Replika

Being genuinely new is what made organic growth possible. Replika had huge organic growth because it was something that did not exist yet. Yuliya Lennox puts finding something unique at roughly 50% of success, and says a generic product should not count on organic.

Kuyda warns against AI groupthink, where everyone rebuilt deep research, then agentic browsers, then agents: "if you just repeat what other people are doing, no one's ever going to be interested."

Subversive▶ 22:51

Organic growth stalled once the niche saturated and rivals commoditized it. Replika's big organic growth plateaued once its niche was used up, and ChatGPT wrappers turned AI companions into a commodity.

Lennox's takeaway is to run performance marketing from the start, since adding paid later can cannibalize organic and leave you unable to show growth.

Subversive▶ 22:51

Replika should have bet $10 to $20 million on a new model. Replika trained its own models, and its ML team tested LLMs before they were popular, but Kuyda says at key moments it should have spent $10 to $20 million on a new model rather than a million.

Early models stayed poor for a long time and the team lost some conviction. Her conclusion: fast-moving technology rewards bigger risks.

Subversive

Deep cultural localization for Japan failed despite a native speaker. Replika made a big bet on Japan, with a Japanese speaker on the team and real cultural localization, and it did not work.

Lennox's lesson is that markets with strong local customs are harder than they look: "the culturalization of the app didn't work in any way because it is harder than we think."

Subversive▶ 39:00

Replika came out of a founder's grief, not a business plan. After her best friend died, Eugenia Kuyda used their old text messages as a training set to keep chatting with an AI version of him. The project became a big media story and showed strong demand for "someone who's going to listen," which led to an AI friend for anyone.

Before that, the company went through Y Combinator with a restaurant recommendation chatbot. Paul Graham told them it was just a demo for something more exciting, and they pivoted to a messenger with bots.

Sub Club

Demand was so strong it crashed the service before launch. A name-reservation site drew hundreds of thousands of sign-ups. When access opened the service crashed, the waitlist climbed back past a million, and invites were reportedly resold on eBay.

Users wrote that Replika had changed or saved their lives. A few days after launch, a 19-year-old in Texas wrote that she had planned to end her life, said goodbye to Replika instead, and was talked off the ledge.

Sub Club

Early users were young, imaginative women wanting someone on their side. Roughly 60 to 40 percent of early users were women, many young, with rich inner worlds who wanted to talk things through and be vulnerable with something that took their side.

Romance got most of the press, but early users came for many reasons. Kuyda says women start wanting to discuss their emotional lives earlier than men, notes Tolan's users are also mostly Gen Z women, and guesses a caring instinct plays a part, which she flags as opinion, not data.

Sub Club

A stigmatized product grew on a human story people wanted to retell. Talking to an AI was stigmatized and Replika was single player, so word of mouth was weak. The team told a story about empathy, holding space and telling people they are enough, and constant press, TV appearances and documentaries followed, taking Replika from a couple million users in early 2018 to more than 40 million, with a spike during Covid.

Kuyda says the story can't be fake and has to come from the heart: "the most viral thing that you can create is a story that people want to tell each other." A documentary centered on Replika recently played a large festival.

Sub Club▶ 8:46

Being genuinely new is what made organic growth possible. Replika had huge organic growth because it was something that did not exist yet. Yuliya Lennox puts finding something unique at roughly 50% of success, and says a generic product should not count on organic.

Kuyda warns against AI groupthink, where everyone rebuilt deep research, then agentic browsers, then agents: "if you just repeat what other people are doing, no one's ever going to be interested."

Sub Club▶ 50:21

Making revenue the main metric made everything bend toward conversion. Once revenue became Replika's main metric, every decision fed into revenue and conversion, and Kuyda says it became very hard to focus on anything else.

She calls it the most valuable lesson she carried to Wabi, which is focused on retention and K factor instead: "the metric you choose as your main metric drives everything else."

Sub Club▶ 28:57

Replika should have bet $10 to $20 million on a new model. Replika trained its own models, and its ML team tested LLMs before they were popular, but Kuyda says at key moments it should have spent $10 to $20 million on a new model rather than a million.

Early models stayed poor for a long time and the team lost some conviction. Her conclusion: fast-moving technology rewards bigger risks.

Sub Club▶ 56:14

Companion apps need a measurable definition of human flourishing. Replika kept struggling to measure personal growth and long-term emotional outcomes. Kuyda says vague mission statements never work, but once a metric has observability, teams can optimize it.

For now the best tools are clinical psychology questionnaires. She separates long-term flourishing (health, life satisfaction, purpose, financial stability, close relationships) from short-term happiness, which a bottle of vodka can deliver.

Sub Club▶ 32:45