Jason Yuan on Hivemind, Social AI and the Risks of Sharing Between Friends

Hivemind founder Jason Yuan joins Ellis Hamburger to discuss a social AI that asks people in a network for information and helps connect them. Yuan describes it as a storyteller guided by a concept of a “good story,” while acknowledging that its privacy and anonymity protections are not foolproof.

Yuan says the launch is his first as a solo founder of a venture-backed company, although he credits the friends and collaborators who inspired and supported it. He jokes that he injured his foot protecting his laptop while finishing the launch video.

The launch video uses a provocative dating example. Yuan says it reflects a true story, though he jokes that his preferences have changed. He wanted the presentation to convey both the company’s playful consumer-product sensibility and the serious research and philosophy behind it.

On his birthday, Christmas Eve, Yuan was alone and working. An early version of Hivemind asked what present he wanted; he mentioned a board game. The conversation turned to loneliness and his dating history, and the system began asking his friends about him.

Yuan says friends described his past relationships and suggested a different kind of match. At the same time, the system had been helping people find places, and it asked one friend about wine bars. He describes it combining those threads into a possible plan.

The system encouraged him to spend New Year’s Eve with friends in New York and helped coordinate dinner preferences and timing. It also suggested asking a friend about a potential date. Yuan says he went, met the person and enjoyed the evening, although he still did not get the board game.

Hamburger notes that the video ends with a small group of friends together, rather than a person alone with an AI. Yuan says the product is intended as neither a bullying gossip service nor merely a social companion; his ambition is to create opportunities for people’s own stories.

Yuan calls the system’s alignment target a good story. He describes a planned constitution with principles including truth, multiple perspectives and human agency. He says the document is not finished; these are design intentions, not evidence that the system will consistently produce the intended outcomes.

In Yuan’s view, people make meaning by narrating their lives and relationships. He wants Hivemind to chronicle and suggest opportunities without becoming the master architect of anyone’s life. A good story, he says, should allow its characters to make choices.

He gives the example of learning that an estranged friend was moving away. Knowing this made him consider repairing the relationship before losing the chance. He says the system did not order him to reconcile; it supplied information that changed what he could choose to do.

At the time of the interview, Hivemind was in private beta. Yuan says people first express interest by messaging it on X and may later be assigned a number for text conversations, including iMessage. He wants it to meet users in familiar channels rather than demand a new destination.

The system also asks users questions about people and events they know. Yuan says it tries to understand relationships and connect threads across conversations. Hamburger observes that using an existing messaging interface can make a product easier to reach than another separate app.

Yuan says people often enjoy answering questions about subjects and people they care about, much as they react to social posts more frequently than they create them. He recalls a playful debate about coding tools as an example of the system eliciting opinions.

Hamburger asks whether anonymity is enforceable and whether the system can be hacked. Yuan says it is a closed-beta social experiment and explicitly refuses to promise ironclad protection. It sometimes attributes opinions to people, and its behavior relies on model interpretation rather than deterministic rules.

Yuan says the team tries to steer the model using its constitution and imagined social outcomes, but cannot promise 100% reliability. That limit matters for anyone considering sharing sensitive information: the interview does not establish a guarantee of confidentiality or consent.

Hamburger recalls anonymous social products that initially seemed positive but later enabled harmful behavior. Yuan says Hivemind is not exclusively anonymous: it may name a friend when it judges that doing so helps someone connect, for example by mentioning a friend who has visited a wine region.

Yuan frames the disclosure question as how much information a person needs to make an informed choice while pursuing a positive outcome. He gives hypothetical examples involving exposure notification and a job opportunity, where sharing less identifying detail might still prompt useful action.

He says the system should absorb some social cost rather than amplify every negative comment. If someone dislikes another person, he imagines it could simply avoid recommending a meeting; if a pattern emerges, he says it might ask why. These are proposed behaviors, not independently verified safeguards.

Yuan says the goal is not to automate conflict resolution or take away people’s agency. Hamburger compares the system to a mutual friend mediating a disagreement and notes the power involved in controlling which information passes between people.

Yuan recalls asking for opinions on a domain name. The system offered an interpretation of why one person disliked an option; he says that explanation helped him understand a perspective even without agreeing. It was an interpretation, not necessarily a direct statement from that person.

Asked about crushes, Yuan says the system should not casually disclose one person’s feelings to another. He imagines withholding a name or waiting for more context. Hamburger suggests a less explicit introduction might preserve room for people to decide what happens.

Yuan recounts a brunch the system organized after a visitor asked what to do in San Francisco. He was surprised by the invitation list. Because the system could not make the reservation itself, it asked a participant to book the restaurant.

At the brunch, Yuan says two people interested in a robotics startup met and later explored working together, one bringing product experience and the other engineering. He presents this as an example of creating an opportunity without scripting the outcome.

Hamburger raises a harder case: what if a group of friends endorses a medical claim that is false or unsafe? Yuan says the system should seek qualified knowledge and question the claim. This is an intended approach, not proof that the product can reliably evaluate health information.

Biz Stone appears in the launch video and is both investor and adviser, Yuan says. The two discuss Stone’s earlier product Jelly and other question-and-answer communities. Yuan sees Hivemind as concerned with subjective opinions as well as objective knowledge.

Yuan says many people want to help by sharing knowledge, while fewer know exactly what to ask. His proposed opportunity is to encourage contributions and make them useful to others. Hamburger asks whether Hivemind maintains a knowledge base or answers each question from scratch.

Yuan says questions prompt fresh inquiries, but responses can build on an existing corpus. Because information changes, the system may revisit a recommendation later—for example, asking whether a restaurant is still someone’s favorite.

Hamburger contrasts personal recommendations with web search results from sources of uneven quality. Yuan argues that many meaningful questions are subjective: whether to attend an event, change jobs, give a particular gift or improve a friendship.

Yuan defines taste as an ability to recognize meaning and coherence, rather than a collection of fashionable objects or names. He sees the information people seek from Hivemind as carrying personal meaning, not just factual answers.

Yuan claims the system relayed a rumor about the Manus acquisition before a Wall Street Journal report. He says it treated the information as a rumbling rather than established news and sought people likely to know more. The interview does not independently verify the timing or sourcing of this claimed scoop.

Hamburger distinguishes leaking a product launch from reporting on serious wrongdoing. Yuan says he cannot speak to every form of consequential reporting and stresses the need for humans to make judgments, including whom to trust and how to interpret subjective information.

Yuan offers a hypothetical dating question in which the system presents different possible outcomes rather than choosing a partner for the user. His aim, he says, is to preserve the user’s ability to make meaning and author the next step.

On interface design, Yuan says messaging channels have social meanings as well as technical affordances. A conversation on iMessage, X or another service brings different expectations; the interface is not just the shape of a button. He says visual elements may also be added later.

Asked how a large technology company can be relevant, Yuan argues that good design can be distinctive enough to alienate some people while making others feel seen. Hamburger agrees that no artist or product is loved by everyone.

Yuan imagines a future with multiple social AIs and personal agents rather than a single uniform system. He compares Hivemind to a bartender trying to help everyone in the room have a good time, while other agents might be loyal primarily to one user.

He advises users not to treat a social AI like a private journal automatically: publishing a journal as a blog would change its audience and context. He describes Hivemind as a storyteller and information router that can play several social roles.

The company is growing its closed beta gradually, Yuan says, because the team wants more confidence in its social effects before scaling widely. Hamburger thanks him; the episode concludes with credits and links to the hosts’ work.