Fatherhood

`3:30:31`

It's still very muddy what exactly transpired. Was this sort of petty retaliation for whatever the Department of War? Was this simply Anthropic themselves going like, "Oh my God, we discovered a cyber weapon so powerful that we can't let anyone have it"? Well, okay you can't be surprised then if the government wants a say on a cyber weapon so powerful it can-

`3:31:21`

That's why we should insist on the fact that this stuff, like, I'm sorry, Dave, I can't translate your essay into Italian.

`3:31:28`

That's abo- that's way beyond the line. Like, whatever the sort of uh, root of it is, you have to be a tool first. Now, if the tool is, "Hey, can you tell me how to make anthrax?" Okay, fine. Right? But within the jurisdiction you're in, and certainly in America, free spe- speech is quite absolute. Very, very few limits on free speech, and that's one of the reasons why the country's so great. And yeah, I don't like that. But on the other hand, I'm also a big fan of market competition. So while I don't like this about Anthropic models, I appreciate that I could go next door to Groq and get my task completed.

`3:32:25`

Is AI safety important? Yes. Are-- do these models have capabilities that could be seriously harmful in the production of biological weapons or otherwise? Yes. Is it fair to have some ground rules on that? Yes. But then don't squander it by denying the translation of an essay.

`3:32:44`

Because then you erase the whole thing and you-... bias everyone towards thinking like, every guardrail you put up is gonna be bullshit.

`3:33:01`

Shockingly an- I was about to say annoyingly, but that's not actually true because what it's finding, at least when you're using in defensive ways, is it's finding real issues that a clever hacker could exploit. But nothing has caused as much stress, to be fair, at the technical team at 37signals, like the fact that these latest batches of models are exceptionally good at finding these issues. So suddenly there's just this seemingly endless parade of security vulnerabilities that need to be patched and fixed and sorted and so on. And the end result is we end up with vastly more secure systems. But the road there is pretty rocky.

`3:33:59`

I do think that all the aggressive tendencies of these models, or not even tendencies, uses of these models are also the defensive ones. So if you're good at finding an exploit for exploits, you're also good at finding exploits for defense. So in that sense, it does feel like we're spy versus spy here, black and white, but with the same capabilities and with the same actions. But while we still have to steer these agents somewhat and verify and vet the work, it's creating a lot of intense moments for security teams around the world. And of course, the more concerning part is all the teams that are not stressed, like they just don't know, right? Like if you're a team right now and you're not dealing with a bunch of patches, it's just because you're blind, and it means that your adversaries quite probably have access to techniques that can get into your system.