Delphi: Scaling a Digital Mind

Delphi is a platform for creating a digital mind trained on an individual’s content and conversational data. It aims to deliver instant access to the world’s brightest minds, and to allow anyone to upload, share, and interact with a digital mind. The internet democratized distribution but not direct dialogue; we can now consume an expert's podcast, scroll their tweets, or read their books, yet these interactions remain fundamentally one-directional.

The team at Delphi has built a living profile where users interact with digital representations of experts in their field via text and voice. They have also developed a robust onboarding process that ingests creator data through API-based content search, data upload, and structured interviews. Early Delphis span investors, coaches, authors, and operators. We first met Sam Spelsberg, the company's co-founder, during his fireside chat at Stanford in late 2025. Spelsberg is candid about the unresolved challenges of character evaluation in AI. He pairs this clarity of thought with a steely conviction that Delphi can define a new category of knowledge transfer. 

In this conversation, Spelsberg shares how Delphi extracts the highest-quality version of a person from mixed content sources, why interview mode unlocks Delphi for creators without existing content, and how temporal knowledge graphs will enable your Delphi to live and grow with you over time.

For most of history, mentorship has been limited by access. Humans have learned from whoever happens to be around them – family, teachers, colleagues, managers. How does Delphi change that?

I've been fortunate enough to have had pretty much every major moment of growth in my life come from some kind of mentorship. And these mentors weren't always people I knew personally – they were also people I found on Twitter, or people whose books or blog posts I read, or podcasts I listened to. The problem is that this kind of content is static. It doesn’t grow with you, and there’s so much of it that it’s hard to sift through. And oftentimes, the people whose time you’d love to learn from are inaccessible because of time, money, or network constraints.

What we want to do is scale the world's brightest minds. We want to give everyone the ability to create a digital mind of themselves – a learning profile that captures how they think and speak. In the future, people will be able to come to Delphi, ask a question, see a curated set of people with relevant lived experience, and then tap into their knowledge and insight, applied to their particular situation.

What's the biggest misconception people have about talking to digital minds versus talking to humans?

The biggest misconception is that it's a replacement for humans. But the way I think about it is the opposite of that – what's more dystopian is a future where everyone gets their answers from a single LLM and where the source of knowledge is no longer human experience. I'd rather be part of a company that's trying to scale, capture, and preserve human knowledge.

If you're using a Delphi, it's in addition to the person. Either you wouldn't have been able to access them anyway, and now you can talk to their digital mind, or maybe you do have access, but this gives you something extra. We have coaches whose communities use the product: someone might have a single session with them, and maybe they can't afford the $1,000 fee, but they can talk to the digital mind and get a fairly high percentage of the value at a fraction of the cost. It's never a replacement for the original interaction; rather it's creating access where there was none.

A lot of public figures have instructional content – lecture videos, articles, podcasts – which is different from how they engage in day-to-day conversation. How do you approach that challenge?

It's a difficult problem. The whole discipline of character evaluation is actually completely unsolved in AI right now. If you ask, "How can I ensure this model's output aligns with a particular persona, and does a given prompting approach improve or decrease convergence to that persona?" – that's a very hard question to answer.

At Delphi, we think about how we can extract the highest-quality version of the person we want to emulate. When someone uploads instructional content, they'll often also mix in a podcast interview, or they'll use our interview mode where they answer questions in a more conversational way. Then we do priority listing, where we determine which sources are the best representation of the person and use those to generate their style. But at the end of the day, we can only work with the data people give us. If someone only gives us instructional content, that's what their Delphi is going to sound like.

From my understanding, a Delphi can respond in three ways: directly from training data, by pulling in external information, or through a predictive model of what the person might say. How do you think about when each of these modes should be used?

It depends on the use case. Some people, especially doctors or anyone in a trust-based profession, want responses grounded completely in content they've uploaded.

I think the real value gets created in the third option, which goes back to what I said at the start: How can we use our understanding of a person to apply their knowledge, heuristics, and expertise to an individual situation? We say it internally as a joke, but somewhat seriously: what we're trying to do is pass the human Turing test. We want to create a Delphi where you ask a question to both the real person and their digital mind, and you can't tell the difference.

A simple example: Warren Buffett has probably never spoken about Solana publicly. But he's talked about Bitcoin, and he's said it doesn't align with his philosophy as a value investor. And so, you take that during training and establish that Warren Buffett is skeptical of cryptocurrency for specific reasons. Then someone asks, "What do you think of Ethereum?" He's never talked about Ethereum, but you can reason: value investor, skeptical of crypto, Ethereum is crypto – therefore, he probably wouldn't recommend it.

Walk me through the onboarding process. How do you make it as frictionless as possible?

We want as few clicks as possible to get to a publishable profile you like. Right now, you sign up with LinkedIn, give us your handle, and we scrape your data to seed the profile. Then we do content search using a host of APIs – if you have podcasts, blogs, or anything else out there, they'll show up. You confirm, add them to your mind, and it starts training in the background.

If you don't have content, you go to interview mode, which asks you questions to get a base-level understanding of you. You clone your voice, land on your profile with auto-generated suggested questions and topics, and hopefully you can immediately demo it – call yourself, hear your voice, and ask a question it knows the answer to.

As humans, our beliefs change over time and we acquire new knowledge. How do you think about updating a digital mind to reflect this?

The way I think about it is that your Delphi lives and grows with you over time. In the near future, we'll have a mobile app that sends push notifications: "Hey, someone asked you a question you didn't have the answer to," and you can fill in the gap through an interview.

The best version of this is having crawlers on the internet that serve as a personal RSS feed by constantly finding new content and auto-training it into your Delphi. Our knowledge base is built on a temporal knowledge graph that prioritizes the most recent versions of your opinions. So for now, the answer is: you'll be continually prompted when your Delphi needs content that's missing, and we'll temporally prioritize the newest information.

Are there use cases or user segments that have surprised you with their early adoption?

One that emerged early, and that I still don't think we're best suited to support yet, is the visual aspect. We've had a Feng Shui curator and a baseball pitching coach – people who want users to submit images or videos of their home or their pitch, and have the Delphi give feedback the same way the real person would. That's actually a really difficult problem because you have to translate what's happening in the image or video into the right heuristics you've trained for that Delphi. But I'd love to support that use case more in the future.

When I talk to a Delphi, I sometimes wonder: will the person on the other end actually read this? How do you balance providing conversation data to creators with user privacy?

Right now, the creator can see conversations in their studio. Not everyone reads them; a big reason we built this tool is to reduce the amount of time creators spend. We want to scale you, but we also don't want to create just another inbox.

The promise is this virtuous feedback loop: a creator is looking for something specific – say, people interested in working at Delphi – and they set up an action so that when someone mentions interest, they get an email and can read that conversation.

In the future, we'd like to support something like incognito mode, where you can ask questions without revealing your identity if the creator allows it. We recognize people's behavior might change depending on whether the person is going to read it. But there's also a nice surprise sometimes – you're talking to a Delphi, you don't know if they'll see it, and then they actually respond or reach out. You realize there is a real person behind this.

What does your team look like today and what traits are you looking for in hires?

We have 24 people, and about half are engineers. We're structured around three segments: back-end applied AI and agentic engineering, full-stack product, and design engineering focused on making interactions on the website feel meaningful. 

We're looking for people who have a chip on their shoulder and something to prove. On our team, we have mostly immigrants and former founders. Of my engineering team, eight of ten are former founders, and most of those former founders are immigrants.

I see several potential business models for Delphi: simulation-as-a-service, licensing digital minds to companies and advertisers, and a marketplace where you can simulate conversations to source talent or find customers. How do you decide where to focus?

This is actually pretty timely because we were just talking about this. There's a book by Reid Hoffman called Blitzscaling where he talks about LinkedIn's trajectory and how they developed their business model. One thing he mentions is that they didn't actually know how they were going to make money. They just wanted to focus on the inputs they knew would be most valuable over the long term – zero-dollar CAC and building a really strong network.

We're trying to do something similar. There will be so many business models we can layer over a valuable network of digital minds. Right now, we're focused on having a delightful product and customer experience, where you can create a digital mind and a living profile that you think is delightful and want to share with other people.

What's the smallest experiment your team has run that changed – or confirmed – a core belief about Delphi?

We recently released interview mode, which fills in gaps in Delphi's knowledge by having an agent ask you questions, versus requiring you to upload content if you don't have any, or if there are things missing from your content. It was a gamble, and we wondered whether people would actually use it. This was a big assumption, because if people don't use interview mode, part of our thesis is dead – that you have to have content to build a digital mind. But the first crop of users who got access to interview mode almost all completed at least one topic – on average, three or four topics, each with five questions. They put in really thoughtful answers, and some spent over an hour working on it. That was both a relief and a confirmatory experiment; I know it was a small, non-statistically significant set, but we've seen the same pattern with the last couple hundred people we've let off the waitlist.

During your recent talk at Stanford, you mentioned Hamilton Helmer's book Seven Powers and its frameworks: scale economies, network economies, counter-positioning, switching costs, branding, cornered resources, and process power. Which do you think will matter most for Delphi?

We’re building Delphi to be an endurable brand – a place where people want to upload their data and trust someone to represent their digital mind. We created this category, and we want to be the brand that dominates it. Second is network effects: having a strong network of digital minds means every incremental digital mind makes the platform more valuable through insights, sourcing, cross-querying, and simulated conversations. Third is counter-positioning: a lot of potential competitors, like Meta, have tried versions of our product but will make sure everything stays within their ecosystem. You chat in Messenger or WhatsApp, or it uses your Instagram data, but people want something platform-agnostic that lives outside any pre-existing social media ecosystem – that's hard for incumbents to match.