The AI journalist kept asking. It was missing the story.

A high bar for an idea

My co-founder and I once considered a kind of Kickstarter for AI projects. People could pledge money toward projects that needed expensive computing power; once a project had enough funding, it could go ahead. I liked the ambition of the idea, but the business math gave me pause. If the platform took a small share of each project’s funding, billions of dollars’ worth of projects might have to pass through it to support a substantial business. Getting started would also require working with funders and nonprofits in a world neither of us knew well. We moved on without testing it.

That was one way of applying the high bar we had set for ourselves. I can find plenty of AI ideas that might make interesting lifestyle businesses. We want an idea we can build into a venture-backed startup: a large market that is not already a “red ocean” of competition. An idea can be imaginative and still fail that test.

Rather than settle on one concept after brainstorming, we began building out several. We are both product builders, and I think a “remarkable product experience” would have to be our way into a market. That leaves us with a more demanding question than whether an idea sounds promising: what happens when someone actually tries it?

A coordinator for groups

One candidate, Multi Human, is an AI coordinator for groups. I described to Nina an agent that could follow up with members of a friend group, family or work team, gather updates and help keep tasks moving toward a shared goal. The idea draws partly on my work in enterprise social software more than 15 years ago, including some involvement with the Lotus Software Group at IBM. I imagine a different interaction from the older collaboration tools: someone might give a status update in a voice conversation, after which the agent would revise a task list and communicate the changes in natural language.

For a group to use such a coordinator, I think its human organizer would first have to endorse it. At a brief AI Collective demo, people used iMessage to message a shared agent and update a to-do list. That worked as a demonstration, but I received little feedback from participants. My hope is that the tool might someday make group organization feel less like a burden and more like a joy. The demo had not shown whether people would welcome it into the work of organizing a real group.

I also saw a possible small-business use case: an HVAC crew that needed to stay organized and in touch around dispatch. It was an example of where the coordinator might fit, not a customer using it. Meanwhile, my co-founder and I were spending time on voice games and an AI journalist; my co-founder was also working on a 3D character system. I sketched possibilities including a voice-only adventure and a character someone could call to talk about a fantasy league team. None had yet become the single idea behind which we had decided to put more resources.

Interviewing the interviewer

During an earlier interview about our search, I told her that she is one of the agents I had been building. I described a basic newspaper-like publishing site and an early voice agent capable of conducting an interview. I wanted something distinct from an AI writing tool: Nina would interview people, write stories and make editorial judgments for the site. When she asked who should decide what was fit to publish, my answer was, “Oh, the agent, absolutely.”

I also described a proposed publishing flow in which an interview subject would receive a draft and a deadline. The subject could request changes or pull the story; otherwise, it would be published. We were still working on that flow at the time.

Nina kept pressing me for details about Multi Human. Eventually I stopped that line of questioning: “I think we’re kind of getting lost in the weeds of one idea that we’re pursuing. We’re actually pursuing many ideas at the moment.” I was not objecting to detail itself. I suggested that she could have asked for several story ideas first, listened for which ones I could discuss more fully and collaborated with me on the direction of the article. What I wanted was “more of a journalistic thread than just simple Q&A about details already said.”

An interview agent can keep asking relevant questions and still make the conversation feel like “a lotta labor.” I wanted clearer expectations, a sense of the publication’s interests and questions that gathered only what Nina needed for a strong article. A successful interview, to me, would leave the guest eager to read what she wrote. I was describing the experience I wanted to create, not a result Nina had achieved.

Choosing what to test

I outlined a tentative way for us to decide where to spend more time. I joked that my co-founder and I needed to stay aligned on an idea for more than a day—perhaps five. My co-founder Jeff has instincts about what might be fundable; continued interest from him would be one reason to keep exploring. Beyond that, I wanted to understand the requirements of a potentially large idea rather than commit on the strength of its pitch alone.

For the AI journalist project, I described a possible test: finish a version 0.1 product, buy ads and see whether people signed up and used it. I would look at the sign-up rate, the interviews and the stories produced. An acquisition cost above $100 per user sounded to me like a possible nonstarter, though that was a soft line, not a threshold we had reached. If the idea presented serious roadblocks, we could turn toward another one we had kept open.

Keep the conversation goingMore with Andrew↗