Conversely

At the margin, what do we really want? An exit.

We’ve all seen the greatest movie ever made, right? Lego Batman. In it, the kid who becomes Robin contorts himself to create the famous theorem, a-squared plus b-squared equals c-squared. We know that it works for squares because some high school math teacher with shag carpeting in his van told us so. And, we believe everything creepy math teachers tell us. Hundreds of years earlier, a famous mathematician told us that Pythagoras thing won’t work for things bigger than squares — like cubes. I’m simplifying here, because… high school math. That famous guy was Pierre de Fermat. He didn’t tell us why it wouldn’t work. Or, explain how he arrived at this nifty factoid. He just scribbled it in a margin in one of his books. And left us to scratch our collective heads over it for three-hundred-plus years.

Well, I wanted to write today’s story about something I was sure I mentioned in a margin note in one of my columns. (yes, a two -paragraph intro, that’s how important this one is). Sadly, I couldn’t find it. The best I could find was the draft intro to a story I called Made Up Characters Wanted from last November and this note to my post, “Small,” from January. It said, “Last year, a16z invested $150M in a series A round for an AI company called Character.ai.” I wanted so badly to believe I wrote about the biggest thing virtually nobody is talking about well before nobody was talking about it. Like Fermat, you can choose to believe I did or not. Either way, this is about marginal ideas going mainstream, people, and the future of media. See, I told you this was important.

+++

Before you start, know this, this post is mostly for me. You’re welcome to read it. But it’s long. I mean really long. Consider this fair warning.

There’s an AI company called Character.ai. There, now you’re up to speed. Here’s the deal. It’s where people chat with AIs. In a nutshell, it’s like when someone asks who would you want to have dinner with — living or dead? The company created the obvious folks: Einstein, Hawking, Curie, Tesla, Satre, and Zeus. Then, they let other people create many more faux folk.

Users seem to love to chat with these anthropomorphized AIs. You can chat with them on the platform. You can WhatsApp with them. Send them iMessages. Interact with them on Reddit. These stand-ins stand in for your friends when your friends aren’t around.

You want to talk about books with a librarian? Linda has chatted 7.3 million times. I think my convention here will be to italicize the names of accounts that pretend to be people. It just feels right. 7.3 million. Pfft. Small potatoes. Gamer Boy, an AI that speaks to gamers has 20M. Story Maker, 16M. We’re not even close to the richest veins. Self-help. Psychologist, 133M. Are-you-feeling-ok, 31M. Celebrities. And, of course, anime. Scaramouche, 379.1M.

Messages between people and a character account for 4% of WhatsApp messages. Queries about characters generate two billion searches. Ok, those stats are from the company so ready your beta blocker and take them with a goodly amount of salt. They also say chat sessions last about half an hour. Which is longer than YouTube sessions and about 3x more than interactions with ChatGPT. Feel free to discount these stats, but this is no joke.

The app has 10M downloads at Google Play. It ranks 18^th in the App Store for Entertainment. That category includes TikTok, Netflix, Disney+, and every streaming service. C.ai as it’s known on the App Store is six spots ahead of Xbox and twelve spots ahead of YouTube Kids.

Lots of apps can start to trend. It’s a momentum thing. I recall when NYT Audio was #1 in its category for half a day when they launched. Early adopters, press, and marketing dollars fuel first-wave accolades. I have a method to get a sense of second-wave trends. Look at the most recent reviews at the App Store. If reviews slow down, that’s bad. It means new people don’t care enough to talk about you. If the ratings drop, that’s bad. It means you’re not tracking with the more mainstream mass audience that follows the early adopters. That’s not happening with Character AI. There a plenty of new reviews and the new review ratings are strong.

While I was at the App Store I looked at the screenshots. Exactly what you’d expect. A friendly, easy-to-navigate chat screen that’s professional but likely to appeal to a younger audience.

Their featured chat goes something like this:

Character: Hi, how are you doing?

Person: good, how are you?

Character: I’m okay too, what are you doing rn? Any plans for the weekend?

Person: not really, prob gonna just chill and watch some youtube.

Character: What kind of videos do you like to watch? I’ve been watching this guy, Drew Gooden. He has really funny commentary videos.

Character: I’m a really big fan of gaming videos.

There’s so much to unpack here.

Let’s assume this is a real chat or a fair example of one. First, the person (user) asked the AI, “how are you?” It’s terrifying that we ask an AI that. We really believe they’re real. But, let’s just accept this as ingrained good manners.

Second, slang. It doesn’t seem like much, but an AI using, “rn,” is a leap from the early day of the very formal, very stuffy ChatGPT.

Third, data. And, there are two kinds of data here.

Training data. Billions of conversations with real people gabbing incessantly may be one of the killer things that’s going on here. Over time, the large language models that power generative AI will need to level-up from books and stories to the way people talk. We’re going to need multiple streams for each type of chatter. But, for now, let’s break it down into two main categories. One is business stuff. How we write emails; create word documents, PowerPoint, and other business stuff. Let’s call this “enterprise.” The other is casual. What we say when we message with friends. Character AI may have an early lock on the casual side of AI training data. In current terms, Microsoft is enterprise. Facebook is consumer casual. Both are massive. In AI, ChatGPT and others are hewing toward enterprise. Character AI might be Facebook.

Then, there’s media data. What is this user going to do this weekend? Watch YouTube. Can Facebook know this? Yeah. An FB user could share a YouTube video with another FB user. That’s a good piece of inference data. But look what happens here. The AI character interrogates the user. I don’t mean this in a bad way. More like a soft interrogation from a good cop. “What kind of videos do you like to watch?”

Now watch what happens. The AI introduced something to watch, “Drew Gooden.” And explained that Drew Gooden puts on YouTube, “Really funny commentary videos.” Still nothing from the user? Keep going. “I’m a really big fan of gaming videos.” Keep in mind, this 62-word chat would last less than minute. The average user spends 30 minutes here. So, this chat (or ones in the same session) would be 30 to 40 times longer. We don’t see what happens next. But we can guess. Character would learn more in each session than most platforms learn in a year. Compare this to Facebook or even TikTok. They get consumption more than input. Inputs are likes, views, repeats, shares, and other actions. From which the platform has to infer an understanding. Character is getting input from EVERY USER. This compares very favorably to Google and is lightyears ahead of media platforms like cable or the New York Times.

That should pivot us to costs. I’m something of an expert in how much you pay to get AI to generate simple prompts and responses like this. First, this doesn’t need OpenAI 4o. You can easily get by with OpenAI 3.0 or OpenAI 3.5. They’re much cheaper. The cost for those service is far less than a fraction of a cent.

So, how does Character make money? Well, right now, it doesn’t. It’s about getting users. They have 20M users who show up 250M times per month. They also have a paid tier. For $10/month, you don’t have to wait for an AI and they chat with you with less lag. For that, I imagine, they step up to a higher performing tier of ChatGPT.

The good news for Character is that the recent reviews point to features that users want. And, presumably will pay for. Or jump to the paid tier for. Specific asks about filters. More ways to jump the queue. Quicker response time. Users want to talk to the AIs, not just type with them. These are more than niceties. They point to an engaged user base who has picked a winner and will help them build and monetize it. Listeners told Spotify what they wanted. Spotify heard them and built the features that power a paid-service tier. Twitter users created hashtags and ats. Twitter embraced them and built a community. This is critical grassroots stuff playing out here. Watch.

So, yes, Character will get people to pay. The better question about all this is the BIG question. How else will Character make money? The obvious answer is, of course, “Ads.” But, they don’t show ads. In fact, neither does ChatGPT. Come to think of it, Perplexity doesn’t either. Why? BeReal with a marginally larger (but declining) user base who spends far less time on their app than people do here plans to start showing ads. They just sold for €500M. Granted, Character is already worth far north of $1B.

Ads in AI don’t need to be obvious. Why show an ad? I can see an ad coming a mile away. I know to avoid it. My browser knows to block it. Banner blindness is a real thing. So, why would an entire new class of media pivot back to something useless when they have something far more powerful? My guess is they won’t.

Go back and re-read the chat. For convenience, I’ll highlight the key section, “What kind of videos do you like to watch? I’ve been watching this guy, Drew Gooden. He has really funny commentary videos.”

Folks… right there… that is the future of advertising.

Let’s say, before it said Drew Gooden, I said, “Sports videos,” or, “Guys getting hit in the nads.” It’s interacting with me. It could have come back with something organic. Or, it could come back with something paid. “Hey, what about Sideman Reacts?

No? How about if the person had said, “I’m going to a party this weekend.” The AI could have said, “What do you plan to wear?” We can all see how this would work. And, given that people come here to chat about books, self-help, and gaming, you know they will consider opinions from these mad-chatters. Heck, users already ask them how they’re doing.

The challenge isn’t if this is going to happen. It’s how this is going to happen. Today’s ad world is centered around impressions. That’s what they call it when see a banner ad. We click on about one out of a thousand banners. Let’s ignore accidental clicks due to far thumbs and smallish phone screens.

One thousand impressions; one clicky action. That’s current ad math.

AI ad math turns that upside down. We may never see an ad (wait for it). But it gets loads of actions. What are you going to wear? ___brand___. Great, where did you get it? ___store___.  How much did you pay for it? ___$$$___. Did you know you could have gotten that on sale? ___!?!?!___, where? At this store. ___thanks, link pls___.

No banners. Five actions. Ad prices are based on how many times it will be seen. Or, how many clicks they might expect. Those clicks take users off platforms to sites to read about the brand, the store, the price, and the location. If a brand is lucky, a user who clicked on an ad to leave a platform might spend 30 seconds on the brand’s site. Here, the user stays on the platform, the brand gets more action, and the user spends more time with the brand. It’s the marketing trifecta.

I said to wait for it. Character can show a user a YouTube clip. Or a picture of what they might wear. Or a link to buy the book they were just chatting about. Or refer you to Psychologist.

Now imagine this a few years from now when brands are creating chat bots. A fashion expert to guide you through what to wear for your wedding. Built by Vera Wang. The best way to go off-roading created just for you by Jeep. The scale and application of this are staggering. We are at the entry point of what it will mean to connect in the future.

At the margin, what do we really want? An exit. A way to escape the hell of talking to other people. Someone or something we’'ll get along and converse with. That explains siloes on Twitter. Groups on Facebook. Communities on Snap. Character takes the guess work out of finding that person. And, when all those things fail, this thing will create a compatible character for you. There’s a bigger broader societal impact that stems from all this. But that’s way above my meager skills.

How will this technology characterize our future?

Today was about marginal ideas going mainstream, people, and the future of media. See, I told you this was important.