Aaron Levie on Why Open AI Wins

**SPEAKER_1** (0:00)
Open-weight AI is often for him because of threat to frontier labs. Aaron Levie thinks that gets the economics backwards. The Box co-founder and CEO joins Theo Jaffee and Sofia Puccini on MTS to discuss why open models could make the AI ecosystem more competitive, the debate around distillation in China, and why America needs more open-weight AI.
They also get into the latest frontier models, how AI is expanding rather than shrinking Box's engineering roadmap, and why model routing could become the default for enterprise AI.

**Theo Jaffee** (0:35)
We are live with Aaron Levie, the co-founder and CEO of Box, which does all kinds of things, cloud content management, enterprise documents, permissions, collaboration, a lot of different AI functions. He has an incredible Twitter account, at Levie, he's been around on the website forever. And he writes about AI and many other topics. We have two huge stories today.
I wonder which should we start with?

**Aaron Levie** (1:01)
Astrology.

**Theo Jaffee** (1:02)
Astrology.

**Sofia Puccini** (1:03)
Of course, astrology.
Yeah, huge acquisition in the astrology image gen space. Something we will be monitoring.

**Aaron Levie** (1:12)
Is it an open source play or not yet?

**Theo Jaffee** (1:16)
I think not yet.

**Sofia Puccini** (1:18)
I think astrology might remain closed source for the time being.

**SPEAKER_1** (1:20)
Okay, that's too bad. We got to fight that. Yeah, we really got to fight that.

**Sofia Puccini** (1:23)
But going to the real story. So today there was an open weights letter that Jensen Huang wrote and Box signed. So can you tell us sort of like your interpretation of this letter? What exactly is it aimed at?
And like, what are you guys trying to shape? What do you think Jensen was trying to shape?

**Aaron Levie** (1:41)
Yeah, is the letter maybe like, it's like a Rorschach test of like, what do you see in this letter? Well, what we saw when we read it was, hopefully, it's like the default stance of the folks that signed it, but basically, almost two full sort of main points. One, the reason why open weights AI is super important is because it actually drives AI progress. We get more innovation, we get more options. You can build on top of these models. You can train them for your own use cases. So like open weights in general, I would argue is actually a very important part of the AI ecosystem. So much so that I think it's actually kind of misframed as zero sum with closed weights. It actually just adds to the number of use cases that people then do with AI. So that's kind of probably part one. And then part two, I think embedded in there was a little bit of a call to arms on the US actually needs to be probably even more invested in open weights models. And we probably want even more companies kind of showing up to the table with this innovation.
And so it didn't seem like it was directly about, we must support all of the things that are happening in China as much as US needs to continue to support open weights. We probably need even more innovation here. I know, by the way, distillation is actually not always bad, and there's lots of use cases around it. So let's maybe calm down on some of the foot on that. We support every one of those stances. Open models are, I think, very important to the AI ecosystem. And I actually think it pushes even the closed model providers to innovate faster and be able to deliver more for the market. So I think everybody is sort of a winner as a result of open weights. To be fair to the other side of the argument, I think there's a category for getting kind of commercial and economic interests for a second of some of the closed players. I think there's a category that obviously has real arguments around the safety elements of open weight at some point in the future of capability. I don't agree with that point of view, but I think it's like a healthy debate to have and to have conversations around it. But I almost unequivocally take the stance that more open innovation and open weights innovation is good for AI broadly and the diffusion of AI.

**Theo Jaffee** (3:29)
So diving into the distillation point, how would you separate distillation that is like just normal, basically normal economic activity from distillation that crosses a line into like some sort of civil or criminal violation?

**Aaron Levie** (3:43)
Well, I don't know if I would make that argument. I don't know where that line would exist. I'm almost deeply on the side of, I think it's very hard to make the argument that AI model should be trained on broadly the public Internet, but another AI model can't be trained on the outputs of an AI model. I think it's a very tenuous argument to make. I don't see how you can kind of pull it off. At the exact same time, I understand. And if I were running a lab, I would want to like lock these things down as much as humanly possible. I'd be trying to find every defensive mechanism to prevent the distillation of my models. But I don't know that you'd be able to kind of credibly argue that there's some ethical kind of line that is crossed, simply because then you would be basically arguing against the underlying training runs of these models. Most people didn't get to opt in or out of their data being trained on. Maybe it's in some kind of terms of service from the underlying provider somewhere in the fine print. But like, this is a thing that sort of like, we are assuming that the general knowledge of the world is going to be trained into these models.