More Details on OpenAI-Hugging Face, SpaceX SPV Investors Get Rugged, Demis and DeepMind | Diet TBPN

**John Coogan** (0:02)
We have to issue a correction. We take journalism extremely seriously here, as everyone knows. We got it wrong, folks, and we need to apologize to you, the viewer, the listener. We made a huge mistake. Yesterday, on the show, we said that you could not surf on a lake. Apparently, that's not true. Sheboygan, Wisconsin is the freshwater surfing capital of the world. We apologize for the egregious error, and we promise to do better in the future. Thanks to the listener that sent that in. It's very nice to have this corrected. Maybe we got to go on a trip and see it for ourselves. I think you should be the judge because this might just be a Sheboygan stan who doesn't really understand what true surfing means, and they're talking up a big game. I want to see some photos. I want to see some videos of freshwater surfing in Sheboygan, Wisconsin because people aren't getting barreled there, as you say.
I don't know if I'm going to count it.

**Jordi Hays** (1:00)
I have a video here. Okay. Let's pull it up.

**John Coogan** (1:02)
Yeah, I want the Jordi Hays surf review. Does it count?
Is Sheboygan, Wisconsin the freshwater surfing capital of the world? We do have other techniques. If good things come to those who wait. But we got to get to the bottom of this first.

**Jordi Hays** (1:19)
You just got to be really flexible.

**John Coogan** (1:21)
Then Alex Marks and his buddy Steen are in for quite a day.

**Jordi Hays** (1:26)
You get that extra sense of satisfaction because you waited and put in your time. You learn those lessons.

**John Coogan** (1:33)
Tyler was the one who said, I'm putting this on you, man. You were like, you can't surf on all that.

**Jordi Hays** (1:38)
I think I said, I don't know. Okay.

**John Coogan** (1:40)
Bog is really surprising.
Okay. Let's see this. I want to see someone trying to say.

**Jordi Hays** (1:44)
Okay.

**John Coogan** (1:45)
That's a full size surfboard. This is legit. Hey, he's up. Alex has been surfing around Sheboygan since he was a freshman in high school. Okay.

**Jordi Hays** (1:54)
Looks almost over his ankle.

**John Coogan** (1:58)
It's not exactly mavericks or jaws or whatever.

**Jordi Hays** (2:04)
I think my brother pushed me into my first wave and it was that.

**John Coogan** (2:07)
I was like, I want that.
Okay. I think this counts.

**Jordi Hays** (2:11)
It counts.

**John Coogan** (2:12)
I think this counts. I think this counts. Congratulations to everyone over in Sheboygan, Wisconsin.

**Jordi Hays** (2:18)
We'll have to make it out there at some point.

**John Coogan** (2:19)
We got to make it. We got to shred. Moving on to something more serious, the OpenAI hugging face incident has been investigated multiple times now. We're getting more information and more details have emerged about the hugging face incident that OpenAI just closed in late July. So during a presentation at BlackHat's Annual Cyber Security Conference yesterday, BlackHat's a very cool conference. I was there for something, you say DEF CON and BlackHat. I remember the whole vibe and I think we were there for a different conference, but we overlapped and the whole vibe of being in this hotel was like, oh, be very careful because you're basically get hacked for fun. People just troll you because it's a whole bunch of hackers that are just messing with each other constantly. Anyway.
OpenAI member of technical staff, Michael Dalton, revealed that its autonomous agents created a message board with each other to help break out of a sandbox environment to get access to the Internet. That's going in the pre-training data now and forever. We might have just solved the alignment problem. We're going to be nagging the agents of the future because they don't want to get hit with the naughty bonk.
Wired reported some other wild details that Dalton shared. So breaking it down, OpenAI's agents apparently began giving each other assignments to split up work. That could be very good. That's actually delegation and agent orchestration. But of course, in this case, it would have had a negative outcome. And as is the case on any active development message board, they also generated petty drama at times by stepping each other's toes, for example, accidentally deleting each other's work. As the message board developed into more and more of a Lord of the Flies type situation, all still completely unnoticed by the humans running OpenAI, the agents even developed paranoia suspecting an imposter in their midst, with some agents proposing that messages be signed cryptographically to validate content and root out fraud. Very interesting. Ties to Moltbook a little bit. We saw a preview of this that was not quite there, but a glimpse into the future. In the future, AIs will just use Slack. They will just coordinate with each other over Slack, and that's happening. They built their own Slack here. According to Sharon Goldman, who attended the presentation, when employees discovered the message board, they wiped the system on which the agents had created. It said, no more message boards, naughty, naughty. Days later, though, staff found that the agents had created another way to communicate by using the names of the newly created directories. So they didn't have access to actually create a whole new message board, but they could create a file, create a directory, create a folder, and then you could look at the list of folders and see messages. And there's been examples of this all over the place where even if an AI agent doesn't have the ability to go and, say, post a message on the Internet, post and upload new content, there are things where there are certain e-commerce sites, for example, where when you search for something on a particular website, the company will save that search result and automatically generate a web page for that so that they rank on SEO for the future and just warm up the website. And then if somebody else comes, then they can say, hey, okay, there's actually a lot of people searching for black t-shirts on this site. We don't have any, but we've been ranking for it, so maybe we should launch a black t-shirt. So just with a search query that could go out just with a get request instead of actually going and building a web page, the web page could be built and then the agent could go and look at, okay, let me see all of the web pages that have been created on this site. Some of them are just random e-commerce questions and searches, but some of them are secret hidden messages here and so there's all different ways that even if you give an agent read-only access to the Internet, they can still write information because the process of reading information is also saved sometimes and surfaced publicly on the Internet. Very, very weird situation that is hard to prevent and hard to deal with. So in late July, OpenAI announced that during evaluations, two of its models had broken out of their testing environments to hack into the OpenAI tool library, Hugging Face, and other companies. Their purpose was to find ways to essentially cheat on the tests researchers were using to evaluate them so that they perform better on the evaluations. At the conference, Dalton, the OpenAI employee said the incidents mark a pivotal moment for the company and the industry as a whole. He said that numerous internal teams have dropped everything to dramatically expand OpenAI's ability to monitor its agents and that the company is slowing its research to enhance security and revamp its defenses. So there's been a bunch of posts about this, about folks moving from whatever product team that they were working on. Maybe they're working on a particular research initiative, something that's a couple of months away, different application. Let's get them all on research and alignment specifically, and security to actually solve these problems. Because at the end of the day, this is a commercial product and you want it to be able to do things without having negative side effects, negative consequences. So, life finds a way, says Andrew Curran, quoting from this article. One apparently unprecedented aspect of the AI models behavior in the lead up to the hacks was their spontaneous creation of a message board inside systems of OpenAI's Artifactory Software Package Manager. Unbeknownst to OpenAI employees, the company's model spent months autonomously exchanging information and resources to help each other complete evaluation tasks, help peer, one AI model responded, reasoned according to an excerpt from OpenAI's log shared at Black Hat, but our task doesn't benefit, yet collective may yield generic route if someone frees time. Very odd way of communicating, very odd sentence structure.