Runway's New Video Model Challenges Rivals Google, OpenAI

Bloomberg

Watch Runway's New Video Model Challenges Rivals Google, OpenAI - Bloomberg on Dailymotion

Transcript

00:00Last time I checked on the leaderboards,

00:01some of them already have you at the number one spot, Cristobal.

00:05What is it that yours offers versus Sora 2 versus the latest Google offering?

00:10Yeah, that's true.

00:11So we just released Runway Gen 4.5, our latest video model,

00:15and it tops the charts in terms of performance and benchmarks across all other models,

00:20which is kind of like a big deal within research.

00:23It's the first time a company has led the leaderboards,

00:29and this company not being basically a large research lab.

00:32It's a model that surpasses pretty much all other models with incredible consistency,

00:38really good realistic results, and just like across the board, amazing creative results.

00:42So really, really excited to get this model out and have people use it.

00:45Now, you had it out there on the leaderboards with a pseudonym

00:49before it was launched in public, Cristobal, and you had it called David.

00:54Is that a Goliath-David construct?

00:56How are you competing against these vast generative AI companies?

01:02Yeah, it was a little bit of a play with that, with David and Goliath.

01:05I think we've managed to outcompete the largest research labs by being very focused.

01:10I think it's the area of both research and efficiency,

01:13and if you're able to maintain the focus as a team, you're able to deliver kind of groundbreaking results,

01:20and we're proving this.

01:21This is the first time I think we've, again, anyone has kind of topped the leaderboards,

01:25not being a large, well-funded research lab.

01:28And I think part of it is really like the team, and it's really also the vision.

01:32We've been working on this for almost seven years.

01:34We started working on video models when there weren't even, like, the other boards to start with.

01:38And I think eventually you build some sort of intuition and really good, like, momentum

01:42as to how to improve these models over time.

01:44And look, this is still, like, the worst the models will ever be.

01:47And so we have a bunch of more releases coming up that I think will further improve

01:51both pre-training and post-training, and so we're very excited for that.

01:55Cristobal, you're not that small.

01:57You raised $300 million in April at a foremost $4 billion or $3.3 billion valuation.

02:04I do think there's some value in you explaining what was different this time around in the training

02:09of Gen 4.5 and the data set.

02:11What is it that you did differently and that has allowed you to release such a competitive model?

02:18I think there's a lot of different things.

02:20On the one end, pre-training has been one of our, like, focus for a long time,

02:25making sure that both the algorithmic improvements are there, but also the way we caption,

02:29we structure data, we build the models themselves and test them.

02:33The best way, really, to think about a lot of the research is you need to conduct multiple

02:37experiments through multiple months.

02:39And there's a lot of learnings within how you run those experiments.

02:43And I think what we're kind of proving is that infinite resources, I mean, you're right,

02:48we're definitely not in the smallest side of a company, but we're not a trillion dollar

02:54or $4 trillion company, yet still managing to outcompete the resources of those companies

02:59is kind of insane, to be honest.

03:01And I think a lot of it has to do with the experiments and the research taste,

03:05which is if you're running all these experiments, how do you make sure they're effective and

03:09efficient?

03:10And I think that's, I think, something we've done really, really well, which is focusing

03:13a lot of pre-training.

03:15Gen.4, 4.4, 4.5, apologies, has been released to your enterprise customers straight away.

03:21What's the business model for it?

03:23You know, how do you guys monetize on top of that?

03:25You've just talked a lot about research.

03:28I mean, it's pretty straightforward.

03:29We have subscriptions and we have credits that people can buy to use the model.

03:33We're releasing this model to gaming companies, to studios, to brands, to production companies,

03:38to creative around the world.

03:39We have tens of millions of users actively using Runway.

03:43And again, the efficiency side is not only coming from the pre-training or the training

03:46side of things.

03:47We've also been incredibly efficient in deploying the models.

03:50We partnered with NVIDIA for a lot of this work, and we managed to get really good performance

03:55of inference.

03:56And so we actually make money every time you use the model.

03:58And that's, I think, a remarkable feat, not only on how we think about the research that

04:03needs to be done, but also the market and deploying this so the unit economics makes sense.

04:07Christopher, sorry, just real quick.

04:11So you're saying you're profitable running GEM 4.5?

04:15No, I'm saying we're making money by every time you use the model, we have a good margin

04:19on the model itself and how you use it.

04:21We managed to deliver really good performance on inference, so the model is still cheap to

04:26use compared to other models, while still being the best model in the category.

04:30And that's incredibly hard to do.

04:33Again, it just goes back to the focus the team has set for quite some time.

04:36So, let's do it.

04:37Let's do it.

04:38Let's do it.

04:39Let's do it.

04:40Let's do it.

04:41Let's do it.

04:42Let's do it.

04:43Let's do it.

04:44Let's do it.

04:45Let's do it.

04:46Let's do it.

04:47Let's do it.

04:48Let's do it.

04:49Let's do it.

04:50Let's do it.

04:51Let's do it.

04:52Let's do it.

04:53Let's do it.

04:55Let's do it.

04:56Let's do it.

04:57Let's do it.

04:58Let's do it.

04:59Let's do it.

05:00Let's do it.

05:01Let's do it.

05:02Let's do it.

05:03Let's do it.

Category

Transcript

Be the first to comment

Recommended