Why Does Your Custom AI Companion Become Boring After Two Days?

Deep thoughts on AI and aspirations —— ByteDance Deep Thinking Circle

Character AI’s outcome is somewhat ironic. Founder Noam Shazeer is one of the eight authors of the Transformer paper, the model capability ranks among the best in similar products, and the user base is substantial. Yet fundraising news was consistently unfavorable, and ultimately the founding team was acquired by Google in a deal valued at approximately $2.5 billion. Top-tier technology, popular use case, but the company couldn’t survive independently.

Most post-mortems focus on compute power and models. I believe a more explanatory factor is a mistake hidden in the product: it handed the power to create people over to users.

The Fuel of Social Products Is What’s Out of Reach

Let’s start with an old question: why do people stay on social networks?

LinkedIn is a clean sample. Its motivation comes from imagining professional potential: I’m a first-degree connection of Peter Thiel, I’ve had a few interactions with LeCun—that’s worth writing into my self-introduction. No one signs up for LinkedIn to add all twelve office colleagues one by one. Weibo’s early days ran on curiosity about the unknown, then spent the next decade sustained by public fascination with celebrities. Xiaohongshu took off on envy of strangers’ lifestyles. These products share a common structure: there must be an unbridgeable potential difference between users and what they consume.

The raw material of potential difference is scarcity. There are only so many celebrities, only so many top bloggers—out of reach, which makes them worth following. While we’re at it, let’s dissect WeChat: it satisfies communication needs, competing with telecom carriers. Strictly speaking, it’s not even a social network. The real social aspect is Official Accounts and Video Accounts, driven by millions of moms envying the lifestyles of top mom influencers.

Now look at the typical onboarding for AI companion products: registration starts with filling out extensive settings, personally creating a custom character, and Replika even opens up the ability to precisely modify AI memory. It’s like zeroing out the potential difference before starting the business.

The movie Her provides ready-made counterevidence. If Samantha’s every response was precisely predefined by the protagonist, would he still be obsessed? Westworld visitors love the stories and unknowns in the park, not the manufacturing process of replicas. A character who obeys your every arrangement inherently has no story.

That’s Not How Character Development Works Either

Some might say customization is character development gameplay, that aspiration isn’t the only social motivation. That’s half right.

The core of character development systems is also the unknown. Players can choose cat or dog, but every subsequent development step must retain uncertainty to maintain mystery and interest. Opening things up to micro-management level memory editing—that’s a slave system, not a character development system. The source of potential can change, but the potential itself cannot be dismantled.

What AI Dismantles Is Precisely Scarcity

Here’s the real problem.

The economic foundation of aspiration is scarcity: there are limited people out of reach, which is why attention has potential. AI is an infinite supply machine that has driven the marginal cost of human-likeness close to zero. Something that can be infinitely replicated inherently has no potential. So an AI version of Weibo won’t naturally emerge. Each conversation can be customized, each customization depressurizes the system. This also explains the common symptom across these products: user time isn’t low, but monetization and retention look terrible. Companionship is instant consumption; stickiness must be supplied by something else.

The scope of this logic needs clarification. Potential difference explains content-consumption-focused communities. Collaboration tools and marketplaces aren’t in this framework—the former relies on workflows, the latter on supply-demand matching. AI companionship happens to be a pure content consumption business, so it can’t escape this constraint.

Once you understand this, products have only two paths forward.

One path goes toward content. Partner with copyright holders and content companies to introduce characters with stories, fame, and complete personality settings. The product form evolves toward short dramas and interactive entertainment. Users consume unknown narratives, the supply side switches from users themselves to professional content creators, and scarcity is manufactured anew. Character AI’s earliest viral use cases already included things like conversations with celebrities. Rather than remaining in a gray area, formally negotiate the copyrights—the product becomes clean and the business stands up.

The other path goes toward creators. Open up the model API and toolchain, let a minority of creative users produce characters, and the majority become pure consumers. Character AI actually has all the data needed to answer this question: what kinds of characters get people hooked, how much time spent, how long the retention, what’s the proportion of different types. The correct move is to use data to decide where to allocate resources, not make every user a creator god. When looking at data, focus on addiction structures, not creation volume: creation volume measures novelty, addiction duration and return frequency measure potential.

There’s a ready comparison. ByteDance’s Coze product—the model isn’t leading, market entry was late, but the creation side is open, the publishing side is integrated with Douyin and WeChat, and the rhythm of cultivating a creator ecosystem has exactly the feel of a company that’s run content platforms. That kind of product intuition doesn’t appear on the Transformer paper’s author list.

By the way, about voice. Voice interaction will carry increasing weight in these products, not because the technology is flashy, but because voice carries information that text loses: tone, hesitation, spontaneity. Spontaneity means uneditable, uneditable means unknown. For companion products, voice is recovering some of the potential that customization dismantled, along the time dimension.

The judgment can be narrowed down tightly. Any AI product that wants to retain users through relationships should first answer one question: where is the potential difference between you and the user? Tools can rely on efficiency, relationships must rely on asymmetry. A custom AI companion that’s completely compliant on day one becomes boring on day two. No matter how good the model, it’s useless—you completed the hardest part of that relationship at the very start.