You might remember earlier this week we talked about a bunch of companies signing the Open Secure AI Alliance. This was dozens of companies coming together to say we are prioritizing cybersecurity and open source in AI. There were a bunch of companies on that list: Microsoft, Palantir, Nvidia, Uber, ServiceNow, Salesforce. Cisco was another company that put their name to it, and I want to bring on Jeetendra Patel, president and chief product officer of the company, for his view on all this. Jeetendra, welcome to the show. It's great to have you back.
How are you, Akash? Good to see you, man.
I've been great. You know, there's never a dull moment.
Never a dull moment indeed. So much has changed, and yet the game is still very much similar to when we last spoke. I want to ask you about this alliance that Cisco signed on to. Tell me about the decision to attach your name to it.
Well, here's the way I've always thought about security. The true enemy is not the competitor. The true enemy is the adversary. And so I think what we have to do is make sure that there's as much sharing of information and as much, you know, kind of collaborating that happens within the industry so that we can make sure that we have a secure posture against the adversary. Now, when the adversary is both humans as well as agents that can come in at machine scale, you need to make sure that you've got the right level of apparatus to go out and combat the adversary, and have defenses at machine scale as well. So it didn't take that much kind of calculus on our side to say, 'Yeah, open source is a good thing.' In fact, we've been promoting open source now for the past, I don't know, a couple of years. We had the foundation security model that we had launched, I think, 18 months ago or something. And so it just seems like it's a very natural progression as you move forward. And given the cost of these models and the token consumption getting so high, having smaller models that are purpose-built for specific domains, for specific tasks that can be available in the open source community is just good for everyone.
So, you know, you talk about adversaries and security and how open source enables that. Dario at Anthropic, I'm going to read you a quote that he put in his letter that he put out. He said, 'I don't agree with the letter's assertions that open weights models necessarily make it easier to develop safeguards or that broad access to capabilities necessarily helps defenders more than attackers.' He, of course, is referring to the letter that was released prior to the alliance that you signed on to. But I mean, his point is basically that he doesn't agree necessarily that open source AI necessarily means more safety. What do you think of that?
You know, so I think this is an area that I wish the market had more tolerance for nuance. And I think what Dario is coming up with is specific arguments that are actually very credible arguments. They aren't worth throwing out of the window. And I think there's very, very credible counterarguments that are worth considering as well. And what we end up doing, and if you look right now at the social media sphere, there is a very, very polarized debate on this rather than one which is stemmed on intellectual curiosity.
Right. What are the new, I mean, walk us through the nuanced arguments for and against them?
The nuances on having an open source model is firstly, you want to make sure that an open weights model specifically, rather than just open source, there's three dimensions that you want to keep in mind. There's intelligence, cost, and control. Okay? So the first one is intelligence. Is the model going to be, you know, effective enough to go out and do the task that you want it to do in the most economic way possible, while providing you the most amount of control. And in some cases, you don't care about control. In other cases, you might really care about control. And so you want to make sure that in some cases, you want to just make sure that your model is running in your data center, not just, and I don't want to go out and support, you know, RSI and aggressive self-improvement for another model and provide my data over there. So it's not just for security models, for any model. You might want to make sure that you have... There's a place for, you know, open source and open weights models. I, by the way, as a side note, actually believe that the United States should have a strong stance in open source. I think companies like Nvidia and NeMo Tron are doing a great job, and we need more and more in the Thinking Machines. Mira Murati has done a good job on just coming out with one, you know, a couple weeks ago and...
Yeah, with Ampere link. Right.
Yeah. And, you know, full disclosure, we are investors in Mira's company, but I think in general, it is a great thing to have closed source models and open source models. There is no downside to having open source models in America. Okay?
Right. So go back to the nuances. So we were talking about arguments for and against the nuanced pieces.
Yeah, so the for and against on the nuance of open source models is like, if I have an open source model, I'm going to one be able to reduce my cost, because like, for example, we did Antares, right? We just launched Antares. We had two models, 350 million parameters and a billion parameter model, which is minuscule, tiny compared to frontier class models. In some particular domains, it's actually performing pretty well in benchmarks. And that domain that we have actually focused on is finding existing vulnerabilities in code that you have. Finding the vulnerabilities that currently exist, not zero-day, but existing vulnerabilities in the code that you have, which typically comprises for, you know, non-trivial amount of exploits that happen. We want to make sure that we provide a very, very efficient way to do it and find those vulnerabilities so you could patch them. Having that available in the open-source community, there's no downside to it. There's zero downside to having that.
So what about the nuance arguments against open-source then? Where does Dario have a point here with the credible arguments?
Against open-source is that everyone's going to have it. I'm going to have guardrails in my models, and I'm going to make sure that those guardrails can actually be done in a way that bad actors can't go out and use that open-source against us. That's the argument, you know, against it. It's not a terrible argument. Now, by the way, if you look at what happened with OpenAI and Hugging Face, Hugging Face actually used open-source models because the guardrails weren't working to go on up and kind of say save that. So these are not like simple positions. And by the way, when you start thinking about policy positions that you apply to it, they get even more complicated and even more nuanced than what happens when you... If you banned Chinese open-source in the US, then what happens? Well, now the cost of tokens and the cost of going out and performing intelligent tasks in the US goes up compared to the rest of the world. That then over the long period of time might actually put us at a disadvantage. If you keep it open, you could have a model that has actually been trained to go create some harm on US soil. So like there's a bunch of these things that we have to just think about very carefully and have the assessment done with the level of nuance that's factored in. I just don't think that culturally we need to start, you know, vilifying one party or the other. I have a ton of respect for Dario.
No, I, and look, I think we're... I'm actually more interested in the nuances of the argument. And so, you know, that's...
I gave you one of those, right?
No, no, no, yeah. And I think it's a good point. You know, one of the nuances that I've been asking folks about on this show is with respect to government regulation, specifically whether or not they think the government should have a role in reviewing the models, at least for these frontier closed-source models as we've seen happen in the past. Should the government have a role in reviewing them before they get released? And, you know, I guess on one hand you could say it's slowing the pace of development. On the other hand, you say it's safer. So I wonder where do you fall on that issue? Should they have a role in reviewing the models?
So, by the way, and this is the ability for the government to say, 'If this affects national security materially, I want to be able to intercept,' should always be something that the government has preserved as a right for them. I feel like it's good for the country. It's good for national security. It's good for the citizens of the country. Now, is their reviewing of the model going to have the level of efficacy to be able to find things that would have not been found otherwise is the thing that you have to think about. And is it better for the government to review the models? Is it better for a third-party agency that actually gets specialized in doing that, that reviews the models that the government make sure that they require for some kind of check before they move forward? Those are all aspects that need to be considered. But if you look back at the first principles and the core intent, the core intent is I want to be able to make sure as the government of the United States that I have a degree of control of what I put out in the open market. If there is a national security risk that I want to be able to make sure that I can opine on, that is a perfectly normal thing to do in situations like this, and you can't just be a dogmatic free market player. On the other hand, you have to make sure that that actually provides the efficacy that's needed and doesn't slow down the market unnecessarily just for bureaucracy. And those are the two things that you have to consider.
So you mentioned Chinese open-source models. Are you using Kimi K3 at Cisco right now with your products?
