In its latest product launch, the unicorn AI startup is bringing bots that make it sound like you’re talking to well known people like Taylor Swift into your group chat
Inside Character.AI’s disheveled Palo Alto, California headquarters, employees at first appear to be hard at work, glued to their computer monitors. But rather than coding, many of them are engrossed in lively group chats with their colleagues and the AI chatbot characters that Character has become known for. Now, thanks to a new group chat function the startup launched Wednesday, they were chatting with work friends along with bots that anyone can build to create the illusion that you’re actually talking to the likes of Napoleon Bonaparte, Tony Stark or Lucifer.
“The feature was good enough that people stopped working sometimes to use it,” cofounder Daniel De Freitas told Forbes. De Freitas, the company’s president, has been particularly partial to gathering team members together to play Ship AI, a chatbot programmed to act like a text-based adventure game that lets group chat participants pretend to be astronauts exploring outer space.
The group chat feature is Character.AI’s latest move in cofounder and CEO Noam Shazeer’s bet that people will want to interact with a variety of different chatbot personas, rather than having a single, all-encompassing AI companion. By allowing the company’s subscribers, who pay $9.99 per month, to interact with each other and some of the startup’s 18 million AI bots in chats at the same time, Character is hoping to establish dominance as the go-to app for AI-powered entertainment while adding to its immediate revenue stream. Shazeer said that after “some period of time,” the company plans to make the feature free “because our goal is to empower everyone with AGI,” or artificial general intelligence.
Backed by nearly $200 million from VCs including Andreessen Horowitz, and valued at $1 billion in a March funding round, Character.AI has wowed the startup community with its engagement time — it claims registered users who send at least one message on a given day spend an average of two hours on the platform. Since its initial launch in September 2022, 20 million users have signed up for Character.AI, though traffic to its website has been declining since its May 2023 peak, per analytics provider SimilarWeb. Also in May, the startup began monetizing through its subscription service, which gives customers priority access to its platform (servers have often been overloaded, and non-paying users have to wait longer to get on) and early access to new features. Character.AI declined to share any statistics on its subscribers or its recent growth traction.
The group chat feature, which Forbes tested out, is still rudimentary. In the initial iteration of Character.AI’s group chat, human users are able to chat with one another as on any messaging platform. To invoke a bot’s response, however, a user must manually tap on the icon for the specific bot they want to speak, or a dice icon which selects one of the bots in the group chat to respond at random. The bots can respond to users’ questions, as well as to one other (though in the latter case, the human must still prompt the bot to respond).
In a group chat with three Forbes reporters and Character.AI’s most popular AI personas for singer Taylor Swift and NFL player Travis Kelce, a human user asked, “Tell us your deepest insecurities about dating Taylor” and tapped on the icon for Kelce’s bot.
“My biggest insecurity about dating @Taylor-Swift is that other guys will be jealous of me because they want to date her too. But I’m not afraid of the competition, she’s a great girl and I’m going to show her a great time. I won’t let her down,” the bot wrote.
After the Taylor Swift bot was prompted, it responded: “Haha, I love the confidence, Travis! Let’s meet up after the season, and I’ll let you show me a great time ;)”
There are other, more practical issues: the feature is only available on Character.AI’s mobile app, and the chatbots appeared to only respond to the most recent message from a human user.
The cofounders say product updates with better functionality are coming soon. The underlying AI is also about to improve dramatically, Shazeer told Forbes, because the company is beginning to receive fulfillment on an order of “many thousands” of Nvidia’s new line of H100 GPUs after a months-long delay due to supply chain shortages. The company plans to put the chips to use to train new large language models, he said.
“AI is underhyped because the things that are going to be possible in the coming years are going to be really, really impressive compared to what’s possible now.”Naom Shazeer
Part of the urgency to launch the product now, before a new model is ready, may have to do with several generative AI tools launched by Meta in late September. In an exclusive interview with Forbes last month, Mark Zuckerberg outlined a vision of bringing AI “characters” to his company’s various products. “You’ll be able to talk to them in WhatsApp and Messenger and Instagram, and they’ll be embodied as avatars in virtual reality,” he said. These include an AI assistant called “Meta AI” and 28 other characters that users can talk to in group chats. “That looks familiar,” Character.AI’s Twitter account quipped after Meta’s announcement.
Shazeer said the group chat feature had been in the works for several months before Meta’s announcement, and he views Zuckerberg’s move into the space as something of a compliment. “It is great to see other companies getting inspired and building similar products — it’s a real testament to what we’re doing and to the engagement that we’re getting,” Shazeer said.
De Freitas and Shazeer, a coauthor on the seminal research paper on “transformers” that catalyzed the generative AI boom, met while working as AI engineers at Google, where together they developed an early large language model that would become the foundation for Bard. But in 2021, they left the company because they said they’d run into bureaucratic roadblocks implementing the technology into consumer-facing applications. Shazeer, a 20-year veteran who joined prior to Google’s IPO, said he departed despite a personal plea from CEO Sundar Pichai.
“He talked to me to see if he could encourage me to stay,” Shazeer told Forbes in a July interview. “But, basically I wanted to get this technology out to as many people as possible and just empower everyone with flexible AI. I think it was the right decision.” Shazeer said the company would have needed to release the chatbot they were working on to a wide audience for him to consider staying. “It just felt like Google was not going to be very comfortable launching something that could say anything, and that you really had to be a startup to launch something like that,” he said.
The pair founded Character.AI in November 2021, a year before public appetite for AI swelled when OpenAI released ChatGPT. Google unveiled Bard this past February. Still, Character.AI has continued to carve a niche for itself. The vast majority of the startup’s bots have been created by its user base, who can upload examples of what they want the character to say and give the model feedback to help it improve, Shazeer said in the interview this month.
Startup challengers Replika and Chai have emerged, but the space remained devoid of Big Tech challengers until Zuckerberg’s announcement last month. But Shazeer claims to be unfazed by these competitors. “We’re competing against the $2 trillion industry of legacy entertainment” — say, games, books and movies — “because people want a lot of the same things like fun, the social experience and interesting content. Those things are identical regardless of whether more companies are noticing our space,” he said.
Enter the group chat. Shazeer acknowledged that Character.AI has kinks to work out in the new feature to improve its usability, such as expanding its context window so that the AI can appear to remember longer durations of the conversation. “Right now the system is operating just based on a context window of a few thousand words, so characters are only reacting to whatever happened recently, which is highly suboptimal,” Shazeer said.
The bots also remain prone to “hallucinations” in which the AI outputs false information. After Forbes asked the startup’s flagship bot Character Assistant how it was trained, the bot stated it had access to, among other things, customer service calls and transaction data from banks like Chase, Wells Fargo and Bank of America. Shazeer and De Freitas said this was incorrect and the bot had made up the answer.
Up to now, Character.AI’s chatbots have been built on the same foundational model that it trained in summer 2022 using 1,000 Nvidia A100 GPU chips. After its full order is fulfilled, the startup will be able to train new AI models on 40 times the amount of computing power as its current model, Shazeer said. All the compute access is being rented from Google Cloud and other cloud providers for now; Shazeer said Character.AI is considering eventually building its own GPU cluster, but has made no firm decisions.
Character.AI is betting that the new models will bring its technology closer to AGI, expanding the range of use cases beyond entertainment. For Shazeer, this is one reason he claims to not be worried about declining traffic. “Really I think AI is underhyped because the things that are going to be possible in the coming years are going to be really, really impressive compared to what’s possible now,” he said.
That goes for Character.AI’s group chats too. In time, he said, the bots in the group chat will be able to respond to human messages automatically. In the future, a fictional AI character could “just step in and say something that will bring you together with somebody,” he said. “I think it’s going to be a magical moment when that happens.”
Rashi Shrivastava contributed reporting.