Victor Riparbelli
speaker
267 appearances
2 recordings
2 series
first heard Jan 2025
last heard Mar 2025
Victor Riparbelli’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsNo recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.
Appearances
OpenAI clearly managed to capture the consumer as in the destination that you go to to use an LLM. I think we'll see LLMs being a part of many different apps.
And I think owning X along with building the LLMs is actually really powerful, especially for the real-time information that will be able to feed into the models directly from X. And the fact that X, of course, already has hundreds of million users that in theory at least could start using their LLMs rather than going to
Clearly, there are still gains to be made from scaling, I think, especially if you look outside the text domain, but like in video, audio, etc., 3D. So you agree that scaling laws will continue? I think they will continue, but I don't think it's just going to be linear, like whoever has the most compute is going to win. The world rarely works in those ways. Something will happen, right?
Someone will come up with an algorithm that is like 100 times as efficient as what it is today. I think compute is important, but I actually also think data is... Actually, I would say compute algorithms and data, maybe. I think in algorithms, what everyone is trying to build into more or less any AI system today is control, right?
We've proven these things are extremely capable at replicating the real world, producing video that looks real, audio that looks real, text that sounds real. But what we all really want to do is get a deeper level of control over these things. In my world, it's like you have some of the big video generation models.
I kind of delineate what we do versus like what Zora does or Runway or something like that. Zora and Runway, these models, extremely capable, extremely powerful, right? You type something in, basically anything you type in, it'll actually spit out. And that's really, really powerful. And it's a great demo.
But to really be able to use this, you have to be able to say, I want this same character in a different scene. I want the character to say this particular, like there's so many layers of control. And those are very hard to build in because every time we try to build them in, it decreases the overall fidelity.
Because all of a sudden, you're trying to make the model do what you want it to do, rather than just replicate whatever a real kind of looks like. And so I think getting those layers of control in is going to be difficult. I think that is going to be a lot of algorithms figuring out, like... To what extent is Grok's lack of layers of control a problem or a benefit?
Well, there's controls like content moderation, and then there's like controlling the output, right? So the thing that I'm talking about here is like the slot machine kind of thing. You type something in, you get something out. Then you try and change the problem, but you pull the slot machine again, right?
And it's very frustrating if you're trying to get to something specific because the model just doesn't really put it out. You want to have some degree of control in there. I think on the moderation, that'll be interesting to see how it's going to pan out, right?
I think we're definitely in the middle of like a vibe shift with Zuckerberg also pulling back moderation, which I think in general, I'm in favor of that. I don't think that like humans moderating content is the right way forward. I think community notes.
I've seen the statement. I don't have the actual details of it, but I think it is directionally correct that having humans sitting down and evaluating content is not the right way of doing it. I think what we've seen with products like Wikipedia, for example, is that the collective power of people working together to arrive at some sort of truth is really, really powerful.
And I think community nodes is kind of like taking that Wikipedia way of thinking about the world and trying to implement that into every single piece of content. And it's not easy and it's not solved yet, but I do think that is the right way for us to have some degree of control of what people do and say. Where it gets really messy is this kind of like... gray content.
We have the problem at Synthesia as well, right? We have basically a product internally for content moderation. The hard thing is you have what we call the green content. The content that everyone agrees is great. That's 99.9% of the content. You have the red content, hate speech, violence. Most people will agree that that's bad as well.
Then you have all the gray middle, and that's where it gets difficult. I think this is the content that Sock is talking about here, where he's going to let off controls. This is essentially like you're making cryptocurrency content, for example. When is it a really enthusiastic entrepreneur who just started a new coin that he or she definitely thinks is going to change the world?
When is it an outright fraud of trying to get well-meaning people to dump their money into something they think is going to go up 10x? Those lines are very difficult to draw, right?
Yeah. So today we are, and that's a decision that we have made. We had a lot of discussion about this in terms of how do you, what's our approach to this problem? And for me, again, it comes back to the customers, right? We're an enterprise product. It's important that the avatar that we have are not seen to be used in all sorts of like wacky content online.
We don't, as a company, have, we don't see ourselves as having to uphold any kind of right of free speech. And frankly, from a business perspective, having someone pay me $30 a month to create very questionable conspiracy content, it's just not good business. So for me, all of those things kind of lined up, right?
Our enterprise clients, they don't want us to be affiliated with content that doesn't kind of like match the brand. And economically, it just doesn't make sense. So we've taken a very strict approach. Which means that we are actually going and being the arbiters of truth. And we definitely have people who are unhappy with that.
I think this is like classical VC brain. Too much focus on the technologies, right? Like, you know, you speak to a... I think there's a big misconception about Synthesia is that we're like an avatar company. And it's kind of a fair assumption to make because you go to a website, if you like, you know, see about us online or like... Advertising is a headline feature for the company, right?
Showing 121–140 of 267 · page 7 of 14
← Previous
Next →