A pair of ex-Googlers landed an outsized $150 million Series A funding this week for a Palo Alto startup they lead that lets people chat with virtual versions of anybody they want, from President Joe Biden to a lost relative or a fantasy character.
Noam Shazeer and Daniel De Freitas of Character Technologies Inc., known for short as Character.AI, spoke to Bay Area Inno about why they left Alphabet Inc. in 2021 after helping to lead development of its chatbot software and what they are building with their small team of 22 people.
Among other things, they said they plan to get up to 100 employees by the end of the year and hope to start bringing in revenue in the “not-too-distant future.”
The following Q&A has been lightly edited for length and clarity.
I am talking to the real Noam Shazeer and Daniel De Freitas, right. Not a couple of your chatbot characters?
Noam: Yeah, for now. Though as soon as we can, we'll probably deepfake ourselves so we can skip these interviews and get back to work.
Why did you leave Google after leading creation of its LaMDA (Language Model for Dialogue Applications) chatbot software?
Daniel (chuckling): We just went out to lunch.
Noam: This technology has so many potential applications and can be useful to billions of people in billions of different ways. We wanted to get this in everyone's hands as soon as possible. It's easier to do that as a startup, so we left Google.
You have a lot of competition from other startups like OpenAI LLC and tech giants like Microsoft Corp., Google LLC and Meta Platforms Inc. What advantage do you think you have over them?
Daniel: We do everything ourselves, from the user interface to training our large language model, from scratch. Very few teams in the whole world can do this. Only one company doing this has a guy who invented a majority of the field — us. That helps us drive our quality way up and our costs way down. It allows us to deliver to the user something which is a deeply personalized super experience that lets them choose who they want to chat with. Some people want it to look like a robot. Others people want it to look like a human character. Others want it to look like relative. People are different and we want to support them in many different ways, sometimes with help on a project, sometimes for entertainment and sometimes for an emotional connection. Because we develop everything in the full stack, we are in a unique position to deliver that.
Running this type of AI can be expensive and, while you just raised a very large Series A round, others have already raised or have access to far more than that. Are you going to need to raise a lot more to get to where you want to go?
Noam: The model that we are serving now cost us about $2 million dollars worth of compute time to train. We can build something much better and more useful by spending some more money and we're working on that. This round that we've raised gives us more resources and Andreessen Horowitz is a really great partner. We are probably going to have to get some additional investment in the short term. But in the long term, the real way to grow this is through funding ourselves, providing something of value to billions of people who will pay for it.
There’s no revenue at this point, right? How far away from getting some are you?
Noam: When we had some brief technical glitches back in December and the site was briefly down, some of our users asked us if we had a GoFundMe they could contribute to. Since our launch five months ago, more than 2 billion messages have been sent by our users and the second billion just happened in the last month after it took four months to get to the first billion. Our users today on average spend more than two hours on our sites interacting with characters. Clearly people love what we have to offer and that will help us start to monetize this in the not-too-distant future.
Will your monetization target be consumers or will you get some by partnering with other enterprises?
Noam: Our first option is to monetize through consumers directly. We want to remain focused on them. We haven't told our users what to do with the technology. We put up something general and they're finding a lot of different value there. We released yesterday an early version of our model that is capable of helping people draft emails and also maybe helping them as a study buddy and for some programming questions. To summarize, they get an emotional connection, support, entertainment, sometimes role-playing, sometimes gaming and Increasingly they get help.
Are there any dangers in fulfilling somebody’s request to chat with a celebrity or somebody like President Biden? Do you worry at all that the real people may sue you over that?
Noam: I think our users understand that this is all fantasy and made up. They know that if they talk to a Joe Biden character, it's not actually representing Joe Biden or what he would say. It's entertainment.
Daniel: A good example might be what happened at the advent of movies in movie theaters. When the early ones showed on the screen a train roaring towards audience, some people ran away. People got used to it. You can have far more intense action scenes now that people don't bat an eye at. Two years ago, it would have been impossible to offer people what we are now. This is an incredible moment and opportunity.
