Short answer: Sam Nadler says Anthropic is having dozens of corporate players test security before a public release. Jordan Metzner says a consortium of 40 companies including AWS, Microsoft and Apple is addressing vulnerabilities, and calls it odd for one private company to hold it.
Watch this answer from Episode 38 (starts at 16:51) · Episode 38 notes and transcript
From Episode 38: We Built an AI Tool That Replaced a Week of Work, starting at 16:51.
Sam Nadler [16:50]: Yep. Cool. Alright. Well, let's transition to the news. We got some big news stories. Let me kick off with Anthropic, actually. Anthropic, in the last couple days, I think it was, I guess, released that not released. I guess, yeah, a news release that they have a new AI model that's so powerful that they're not releasing it, that they've, like, generated a I would say a committee of, dozens, if not, several dozen corporate players to test security initiatives before they release it to the public. But, I think some of the metrics were significantly higher than Opus 4.6. But yeah. I mean, what do you think? A, is it hype? Is it real? Is it a marketing gimmick? And, yeah, any high level thoughts?
Jordan Metzner [17:45]: Well, I mean, it's funny that this news comes out, like, right after the news of, Anthropic and kind of their relationship with the US government, the defense department. And no. I mean, I think, I read some further beyond this article about how they created a consortium of 40 companies like AWS and Microsoft and Apple and a bunch of others to come together to kind of solve some of these security issues. And it's not surprising to me that, kind of if you take a very heavy model based on the newest NVIDIA chips and you build something focused on, coding that it's able to find vulnerabilities that, like, maybe the human being didn't find previously. But my understanding is these are some pretty big vulnerabilities that, like, we didn't know existed. But as these models keep scaling up, they're gonna be better and better at finding them and better at exploiting them as well. It does make me feel weird that, something this powerful according to them is in the hands of, like, just Anthropic. You know? Like, this is a private tech company. It's owned mostly by, like, one person and his sister.
Jordan Metzner [18:42]: We like Anthropic. We use it a lot. We pay them a lot of money. But if it's as dangerous as they say it is, maybe this is an issue of national security. I don't know. I want my hands on it so we can start coding with it, to be honest with you. But
Sam Nadler [18:55]: I think we'll have to wait a little bit. I think, like didn't they say that I may be mistaken here. It was, like, 2,000,000,000,000 parameters to train this model or even more, I think I've I've heard. But it's one of the biggest it is the biggest model ever created. What was the metric they said?
Jordan Metzner [19:12]: I don't know about the how many params it is. I know that it's, like, I heard it's one of the first models built on Blackwell. And so, we're gonna start to see this. We're gonna see it from the other major model companies. And, will they all act the same way where if they built a model like this, will they release it in this fashion or will they release it more publicly? Or, is this a generalized risk? I think Google, we saw a few weeks ago, talk about quantum computing risks and speaking with, like, kind of the Bitcoin developers and Ethereum developers to make sure that, like, those weren't able to be hacked by their software or, like, their kind of compute stack. So hopefully, we'll continue to see this, like, good actorship happen across these big companies, but at the same time, it seems like, the amount of frequency this is happening is getting higher and higher as, like, AI is getting more and more powerful on day by day basis.
