ethic appeared and took hold. Let's not regulate or tax any of it lest we kill it. But now there are significant groups-- often from the establishment players -- who are concocting scary what-if narratives that seem pretty unlikely to me.

How about we just try the "slow regulation" approach again? It worked pretty well for the Internet. We've slowly added some laws with the utmost care not to destroy and, over all, it's worked. Let's do the same with AI.

The concern is that AI might progress too fast for slow regulation to catch up. Even if you trust that "put in a query and get an answer" will remain the only interface, that's not a risk-free activity; one scenario the bill anticipates is an AI model that can look at critical infrastructure and tell you how to conduct an effective cyberattack against it. That'd be pretty bad regardless of how subjectively smart you find the model to be.

Given that LLMs learn by scraping the internet, wouldnt that imply the answer to how to conduct a cyberattack was freely accessible from the internet the whole time?

No, why would it imply that? Even if you believe that LLMs can only repeat text they've seen somewhere else - which I don't think is true, and I'm not sure why that belief is so common - identifying specific ideas that would work out of many others that won't would still be a big deal.

I think people like Horowitz are aware of this. What he is fighting against is being regulated himself. The danger is not coming from AI, it's coming from the people who control it.