Metaphysic Deep Fakes TED

Tom Graham, CEO of the three-year-old Generative AI startup, Metaphysic, astounded the audience at the prestigious TED conference two weeks ago when he used his company’s AI tools to make a perfect copy of host Chris Anderson, and map it over his own body. What’s more, the AI got the voice right, too, with all its hidden nuance. On video, it looks like Anderson is talking to himself.

Metaphysic broke into pop culture last year with its “Deep Tom Cruise” deep fake video, which was a TikTok sensation last year. They raised $7.5 million in seed funding in a round led by Google Ventures, and Section 32, with participation from Winklevoss Capital Management LLC, and Meta. They are now working on director Robert Zemeckis’ new film for Disney’s Miramax studio starring Tom Hanks and Robin Wright who age from their twenties to their 80s, a pretty hot run for a company he founded in 2020 with Chris Ume​

We sat down with Graham the week following his TED talk to learn a little more about his vision for the future of Generative AI. This interview was edited for length and clarity.

Charlie Fink: Tell us a little about yourself.

Tom G: I went to Harvard Law School and then I was a lawyer, then went back to do a PhD. And then instead of doing that PhD, with a friend in the from MIT, we started Heavy.ai, a GPU and memory database company, that is still going. That was heavily integrated into MIT because it’s kind of like the professors who, you know, a few years later won the Turing award for their work in big data and databases, etc. They were our mentors on that company.

So that was about like getting towards 11-12 years ago. So have had a couple of other companies, analytics companies, one in London, which is just not announced, but just in the process of being sold right now, and then ran a large crypto advisory firm and fund through kind of like starting in 2016 and moving through kind of 2020 or so.

So the kind of the through line for all of the things that I’ve done in technology are big compute stuff with GPUs, doing machine learning, very heavy data, heavy analytics. But Metaphysic is certainly the most fun because the output is really visual, and the impact on society is very profound, I think.

CF: As a result of the deep fake Tom Cruise you came to the attention of CAA. Was that your entree into Hollywood?

TG: Deep Tom Cruise lives in the minds of people. It’s like a big step forward in AI content creation, the first hyperreal thing. A lot of people like Jared Leto reach out to us randomly. The kind of our strategic partners and people that we’ve done things with covers a large spread of Hollywood. And then some of the investors are these big family offices, who own sports teams. We’ve got a really good rep in LA because of the high level creative work.

If people in the entertainment world think that you’re competent, creatively, you get to have different sets of conversations than if you’re like a tech person or a lawyer or a business person or something like that. And so that, I think, has been really hugely helpful for us as a company doing interesting creative work.

CF: How much did it cost to make a digital Tom Hanks you can age and animate?

TG: Well, there are different levels. So let’s say the Tom Hanks model for a face that only costs a few thousand dollars to train because we only put Tom Hanks in it, right. And so if you want to do a Chat GPT, you are going to have to spend $30, $40 million on GPU compute to troll through all of the data in the stick it into a big generalist model. But a lot of the time when you think about visuals, what you look at, right, you can have zero hallucination. So actually, very specifically targeted for zero hallucination.

Even stability, AI, the open source version of stable diffusion that they released that only cost $600,000 for them to train. It’s hundreds of millions of dollars, maybe, but it won’t ever be the amount of money OpenAI spent ever again.

CF: So what and where is the company? You’re London-based? And how big is the company now?

TG: We started during COVID so everything was distributed remote. And the largest hubs are UK, but also across Europe. So like our CTO here is in Lisbon, so there’s a disproportionate number of Portuguese. US people are mostly client facing or business focused. There are some great tax incentives for being in the UK and Canada, with lots of incentives for growing VFX style businesses. In all we’re at 80 and we’re probably going to triple or even quadruple in number in the next kind of year.

CF: And that would be to work on entertainment and games using GenAI?

TG: Yeah, I think that the ultimate goal, if you could imagine where it all leads: to virtual production that is entirely AI driven and animated. So I could have swapped faces and now I’m Tom Hanks or Tom Cruise or something like that. In the Tom Hanks movie, we’re doing kind of face and body swapping, but also backgrounds, and foregrounds.

At Metaphysic we do all kinds of generative AI models, all these neural nets, rely on the process of taking some kind of real world data, some kind of input data and training, and iterating on that. So in the case of, say, young Tom Hanks, what we actually do is kind of build a set of models of how he would look and then generate a synthetic data set so that we have really complete data.

It’s kind of like a 3D scan with all those cameras going everywhere, but much more focused on the front and much more focused on getting all of the interstitial movements. So the AI cares about the difference, but not just the difference between no expression and smile. It needs to know all of the subtle facial movement in between all of that movement is encoded.

CF: Let me sort of zoom back and ask you a question. What happens when we can’t tell what’s real?

The medium is not the message, right? The message is the message. And if it’s so real that we process it like reality, the medium starts to melt away into the background chatter, the background noise of the thing that is happening. And that’s really important, I think, because what we’re trying to work out is whether through photorealistic hyperreal AI generated content, we can generate experiences which live in the mind with the same fidelity as real world experiences.

If you have the power to make reality programmable in some sense, that is tremendous power because like we’re talking about recreating human experiences, right? Kids are not going to learn about the 26th presidential election by reading about it in a textbook. They are going to consume content that puts them on the podium during the inauguration, right? They are going to be in voter booths in the south, in Georgia, like living those experiences through virtual reality, through recreation. They’re going to be talking to kind of finely tuned bots of the candidates. We’re not going to do history like we did in school with a crazy three paragraphs in a book and read an article or something. We’re going to experience it.