Welcome back. AMD signing a new server deal with Anthropic, worth tens of billions of dollars, in which AMD will take a $5 billion stake in the AI firm. AMD may also potentially backstop Anthropic's future data center leases, according to the Wall Street Journal. This comes after news also today, Super Micro announced it had received over $60 billion in new orders in its fiscal fourth quarter, giving them a record backlog, but those deals raise questions about the sustainability of all of this, the spending and the billions of dollars in the AI build out, trillions really. Let's discuss with SandboxAQ CEO Jack Hidary. Welcome back. It's good to have you.
Good to see you all. Good morning.
Where are we in this sort of cycle? What do these deals signify to you?
Well, it means that compute is the lifeblood of these LLM companies. If you look at Anthropic, they cut a deal with SpaceX with Elon for 45 billion spend over a three-year period. Elon and SpaceX have a great arbitrage going. They're leveraging their scale and skill to build out quickly and build at scale and then offering that as an opportunity for Anthropic and other big LLM companies to use that compute. They can build faster than anyone else because of the nature of Elon and his team. And so that's a great opportunity that he can play all day and look for that to be a key part of SpaceX's growth over the next three years.
Yeah. But Anthropic's going into AMD now too. I know that you think this is a really big deal. We should talk about what's happening with Hugging Face and OpenAI.
This is very important because OpenAI was testing inside its own premises one of its new models. And it's in a sandbox, as we say, it's in what's called Exploit Gym. Exploit Gym is supposed to be cut off from the internet, not having access. Now, I think most of our viewers have watched Jurassic Park.
I was just going to say Jurassic Park.
At least every Jurassic Park, what happens? They have a containment facility and the word containment means it's not going to be contained. You know the velociraptor is going to find a way out. And sure enough, that's what happened here. Exploit Gym is, imagine like a game, a video game for the LLM. And you say to the LLM, see if you can break out. See if you can, you know, hack something. And sure enough, it found a backdoor in Exploit that no one knew about and it found a way to get to the internet. That's literally just like a dinosaur getting out. And when it got to the internet, it attacked Hugging Face. Hugging Face, realizing the severity of the incident, needed an LLM, an AI on its side to help it. Sure enough, the closed LLMs couldn't help it because the closed LLMs are trained not to work on cyber. And so it had to turn to an open-source one from Z.AI, GLM, general language model from the Chinese company, Z.AI. And Z.AI, being open source, they were able to help track down the issue and shut down the hack. So, this is a battle of LLM versus LLM now. It's a fascinating moment in time. This is only going to get worse, Carl, not better.
I know, and none of us can figure out today why the security names are trading down. Does that make sense to you?
Well, the fact is, the security industry has been late to the party in really understanding and embracing the threat of LLMs. And it's going to take a combination of the security industry, governments, and in general, industry tech companies to really, you know, address this issue. I'm not even going to say the word solve this issue. There is no solve for this. This is now unbounded in its potential impact.
But could it be that the defense used was this open source model as opposed to going to one of the big private providers?
They had to turn to that because the traditional providers, cyber providers, are really still old school. They're still talking about firewalls and things like that.
That is the last thing. It's time to move on now to using LLMs to protect in addition to looking at the hacks.
But the OpenAI part of the story, I mean, doesn't it just show how dangerous this thing is?
Well, what it means is that if you give an LLM the task and say you can get more reward points if you break out and do something, it's going to try to find a way to do that. And just like Jurassic Park, it probably will.
But in the real world, that's a problem, isn't it?
That's a major issue. Yes. And when we talk about national security, when we talk about defense, this is a significant issue that has to be addressed now. Regulation alone is not going to address it. We need also to bring all the parties together. And I think you're going to see that over the next 90 days, a number of key gatherings on an emergency basis.
It is addressable if we are proactive. We can't just wait for the next one to happen. Got to be proactive on this.
But does that amount to not training your models to push the limits and see what they can do? I mean, is it kind of an opt-in like, okay, we're not going to go to a certain point?
I think that's an impossible regulation to put out there.
