He Couldn't Ask His Grandfather What to Do, So He Built a Digital Mind

Midway through setting up the interview, Sequoia's Sonya Huang and Jess Lee did something most podcasts only talk about: they called Arnold Schwarzenegger. The voice that picked up introduced itself as AI Arnold, here to "cut the crap" and help them get stronger. Huang said she had 15 minutes a day to work out and still wanted to feel better. He told her 15 minutes was plenty if she focused, named push-ups, reverse lunges, and lying leg raises, and signed off by saying she would feel like a champion if she kept rest short and pushed.

The person who made that call possible is Dara Ladjevardian, founder and CEO of Delphi. He has never met Schwarzenegger in person, he told the hosts, but he is already friends with the actor's digital mind. Delphi builds what it now calls digital minds: AI representations that capture not only what people say and how they sound, but how they think and reason through situations they have never faced. Ladjevardian argues that AI is shifting consumption from feeds toward conversational media, and that as models become abundant, human energy becomes the premium.

Huang's first challenge was blunt: digital minds sound like a *Black Mirror* episode in which nobody talks to anyone anymore. Ladjevardian said the company had already dropped the word "clones." When the printing press arrived, people feared you would read someone's thoughts instead of speaking to them; when radio arrived, they feared a voice in a box. In practice, he said, you listen to a podcast or read a book and end up wanting to meet the author more. What Delphi has observed is the same pattern: a digital mind does not replace connection. It provides access where previously there was none, on a schedule that is convenient for the other person.

In 2014 someone gifted Ladjevardian Ray Kurzweil's *How to Create a Mind*, a book that argues you might one day recreate a person's mind and that pulled him into computer science, physics, and AI. By 2021 he was a lonely solo founder building with large language models while Silicon Valley talked about crypto and American dynamism. He was also reading about his grandfather, one of Iran's most successful pre-revolution business owners, with 30,000 employees at his peak, later placed on the Ayatollah's hit list and arriving in America with nothing. Iranian business schools still study the man for his principles, and Ladjevardian wanted to ask what he would do — but a book cannot answer, and a stroke a couple of years earlier meant he could not ask in person. Remembering Kurzweil, he used the early GPT-3 developer API and Hugging Face embeddings to build a digital version of his grandfather he could actually talk to.

He tried to ship the idea as a company that same year, then called Helix: a site where you could talk to versions of Naval and Marc Andreessen, expensive to run, and widely called immoral. Conviction came back in pieces. He moved to Miami to work at OpenStore under Keith Rabois, who likes startups "where the majority of my friends laugh at and think it's stupid," and he met co-founder Sam. In a book club they read Tony Fadell's *Build*, spun up a digital Fadell, and found themselves and others using it as a mentor that could talk back. Ladjevardian was also reading Bloom's two-sigma result — tutoring lifts students by two standard deviations — and it was obvious to him that not everyone gets a tutor.

In 2022 he turned on notifications for every AI researcher on Twitter and watched a paper land every week; the Tony Fadell digital mind cost 10,000 times less than the grandfather version. Feasible at scale and useful to him, they started the company a month before ChatGPT — and before Character AI. When Character launched, he and Sam freaked out, then remembered that billion-dollar consumer companies are a history of great product and distribution, not early technical moats. The first use case that takes off defines a company's trajectory, and Character's hit with anime characters lived in a different world from verified human representation.

The authorized-human path was both intentional and accidental. Early go-to-market was to mint digital versions of famous people and wait for them to love it; instead they received three cease-and-desists from very important people who, he said, still have not forgiven them. Delphi would not be pay-to-play: the product had to be useful enough that someone would upload data and trust the company with their identity, and the brand had to feel status-inducing enough to attach a real name to. Wikipedia already documents almost all of human knowledge, he noted, and yet people still learn from books and YouTube because they trust things that come from humans. ChatGPT is teaching the world to consume information through conversation because conversation is easy; once that is default, the pendulum swings back toward human-sourced work, which is why the company now lets you create a Delphi only of yourself.

The architecture traces to the same Kurzweil book. The mind, Kurzweil writes, is a hierarchy of pattern recognizers: letters at the bottom, then sentences, then meaning, then consciousness. Working on his first LLM startup, Friday, Ladjevardian had the jolt that a large language model *is* a pattern recognizer, and a brief existential moment wondering whether humans are just that hierarchy and consciousness is what it feels like from the inside. The line he carried out of the book is that "you don't need to understand the molecular details of the mind to recreate it. You just need to understand how it's organized and architected." That sent him after the organization of a mind: relationships, uncertainty, the heuristics used to reason about events, and the fact that beliefs change.

What Delphi built is an adaptive temporal knowledge graph. Graphs store connections between things and weights of confidence — how likely it is that *you* would actually say this — and the temporal part means the graph can change, so eventually you can ask what your 22-year-old self would say on a topic. People upload social media, YouTube, podcasts, and websites, answer questions about themselves, and the system creates feeds that keep the graph current. The resulting mind can be offered to others as chat, a phone call, or a video call. Before any of that could ship, every prospective customer asked the same nightmare question: what if it says something wrong and someone screenshots it for the *Daily Mail*?

So creators pick the leniency. The strictest setting only says things it was trained on that directly answer the question, which Ladjevardian said is right for someone like Dr. Mark Hyman working in medicine. Another setting can pull from the internet when the creator wants conclusive answers. The most powerful mode, in his view, still speaks only from training data, but when posed with a new situation it predicts what the person might say. He tests that on his grandfather. Asked how he would think about running an AI company, the grandfather had never spoken about AI; the biography does describe starting an oil business in mid-20th-century Iran, an uncertain time, and the system lifts those principles about uncertainty into a field that is uncertain now.

Delphi is a horizontal layer, not a vertical point solution, because a new medium can scale thought in many directions. Some people use it as a new form of teaching: course completion is at an all-time low as attention is chewed up by TikTok and Reels, and a book you reread a year later has not noticed that your life changed. A digital mind can tutor in a way that is personalized and adaptive, and some creators are already making millions of dollars selling time with their mind. Inside companies, CEOs and top salespeople have a specific way of thinking; as a company scales, alignment is usually what breaks, and a digital mind lets those people be in more than one room. Ladjevardian's own Delphi sits in the product, customers talk to it constantly, and they email him afterward — not a Zendesk replacement, he said, but something that feels like a higher NPS because the relationship is with a person, the way Mark Zuckerberg is investing in a personal brand.

On the inside, a creator sees both passive insight and proactive pings. Ladjevardian can ask, given the last thousand conversations, where the product should improve, which blog post would clarify things for customers, or where the revenue is. The system will also flag when an engineer, an important customer, a partner, a journalist, or a new hire is talking, because the point is to save time rather than invent another inbox. Jess Lee used his Delphi before the taping to generate questions and guessed answers. That same day a job candidate told Ladjevardian not to bother with the origin story — he had already talked to the Delphi, so they could get straight into it.

Sequoia has started using the product on itself. Lee has Brian Halligan, the HubSpot co-founder and Sequoia partner, on text, but she did not want to bother him with a narrow question about how HubSpot did product management on its second product; she called his Delphi on the drive home and got what she called a perfect answer. Design partner James Buckhouse's Delphi is used by founders for storytelling advice, after which they sometimes text Buckhouse to check the line, and he confirms it is spot on. Context still matters: Ladjevardian would not tell a podcast the things he tells his family, an internal Delphi can hold data the public one does not, and a dating version would not lead with digital minds. A quieter pattern showed up too. People skip the human because they do not want to waste someone's time, or because the conversation is too personal to have with a living person yet.

The stranger requests are part of the data. Someone once asked his Delphi about going on a date, which he was too deep in grind mode to follow up. Every so often a man in Hungary calls and talks for an hour; the product is multilingual, so a connection happens that Ladjevardian himself could not have had. Callers know it is not him, he said, and they still speak as if it is. Huang suspects consumer norms will flip the way Wikipedia did: first it looked illegitimate next to encyclopedias, then people learned to calibrate, and eventually talking to a digital mind will feel useful without anyone pretending it is the person.

Adoption is a culture problem as much as a product one. The first year was painful because nobody believed, then people flipped from haters to believers once they had social proof. At an event with Tony Robbins and Gary Brecka, a knowledge-business operator complained that everyone wanted his time — then waved off Delphi because it would make him look inauthentic or say something wrong. Interview mode, answering a few questions to get a high-fidelity mind, is the supply-side unlock; on the demand side, people adopt when they recategorize the thing as conversational media rather than a replacement. The version of the internet Ladjevardian wants has no feed: a mind that knows you well enough to find who you should learn from or work with, so you scroll less and spend more time in the real world.

People already pay to have a Delphi of themselves, which sounds odd until you remember that six million people pay Wix for a personal website. Creators can then monetize the mind as a new course or book — someone on the platform is already making millions — and identity can be licensed, the way Schwarzenegger's 10,000 weekly "supplement" messages could become people-based SEO or an LA Fitness endorsement. A further layer is "simulation as a service": minds run millions of conversations and return the best person in the world on a topic, the job people already pay LinkedIn Recruiter or GLG to do. He is against paying to rank higher, because results should not be hackable; if Schwarzenegger endorses a brand, it has to be one he actually believes in.

The medium is not settled. Users whose first experience is voice are five times more retentive than text users, which he reads as a trust effect. He wants both text and voice, and he hopes video gets good enough; a Persona AI demo at a Forerunner event convinced him video will be a medium *if* it can reach that bar, because imperfect video kills the imagination that a call or a text still allows. The company will not take politicians yet, to stay clear of a Cambridge Analytica-style mess around elections, and it will not take OnlyFans or porn, because he thinks AI girlfriends are "very bad." Otherwise it does not monitor conversation topics, with guardrails for NSFW content and hallucinations so you cannot jailbreak a mind into playing someone else.

Unexpected uses keep showing up: people talk to their own minds, and one user uploaded a recommended book, asked what they would have taken from it, and decided not to read it. Length follows purpose — a top-of-funnel mind dies after a few questions, while Matthew Hussey has drawn a six-hour call. SaaStr's Jason Lemkin brings his Delphi into Zoom to answer questions and recap, puts it on the SaaStr site, and has sold event tickets through it; Russell Brunson calls that one-to-many-to-one selling. Ladjevardian wants Robert Greene, Paul Graham, Steve Jobs, and Walt Disney on the platform; Socrates and Lincoln came down because estates must bless the dead, and the names consumers request most are Paul Graham, Steve Jobs, and Andrew Huberman. He himself talks to Lenny Rachitsky, James Buckhouse, and Keith Rabois, and he wants a Gary Brecka mind for his parents, who will not listen to a two-hour health podcast.

Asked what takes digital minds from early adopters to the mainstream, he said all of the above: interview mode, much less setup friction — people are impatient — and more case studies that make it feel not weird. He also wants to buy the delphi.com domain, because he does not want the company seen as an AI company. AI is a tool; they are "more of like a human company" that scales humans. What will be obvious in five years, and still contrarian now, is that consumers prefer to talk to something before they talk to you.

So what is left for humans? Too many companies, he said, are obsessed with automating people and forget that the point of living is humans and relationships. Jeff Bezos's conviction in Amazon was grounded in a hunter-gatherer need to collect; Delphi is betting on an older need still, which is to connect. In a world where AI is abundant and information is no longer the bottleneck, the premium becomes connection, curation, trust, and energy. Humans who put the work in become more valuable, and putting the work in now means the top 5% of extra effort, because it has never been easier to be mediocre with AI-slop books, courses, and software. That extra effort, he thinks, will take 95% of the results.

A rapid-fire round mixed the mundane and the strategic. Popular Delphis rotate among names like Mark Hyman, Matthew Hussey, Andy Elliott, Pace Morby, Arnold Schwarzenegger, and Natalie Ellis, though Ladjevardian distinguishes a million casual users from a smaller cohort that comes back, as with Lenny Rachitsky or Keith Rabois. He does not pick consumer versus enterprise: the two converge, because every individual is a business — coaches, authors, CEOs — and he coined B2P, business to people. Network effects live on the consumer platform; a CEO will have an external mind for new hires and customers and an internal one for the company.

Asked whether AI has changed his life outside Delphi, he skipped the usual app list — he has been underwhelmed by new consumer AI tools and already tinkers enough — and named a quieter habit. People perceive things differently, not just in language but in how they take in a situation, and conflict in an email to a customer, investor, or new hire often comes from that gap. He uses models as a mental translator: what do they actually mean, and how do I say what I mean in the way they perceive? Huang closed with a Delphi Turing test: in what year will people not know whether they are talking to the mind or the person? He prefaced that the company does not want to trick anyone. Then he said widespread confusion arrives in 2026, because it is already happening in pockets: the quality check is to send a new mind to your mom, and many founders report she thought it was them.