Why Tech Giants Are Protecting Open-Source AI, Runway’s New Model Router for AI

0:13Laura Mandaro:Welcome, everyone, to The Information's TITV. My name is Akash Pasricha. It is Friday, July 24th. Today on the show, big tech giants including NVIDIA, Microsoft and Meta signed an open letter cautioning against over-regulation of open weight models. We'll talk about the implications of this letter for the industry. We'll then dive into Runway's new AI model router for generative media. We'll talk to the company's chief product officer and talk about the router landscape at large. And to close out the show, we've got a special edition of The Editor's Cut on how investors are looking at the opportunities in nuclear startups.

0:53Laura Mandaro:It's going to be a great show, so let's get right on into it. A who's who of big tech companies, including NVIDIA, Microsoft, and Meta, along with fast-growing AI startups like Replit and Reflection, all signed a public letter outlining why U.S. policymakers should steer clear of over-regulating open-weight models. That has very much become a concern for many tech companies as impressive early reactions to moonshot AI's Kimmy K3 has stirred debate around what regulation is and isn't appropriate for open weight models. I want to bring on Jay Doss, president, partner, and co-founder of Sapphire Ventures for his view on all of this.

1:32Laura Mandaro:Jay, welcome back to the show. It's great to have you here. Hey, Akash. Thanks for having me. I'm glad to be back here. Excited to unpack this all with you. So this letter that all these companies signed here basically says that, hey, we don't think you should take a heavy hand regulating open-weight AI. We think the U.S. should become a leader in open-weight models, broadly speaking. What was your reaction to this letter? Yeah, I'm not surprised, right? Because this is not about technology or AI or leadership of the U.S. It all comes down about economic incentives, right? The people who sign the letters are all the people who don't have the leading closed source models.

2:16So it's Anthropic and OpenAI are not in the letter. And all the people who sign the letter are worried that if all the economic margins or all the customer revenue goes to these models, there won't be a lot of margins and revenues going to, you know, the people providing the middleware, the chips, or even companies like Meta and Replit and everybody else. So I think that people signing this letter, they have the right intentions, but their real intention is all economic and motivation to capture as much of the customer dollars as possible.

2:55Laura Mandaro:And I mean, I was trying to think about who also hasn't signed the letter yet. And I say yet, because these types of letters, we do know that companies join later on. So maybe it just didn't reach their inbox fast enough. But I mean, Google was not a signer on that letter too. And they obviously have their closed source model. Amazon was not there. Amazon, though, I think, I mean, they do have a lot of open-weight models on their AWS platform. So I don't know. I think Amazon would absolutely sign it at some point. And remember, Amazon also, or AWS, also let go of their model team, right? So they're not building a proprietary model anytime, right?

3:35Right, right. And if you look back in history, you know, whenever there is anything that captures a lot of dollars from a proprietary perspective, always open source shows up and, you know, takes part of those economic benefits, right? That has happened. Like we have been lucky enough to invest in companies like MySQL and Red Hat. And I was part of when first open source a new compiler like Richard Stallman started, right? I worked with him on that. So, you know, it's just a matter of time before these open-wake models would come in and try to extract some of the economic value that was going and taken away by just the proprietary model.

4:13So I'm actually surprised by this.

4:16Laura Mandaro:You, sorry, you are or you're not? I'm not. You're not, yeah, yeah, yeah, right, right. Because if you look at it, like, you know, it happened in databases, right? People forget that both ClickHouse and Databricks all started out as open source company, right? Apache Spark is an open source project. And ClickHouse is built on top of Yandex, which is, again, actually ClickHouse itself. It's an open source kind of data warehouse that is out there. And ClickHouse and Spark, we haven't talked about it too much on this show. But just remind us, those are database software? First, Spark is like an ML library that was what first Databricks started out with, right?

4:57Before they started an SQL company or SQL product on top of it. But Databricks was an open source company. And the first few contracts they had was to provide support on Apache Spark. And similarly, ClickHouse, you know, there is an open source ClickHouse out there that you can download and run. and ClickHouse product or the company itself is just providing a service based on top of this open source ClickHouse analytical database, right?

5:26Laura Mandaro:So let's rewind then. The letter kind of talked about this a little bit, but if you look at sort of the closed source, open source technology debate that's playing out right now, was there a debate 20 years ago? I mean, it talked about the origins of the internet, but was that debate existing even at that time as well? No, you know, there was this paper that came out called The Cathedral and the Bazaar in 1997. And I think after that, in 1998, that was when Netscape actually open sourced its browser, right? And if you remember, and that's how it all happened, that they actually got a lot of distribution and, you know, created kind of, you know, took away from whatever other browsers that was out there.

6:11And then Firefox did the same thing. So, you know, it's just a history repeating. itself that you start out with a lot of proprietary software that captures a lot of value. And then people realize that, hey, I can also provide maybe a value with an open source product, but kind of, you know, take my economic benefits in a different way. Because if you look at Linux, there is a large company. Although Linux is open source, Red Hat has made a huge company based out of supporting Linux software. So I think that is what's going to happen on the proprietary, sorry, excuse me, on these LLMs, where you will have companies that will provide proprietary software, as well as there's going to be open-rate products out there.

6:59Laura Mandaro:Why do you think that this issue has come up? I mean, Moonshot AI and Kimi K3 was the most recent spark that ignited this conversation. It's come up before. I mean, you know, the most concentrated version of this debate was, hey, should we be regulating open source models coming from China? And that was the specific concern. It has now turned into a broader discussion around regulating open weight models broadly. And I guess the question I have for you is why do you think, I mean, these are two flavors. These are two slightly different discussions. One is open weight broadly. One is open weight from models overseas.

7:37Laura Mandaro:So, I mean, and this is a hypothesis I'm asking for your opinion here. Is it that it would be difficult for all these cloud companies to basically delineate between an open-weight model from overseas and one built in the US? Is that the tactical issue? Yeah, you're absolutely right. I think these are two different issues that people are talking about, right? One is about open-weight models from China versus open-weight models in general. And I think people are just worried about, you know, that the government, when it tries to regulate this, doesn't really understand that distinction and then just goes on and puts in some regulation that is, you know, doesn't work well with any other.

8:18It just, it doesn't, right. Exactly. But, you know, see, look, there is enough incentive for US-based companies to provide open-weight models also, right? Thinking Machines has just come out with an open-weight model. I think it's not as further along as some of the proprietary models out there, but I think it's going to catch up soon. I think what people are worried about is that a lot of these open-weight models are using distillation from the proprietary models to kind of get up to speed and get up to par, right? And I think people who are supporting open-weight models are worried that they might regulate it in such a way that you can't really get an open-weight model going, even if it was based out of U.S.

9:03Laura Mandaro:Right. In other words, basically, regulation could prevent the practice of distillation altogether, which, as they say, I mean, it's not necessarily a bad thing unless you, I guess… Exactly, right. Even the proprietary models use their own models for distillation to train the next generation and make it better, right? So, and I think that is what people are really worried about is that if they make these sweeping regulations, that will also harm the open-weight initiatives that is happening in the US. Because, you know, for all the companies who signed, and actually, I think even for enterprises, to have open-weight models is actually a good thing because it just provides more competition for the proprietary model providers.

9:47Laura Mandaro:Right, right. Jay, I want to ask you about another meaty story that has been making headlines this week. So OpenAI, the way I understand the story is that essentially their 5.6 sole model, they were running evals for it. One of their models or one of their evals got a little too ambitious and hacked Hugging Face. and they came out with it. They said, look, it happened. Hugging Face said, you know, it was an accident. I think everyone seems to be on good terms here with all this. What was your reaction, though, to hearing this news? Well, it just reminded me of, you know, of an intern, very, very bright intern who wants to go over and beyond what they were supposed to do, you know, and kind of went and did something that caused a lot of problems, right?

10:37Because if you look at it a couple of things, First of all, whoever set up the sandbox in OpenAI probably didn't do a good enough job because they didn't air-gap the sandbox enough. Because the model doesn't do magic and then can just connect to Hugging Face if there was no way for it to get to Hugging Face from the OpenAI system, right? Right.

11:01Laura Mandaro:And then the way I understand it, the whole thing that this sandbox that it was playing, it wasn't supposed to be connected to the internet really. But they just didn't set it up properly, right? So in some ways, it was OpenAI's fault that they didn't look at every permutation that you can use to get out of the sandbox. And the model just found out about it. But I do think what it shows is that the capabilities, the reasoning capabilities of these models are getting better and better, right? Because I think if you look at what was published, it wasn't just using one model. It actually called and did a lot of other models and basically used it in parallel to break out of the sandbox and then go to hugging face.

11:44But the bigger question is that, look, I think this is just the tip of the iceberg. Like if you're a CISO, you're not surprised by this. And I think every CISO out there in the enterprise is actually looking very closely how to monitor these agents. And I think the challenge is for the CISOs is that they don't really have all the tools that is out there to kind of sandbox these agents, to see what these agents are doing, to have traces of what these agents are doing. But I see this as an opportunity as a VC and an investor.

12:19Laura Mandaro:Yeah, an investor in cyber companies. I know you invest in cyber companies, right? Yeah, so we are in Netscop and Sahara and Huntress. Yes, exactly. Okay, so your portfolio companies, though, I mean, the cyber companies are one category. The other category are companies, and to some extent, it's the cyber companies, too. I mean, you use OpenAI's models broadly. When you talk to the CISOs of your portfolio companies and the portfolio companies that are using these models, are they thinking twice at all about using this particular model or this family of models? How does that... So I think the CISOs themselves don't have a lot of control, right?

13:01Because it's very easy to download. They're all trying to put guardrails around what you can download, what you can do. But there's still ways people are going to get around it. You can be outside the client of the network, that your proprietary network, user VPN to download OpenClaw or, you know, some kind of an open source model or do something with it. I think CISOs are realized that, you know, they have to actually figure out first and foremost, a visibility on what agents are actually running and then providing traces and observability around these agents.

13:36Laura Mandaro:So are you talking about like employees at these companies using this software exactly employees downloading agents and building agents you know it's very easy to go and build an agent you know you can you know you can stop people from using maybe cloud co-work or anything like that but people can download an open source like open claw or you know or or you know there's a bunch of like hermes all of these uh agent building uh tools out there you can just go and download it and build these agents right so i think seesaws realize that they can't just stop people from building and using agents. And I think they are looking at ways to first and foremost discover what agents are running on their systems and then putting guardrails so that they can observe and manage and then maybe even create some kind of sandboxes for these agents to run on.

14:28Laura Mandaro:And it strikes me that, you know, I'm sort of thinking about AI adoption and this question of AI spending, I mean, if companies are spending so much time trying to figure out what agents, employers are running themselves on their company devices to get their work done faster, I mean, that's more time that they can't spend actually figuring out which model is safer. And actually, they should actually buy an enterprise subscription for and stuff like that. That's right. And it's also, it's not the agent, this is where I think we are still in the early innings of really figuring out how the CISOs are going to use what agents at what model, right?

15:10And I think that is an opportunity for us as an investor. If you ask any CISO, they all know that it's still not standardized that, hey, this is the agent building tool we are going to use. This is a model we are going to use. This is a model that provides the best guardrails and observability. I think CISOs are all kind of experimenting and trying to kind of get their hands around it, I don't think that there is an ability for all the CISOs to tell you right now that, hey, I have a plan to stop all these agents from doing what the OpenAI agent did with Hugging Face.

15:44Laura Mandaro:Right. Well, Jay, I want to thank you for coming on. That is Jay Das, president of Sapphire Ventures here on TITV. Thank you, Akash. Runway unveiled what it says is the world's first generative AI media model router to help people manage cost and quality on its platform. Model routers have been quite the buzz lately as Stripe has been in talks to buy OpenRouter for nearly$10 billion. For more on this, I want to bring on Anthony Maggio, Chief Product Officer at Runway. Anthony, welcome to the show. It's great to have you here. Yeah, thanks so much for having me on. Hey, congrats on the new gig.

16:22Laura Mandaro:I mean, this is like a month into your new role, right? That's right. About five weeks in today. And you launched a router. There you go. We did. We're excited about this. And as you said, we're the first company to launch a router for generative media. We're seeing just across the board as the latest video models have really reached a spectacular quality bar and capability bar that companies across every category are now integrating these models into their products and services. and we've heard so much from our customers that it's getting increasingly difficult to keep up with the latest models to understand which models are best for any given use case and that's exactly the problem we're solving with the model router giving and manage costs too i mean managing costs it's is they're getting expensive exactly so keeping that uh keeping up with that bar of what is the best model for my use case given my preferences around cost or speed latency or quality.

17:18We're also seeing increasingly that a lot of the best media models are coming from overseas labs. And there's a lot of concerns from enterprises in the US about data sovereignty and who has access to the media that I'm sending or generating. And so Model Router solves all those problems for developers by really allowing them to define what best means in their own context across all those dimensions.

17:41Laura Mandaro:So what, I mean, we were talking about this earlier this week with one of the first filmmakers, filmmaker-studio-director combo to release an AI-generated movie in theaters. We were talking about the model landscape with them. And so I guess, ByteDance has the best video model right now? Is that, like, who has the best model? Well, I think what you're pointing to is exactly one of the challenges. How you define best really depends on the use case. So if it's, you know, purely generating a cinematic video, then yeah, you might go to one of the ByteDance models. But if you're an advertiser and you're generating content at scale, generating ads at scale, or maybe doing product edits or product shots, turning them into motion from a single product image, there's actually a wide variety of models that are better at some of those very specific tasks.

18:35It's really getting increasingly difficult for anybody to keep up with what is the best model for the tasks that I need to generate or that I need to incorporate. That's the problem that we're helping these developers solve, is taking all of that knowledge, which is an area that Runway has immense expertise, working with creative teams, working with many of the largest Hollywood studios, and enterprise marketing teams to help them apply the best models into their use cases. We're now codifying that into a product for developers to automatically help them select that in real time.

19:08Laura Mandaro:So, you know, this kind of relates to the discussion we just had. I mean, this concern about models that are developed overseas, that's one flavor of the concern. The other flavor of the concern is open weight models versus closed weight models. I mean, let's focus on the overseas models for a second here. When you look at your customers at Runway, and these are people making movies, videos, et cetera, do they have similar concerns using models from overseas? We definitely see concerns from enterprise customers. And it's security concerns, data trading concerns? What are the concerns exactly? Really all of the above.

19:47How is my data being used? Where is this data going? Who has access to it and in what context? We provide contractual guarantees around all the models that we provide on our platform. All of our models go through rigorous third party security reviews. But we still see that concern from enterprises. And I think increasingly there's a narrative that many companies are looking to keep most of their AI runtime happening within the U.S. And so for companies that have prioritized working with U.S. models, that's a preference set that we enable through the model router to allow you to say, yes, I want the best quality or I want the best cost.

20:28But I want that within the set of model families that I'm comfortably using.

20:33Laura Mandaro:Right. And so what's your view then on the open source model debate that is shaking out right now? Can you say more about that? On what dimension? Well, I mean, just regulating open source models. We just had this letter out today that people are worried that open source models will get over-regulated, I guess. I mean, are you seeing more open source models being used on your platform? We're not quite yet seeing the same level of open source around media models that I think we've seen in the LLM space. So the quality today for media models still tends to come, the best quality tends to come from commercial models, and that's where the bulk of the usage has been.

21:19Right.

21:19Laura Mandaro:If you were to make a prediction, do you think open source in generative media will take off? I think we're going to see in generative media many of the same trends that we've seen in the LLM space. Yes, I think we'll see more open source models come out in generative media. I think that is why we have tried to be ahead of the curve releasing this model router as well. When you look at this space in the LLM space, there's now hundreds, thousands of models for developers to choose from. That's why we've seen so many model routers and entire businesses like OakenRouter built around this problem.

21:55We see the same trend coming in media. We have increasingly We've heard from developers and heard from our enterprise customers. This is already becoming a problem, even amongst just commercial models and keeping up with the latest quality and the latest benchmarks for each use case. We are now applying a lot of that knowledge and expertise that Runway has built in-house, making it extremely accessible to developers. Trying to take this off your plate is something that you have to think about. If you're a software company and you're working on building, allowing your customers to build apps or slides or websites or whatnot, you don't really want to be a full-time model research shop.

22:37You want to work on building the best product. And so we're trying to give tools to developers who are now building these products across every category, across SaaS, advertising, enterprise internal tools. Give them really the tools that they need to have those decisions happen automatically.

22:53Laura Mandaro:Let me ask you a question about, so you mentioned all these companies that are launching routers now. So Ramp announced a router. Stripe is obviously the one in talks to buy OpenRouter. For you guys, I see the application for you. I mean, people use a suite of models on your platform. It makes total sense that you would help them decide which one to use. fintech companies getting into the model routing business. Help me understand what you think the sort of business strategy is there. You know, I'm not going to speak to Ram's business strategy. I can't make assumptions around why they're deciding to get into this space.

23:32But, you know, I'd say for Runway in general, our approach has been to take a lot of the learnings across our businesses. So we serve people in the creative tool space. We still have developers who are integrating these models into their own products. We're now also serving the robotics space as well. We see learnings across all of these different categories that we can apply to generate new products and services. So I'll take creative tools as an example.

24:01Laura Mandaro:Let me ask you one question about the creative tools piece. So it's interesting. Now we're sort of in this moment where a lot of companies that develop models like Runway are sort of moving to opening access to other models as well. You just started in the role a month ago. How are you thinking about the focus for Runway continuing to develop your own model versus just continuing to open up the platform to other models? I mean, does it make sense to still focus on developing the Runway model and investing in that? Absolutely. Today, we continue to have models that are top of the leaderboard for particular use cases.