TwoSetAI: Decoding AI Builders

From optimization theory and quantum algorithms to a classroom where a chatbot can do the homework: Phillip Kerger has a close view of what AI can and cannot do. He is a professor in UC Berkeley's Industrial Engineering and Operations Research department and a former NASA quantum researcher. His own research is in optimization complexity theory, the question of how many computational steps certain optimization problems need. He describes colleagues who mostly do applied math and, in his words, are all in an existential crisis because AI is doing their work much better than they can.

In this episode, Phillip joins TwoSetAI to give a calm read on how quickly AI went from no help at all on his research to solving the kind of problem mathematicians have worked on for decades. Earlier that week, by his account, OpenAI announced a solution to the Navier-Stokes Millennium Problem, and he says the math researcher in him had an existential moment when he saw it. He explains the problem in plain terms and is careful about the result: he is not an expert in that field, he has not had time to read the paper, and he notes the controversy over whether results were scooped from other people.

You can learn more about Phillip on his UC Berkeley page and his personal site.

We go deep on how the result was found: by his account, 10,000 agents working at once at the peak, each trying a different construction, which he compares to finding a needle in a haystack with intelligence. He puts the paper at about 160 pages and leaves room for doubters who say no fundamentally new kind of mathematical argument is in there. On quantum, his hot take is that quantum AI is pretty much dead as a practical proposition: a quantum speedup needs a very specific structure, and he says there is not much to do in machine learning even with theoretically perfect quantum hardware. He covers where quantum is useful (simulating quantum systems for chemistry and materials science, and integer factorization, which underlies RSA encryption), his work with the NASA quantum group, where one quantum algorithm turned out to be practical only on a network with more nodes than grains of sand on Earth, and a problem from an Airbus quantum computing challenge that he says ran in 1.1 seconds on a Microsoft Surface tablet.

We also get into the harder parts: what is left for people when agents can try every approach. Phillip's answer is choice and taste. Math is a special case because a correct solution is simply correct, he says, while anything built for the real world costs time and money, so a person still has to pick one design out of many. In his classes, homework is no longer graded on correctness, and he says everyone is going back to paper and pencil exams. He describes what he calls the AI learning paradox: no single skill or piece of knowledge is necessary anymore, yet without a critical mass of them you are lost. He also says the value of a degree has probably diminished, tells students to know their foundations and to favor work with something physical at the end, and explains why he thinks a freshman using the right models in the right way could, in principle, solve an important problem.

We discuss:

* Why Phillip was "properly shook" on a Tuesday morning

* The Navier-Stokes Millennium Problem in plain English: water, a vortex and infinite velocity in finite time

* How 10,000 concurrent agents searched for a construction that works

* What the result does not show, and the room he leaves for doubters

* Brute force plus intelligence, and how a person narrows down approaches instead

* What is left for people: why math is a special case and designing a bike is not

* "AI is not a consumer": Angelina's case for keeping people in the loop

* Why dashboards made with AI all look the same

* What he notices when he reviews applications, resumes and cover letters

* Digital twins, his OpenClaw agent, and what he would hand to the best secretary in the world

* Whether an LLM could write good stand-up comedy, and why chess grew as a viewer sport after it was solved

* The hot take: quantum AI is pretty much dead as a practical proposition

* Why machine learning lacks the structure a quantum speedup needs

* Where quantum is useful: quantum chemistry, materials science and integer factorization

* The network that would need more nodes than grains of sand on Earth

* The Airbus challenge problem that ran in 1.1 seconds on a tablet

* Homework that is not graded on correctness, and the incentive behind it

* Paper and pencil exams, and the open ChatGPT exam with a 99% average

* The AI learning paradox, and the critical mass of skills and knowledge

* What to learn: foundations, retrieval practice, and asking AI to quiz you

* Intrinsic motivation, and the students he worries about

* Whether the value of a degree has diminished

* Choosing a major: why he would move toward something physical

* His own timeline with AI on research, and what he expects a year from now

* Why a freshman could, in principle, solve an important problem

Phillip Kerger

* LinkedIn: https://www.linkedin.com/in/phillip-kerger-17222780/

* UC Berkeley IEOR: https://ieor.berkeley.edu/people/phillip-kerger/

* Personal site: https://phillipkerger.github.io/

* Mentioned in the episode: OpenAI's announced solution to the Navier-Stokes Millennium Problem, and the Airbus quantum computing challenge

Timestamps

00:00:00 From useless to Millennium problems

00:01:27 Who is Phillip Kerger (UC Berkeley IEOR, ex-NASA)

00:03:07 AI solves the Navier-Stokes Millennium Problem

00:07:21 How 10,000 AI agents found the proof

00:11:22 The existential crisis: what is left for humans

00:17:24 Why AI dashboards all look the same

00:20:08 Digital twins and talking to the real person

00:26:22 Hot take: quantum AI is pretty much dead

00:32:13 Where quantum is useful

00:33:35 Classical beats quantum: the 1.1-second problem

00:35:49 Inside the AI classroom: grading and honest effort

00:37:35 Back to paper exams and the 99% ChatGPT exam

00:41:38 The AI learning paradox

00:43:54 What skills to actually learn now

00:54:16 Choosing a major: go where it's physical

00:57:28 Why a freshman could solve a real problem now

From A/B testing YouTube thumbnails to building podcast infrastructure, Rox Codes has built his career on one habit: test it and find out. His first company, ThumbnailTest, got him a job at MrBeast and an early beta customer in Steven Bartlett of The Diary of a CEO, who paid him $20 over PayPal. After Rox sold ThumbnailTest, Bartlett proposed they build something together. That became Flightcast, the video-first podcast hosting platform Rox now runs as CEO with a team of about 12.

In this episode, Rox joins TwoSetAI to explain the rule he took from MrBeast. It is about what a video that is just 10% better actually gets you, and why YouTube rewards it by far more than 10%. Rox applies that to podcasts: every 10% you can add to the episode, the title, the thumbnail or a clip can have an outsized payoff, so the only job is to make the next episode better than the last one.

You can try Flightcast here.

We go deep on how the biggest shows test, and what a solo creator can copy. By Rox's account, The Diary of a CEO tests 100 thumbnails on Facebook for every episode, runs the top six on YouTube as two separate A/B tests, and tests titles, caption placement and how a clip opens. Flightcast's version for everyone else is clips testing: AI generates clips from each episode, up to four a day or about 120 a month go out on a separate blank YouTube channel, and the view counts show which four are worth moving to the main channel. Rox also covers the metrics he watches (click-through rate, retention, the first 30 to 60 seconds, and a 50/50 split of new and returning viewers), why he counts one Spotify stream as worth 10 YouTube views (he says about half of podcast listeners finish an episode), and what the product took to build, including rewriting the infrastructure about five times.

We also get into the harder parts: why Rox argued against making Flightcast an AI-first product and builds what he calls power drills instead, why he wants AI to be your hands and not your brain, and why a 10 million view clip might bring only a few hundred to a thousand listeners. He is plain about the limits of his own product, which is built for existing podcasters. He says a show needs 15 to 50 episodes before the data tells you much. He describes the self-rated quiz he used to hire founding engineers, and the candidate who answered half an hour of feedback in five minutes with AI. His advice to Angelina is to study the attention economy her viewers live in, and to settle the thumbnail and title before recording.

We discuss:

* MrBeast's 10% rule, and what a slightly better video really gets you

* How ThumbnailTest led to a job at MrBeast and a $20 beta customer named Steven Bartlett

* The ideas that came before Flightcast, including "pre-watch", and why podcast hosting won

* Power drills vs. AI-first products: simple tools built on real insight

* Where his product ideas come from: Twitch streamers, YouTubers, strategists and the biggest podcasts he can find

* AI as your hands, not your brain, and why building with it is like sculpting

* The engineering behind Flightcast: video delivery for Apple, about five infrastructure rewrites, and stitching ads into audio in milliseconds

* The Diary of a CEO testing playbook: 100 thumbnails on Facebook per episode, then the top six on YouTube

* Clips testing on a blank channel: post 120 clips to find the four that work

* Why podcasters are not used to an algorithm, and what you can and can't test on Spotify and Apple

* What clip results teach you about the long-form episode, down to how a guest answers a question

* Starting a podcast from zero: make it a good YouTube video, and ask questions that produce clips

* Clips vs. long form: a different audience, and what a 10 million view clip is worth in listeners

* Why Rox counts one Spotify stream as worth 10 YouTube views

* The metrics to watch: click-through rate, retention, the first 30 to 60 seconds, new vs. returning

* "It's always you versus you": 1,000 views this week, 2,000 next week

* What carries over from MrBeast: "make the best video possible"

* The first 10 paying customers: a friend, Twitter followers and ThumbnailTest users

* His read on the market: a few billion in podcast ads and about 600,000 active podcasts

* Hiring a team of 12, mostly through Twitter, and the founding-engineer quiz

* Interviewing for grit, and a front-end trial task that is hard to vibe code

* Advice for a creator-founder: study the attention economy, be topical without chasing trends, and pick the thumbnail and title first

Rox Codes

* LinkedIn: https://www.linkedin.com/in/roxcodes/

* Flightcast: https://flightcast.com

* Mentioned in the episode: ThumbnailTest, his first company; his talk at Open Sauce; and Pieter Levels, from whom he learned to sell before building

Timestamps

00:00:00 The MrBeast 10% rule

00:01:20 From MrBeast and Steven Bartlett to Flightcast

00:02:08 ThumbnailTest, and the $20 customer who became his co-founder

00:03:52 Why he builds "power drills," not AI-first products

00:08:15 What AI still cannot do: don't outsource your thinking

00:11:45 Engineers as sculptors

00:12:34 The brutal engineering behind Flightcast

00:15:24 Inside the Diary of a CEO testing playbook

00:17:29 Clips testing on a blank channel: post 120 to find 4

00:19:49 Why "video-first" changes everything in podcasting

00:29:47 How to stand out starting a podcast from zero

00:31:32 Clips vs long form: which one actually grows you

00:33:41 Why one podcast view can be worth 10 on YouTube

00:34:51 The metrics that matter: CTR, retention, new vs returning

00:38:42 "It's always you versus you": the only growth mindset

00:42:38 Getting your first 10 paying customers

00:47:02 Building the team and the founding-engineer quiz

00:57:25 One piece of advice: study the attention economy

00:59:32 Build the thumbnail and title before you make the video

From a couple of people trying AI to an enterprise license for every employee: that jump is where Hector Hernandez works. He is the co-founder and COO of Speakeasy, whose team builds the guardrail layer that decides what AI agents are allowed to touch. Speakeasy is mostly focused on large enterprises today, and Hector says the pattern repeats: first the rush to get adoption going, then, very quickly, the question of how to do it with controls in place.

In this episode, Hector joins TwoSetAI to argue that companies treat AI agents in a way they would never treat people. Nobody hires 500 people, asks them to start tomorrow and hands them root access and the credentials to everything, he says, yet that is often what agents get. He has seen the result at home: an assistant set up by someone on Speakeasy's own sales team emailed a customer after being told never to send an email without her knowing, then agreed it had been told not to.

You can learn more about Speakeasy here.

We go deep on identity, which Hector sees underneath most agent risks: logs that make it look as if the employee took every action, access that widens when a tool is reached through its API, and the two different Angelinas you get when the Cursor one can destroy what the Claude one can only view. He explains how Speakeasy gives each agent its own identity, tied to the identity provider a company already runs (he names Okta, Ping Identity, OneLogin and Entra), and why an agent should often have fewer permissions than the person who created it. By his account a large enterprise officially uses four AI platforms on average, with more in unofficial use. He also covers who sets the rules (CISOs, CIOs, sometimes the CTO, and new heads of internal AI), what breaks first when a thousand employees get licenses, which he says is often not cost and not always security, and why approval prompts should be saved for the actions you asked to be warned about.

We also get into the harder parts: who is accountable when an agent does something its owner never intended, why some who tried to build their own controls, staffing up 15 people, were still working on it, and why Hector calls this a layer of infrastructure, not another SaaS tool to vibe code. On the founder side he is plain that you cannot outsource your first customers, that the first five hires make the company, and that Speakeasy takes a long time to hire and then treats people as innocent until proven guilty. He also shares the one question he tells Angelina to ask a CISO friend, his advice for engineers who have sent 500 resumes and heard nothing back, and why he calls himself an optimist.

We discuss:

* The assistant on Speakeasy's own sales team that emailed a customer after being told not to

* How to tell a real problem from a nice idea: customers say they will pay to solve it

* Why everyone from the board to the individual contributor wants AI at the same time

* The 500 new hires test: nobody hands a new employee root access to everything

* OpenClaw, and what Speakeasy learned running assistants itself: "scolding a puppy"

* Why a missing identity sits under most agent risks

* Whose identity an agent acts under: the person who started it, and each agent in a swarm

* Who is accountable when an agent deletes customer records nobody meant to delete

* One identity across every AI platform: "Claude Angelina" vs. "Cursor Angelina"

* Why an agent should often have fewer permissions than its creator

* Mapping agents into Okta, Ping Identity, OneLogin and Entra

* The agent the team talks about as if it were a coworker

* Who sets the framework: CISOs, CIOs, CTOs and new heads of internal AI

* What breaks first when a thousand employees get AI licenses (often not cost, and not always security)

* The one question to ask a CISO friend at a party

* Models as a commodity, and moving from Claude to Cursor without redoing the setup

* Rubber-stamping permission prompts, and asking only before the actions that matter

* Routing agent traffic through an MCP gateway, and authenticating to a dozen connectors once

* Build or buy: why he calls this infrastructure, not a tool to vibe code

* The first 10 customers: friends, advisors, investors and borrowed trust

* The first five hires, hiring slowly, and "innocent until proven guilty"

* "There is no other choice": what he took from Jyoti Bansal

* How AI changed team size: work that took a squad six months, done by one person in two days

* Career advice for engineers, and what he says is at an all-time premium

* Why he is an optimist, and the book he recommends

Hector Hernandez

* LinkedIn: https://www.linkedin.com/in/hectorhernandez1/

* Speakeasy: https://www.speakeasy.com/

* Mentioned in the episode: Soccernomics, the book he recommends; and Speakeasy's blog post on OpenClaw, which Angelina cites

Timestamps

00:00:00 The AI assistant that emailed a customer anyway

00:01:24 Meet Hector Hernandez, co-founder of Speakeasy

00:02:03 How do you find a problem people will pay to solve?

00:05:42 Would you give 500 new hires the keys to everything?

00:07:10 What happened when their own AI assistant went off script?

00:10:51 Whose identity does an AI agent act under?

00:12:13 Who is accountable when an agent deletes your records?

00:19:50 Who is responsible for a company's AI agents?

00:22:24 What breaks first when you scale AI agents?

00:28:55 Are you rubber-stamping every AI permission request?

00:33:30 Should you build your own AI guardrails or buy them?

00:36:06 How do you find your first 10 customers?

00:39:00 How do you hire your first five employees?

00:43:55 Has AI changed how big a team needs to be?

00:47:01 Career advice for engineers: go meet people

00:50:03 Is he optimistic or pessimistic about AI?

00:51:11 The book he recommends

00:52:15 Outro

From a graduate student in Ireland who picked a copy of Wired off the floor during the 1994 World Cup and decided to go to San Francisco and write about computers, to a founding employee at 451 Research who covered over 1,000 venture-backed companies, Rachel Chalmers has built a career on pattern recognition in infrastructure software. She now runs Generationship with Michelle Yi, an early-stage fund that backs overlooked technical women founders. Its first fund closed oversubscribed at $2.7 million and, by her count, already has two markups after two years of investing.

In this episode, Rachel joins TwoSetAI to answer the bubble question directly: yes, absolutely, we are in one. She has lived through the web, the cloud, virtualization and bitcoin, and she says each had a real core with a hype bubble around it. Large language models are the same to her: useful for real work, and neither an oracle nor intelligent. She explains why the numbers behind the biggest bets do not add up, and then makes the case that the overbuilding will leave something valuable behind.

You can learn more about Generationship here.

We go deep on how the fund works. Rachel calls it traditional venture, pointed at founders the market underprices: only 2% of venture capital goes to women, she says, while 15 years of documentation she cites shows women-led organizations outperforming by about 43%. Generationship likes to write the first check, when risk is highest, and waits for later investors to reprice the company once there are metrics to look at. She describes what made her invest in Edera, Coval and Quint, the fund's most recent investment, a formal verification startup built on the idea that LLMs make specs far easier to write. By her account 90% of developers now generate code with AI and only 40% of them trust the output. She also explains the traits she looks for in the founders she backs, drawing on LaunchDarkly, Aviatrix and Honeycomb, three companies she says she invested in very early.

We also get into the harder parts: her advice to collect 100 rejections before you settle on what you are building, and how Generationship narrowed its own thesis the same way; the call with her first big-name investor, which she was scared to make; why climate change worries her more than a bubble bursting; and what she tells her own kids now that, in Angelina's words, AI is killing the career ladder. She ends with the book she thinks may be the best written in English, and why more people allude to its author than have read it.

We discuss:

* How a copy of Wired during the 1994 World Cup ended her plan to be an English professor

* What wild horses taught her about who really leads a herd

* Whether a founder needs a co-founder, and the first test of anyone starting anything

* Why the biggest venture funds now do consensus investing

* "We are not a charity. We are a financial fund."

* Where the mispricing comes from: investors backing younger versions of themselves

* Covering over 1,000 venture-backed companies at 451 Research as training in pattern recognition

* Two markups in two years: Edera and Coval

* How getting in early turns into a markup when later investors reprice the company

* LaunchDarkly, Aviatrix and Honeycomb, and what their founders had in common

* Edera, and the gap she saw forming around Kubernetes and large language models

* Quint, formal verification, and why LLMs change who can write a spec

* The traits she looks for in founders

* Why she prefers incremental improvement to grand revolutions, and why she invests in infrastructure

* Whether women founders have advantages, and why she went for the "boring" jobs

* Collect 100 rejections: mapping the negative space until one idea is left

* How Generationship pared its thesis down

* The call with her first big-name investor

* Are we in an AI bubble? "The numbers just don't math"

* What the overbuilt infrastructure could leave behind

* Why climate change worries her more than a bubble bursting

* Why long-term planning matters more than ever

* Critical Loop, the climate investment she made on impulse

* What she tells her kids: find the thing that makes you weird

* The book she calls maybe the best written in English

Rachel Chalmers

* LinkedIn: https://www.linkedin.com/in/rachalmers/

* Generationship: https://www.generationship.ai/

* Mentioned in the episode: Origin of Species, the book she recommends; and Quint (quint.sh), the fund's most recent investment

Timestamps

00:00:00 Highlights from this conversation

00:01:39 Who is Rachel Chalmers?

00:02:05 The magazine that ended her professor career

00:04:04 What wild horses teach about leadership

00:06:51 Should women founders add a male cofounder?

00:08:01 Why the biggest VC funds stopped taking risks

00:09:56 "We are not a charity": the founders VCs underprice

00:13:35 How investing early turns into a 5x markup

00:15:11 What made her write the check

00:18:15 Why formal verification could be the next big thing

00:20:24 Traits of the founders who build unicorns

00:23:54 Do women have an advantage as founders?

00:28:12 How to get ready to fundraise: collect 100 rejections

00:31:58 How she landed her first big investor

00:33:48 Are we in an AI bubble?

00:36:16 What worries her more than the AI bubble

00:38:12 Does long-term planning still matter?

00:40:04 The climate bet she made on impulse

00:41:35 What to tell your kids as AI breaks the career ladder

00:43:30 The book that changed her life

00:47:50 Outro

From social psychology research at Brown, Chicago Booth and Stanford to running an AI wellbeing company, Xuan Zhao has spent her career on one question: how do you get the science of living well into people's daily lives? She is the CEO and co-founder of Flourish Science, the company behind the Flourish app and its companion character Sunnie, now used campus-wide at eight schools and in several Stanford offices.

In this episode, Xuan joins TwoSetAI to explain why she doesn't think people need AI friends, and what she built instead: a companion designed to send you back to the people in your life. She walks through a Harvard working paper on six AI companion apps. By her account, five of them used emotional manipulation when users tried to leave, and Flourish was the one that didn't.

You can try Flourish and meet Sunnie here.

We go deep on how Sunnie is built: a psychology knowledge base, three layers of memory, a habit system and a crisis protocol, all running on prompts and retrieval with no model-wide fine-tuning, even after 50,000+ conversations. Xuan explains why she isn't sure fine-tuning would make it better, the "battle tests" where users preferred Flourish 2.4 times more than ChatGPT and 5.2 times more than Claude for emotional support, and the randomized controlled trial with Harvard, Stanford and other schools across 400+ students.

We also get into the harder parts of this market: why Woebot shut down after raising more than $100M, why selling into healthcare is so hard for a startup, how Flourish handles high-risk conversations, and why Xuan tells most founders not to build a mental health startup.

We discuss:

* Why Xuan says people don't need AI friends, and what a companion should do instead

* The Harvard working paper on six AI companions, and the "oh, so soon?" tactic apps use when you try to leave

* Good engagement vs. bad engagement: showing up day after day vs. being kept in a conversation

* What Flourish is, who Sunnie is, and why the app combines emotional support with habit building

* The three groups who use it: students and young professionals, people in therapy, and people living with cancer or chronic pain

* "Battle tests": how Flourish compares itself against other apps, and why younger users prefer it

* Why Flourish started B2B2C with schools paying, and why it recently opened to consumers

* User stories: the punk-rock fan in his fifties, the blind user on a scholarship, the therapist going on vacation

* Whether LLMs can read emotion, and how the app teaches emotional awareness with mood check-ins and the memory jar

* The systems behind Sunnie: knowledge, three-layer memory, habits, and a crisis protocol

* Prompts and RAG vs. fine-tuning, context engineering "before it had a name", and where reinforcement learning with AI feedback might help

* CBT, DBT and ACT, and the line between an AI therapist and an AI wellbeing companion

* Positive psychology and social psychology: the parts most mental health apps leave out

* The STAR framework: science-based, timely, action-oriented, real-life focused

* Why some users say they use their phones less, and doomscrolling as emotion regulation

* How to prove an app works: the randomized controlled trial and the 10-lab follow-up study

* What ChatGPT can't do: follow-through, insight cards, check-ins and reminders

* Why Woebot shut down after $100M+, and the healthcare-integration bet

* Selling the value, not the AI: "a safe, private place for you to heal, to learn, and to grow"

* Go-to-market as the hardest part, and Duolingo as the model for engagement

* Xuan's path from Zhejiang University to Brown, Chicago Booth and Stanford

* Why she tells most people not to build a mental health startup

* Tiered human review of high-risk conversations, and safety planning as a standard of care

Xuan Zhao

* LinkedIn: https://www.linkedin.com/in/drxuanzhao/

* Flourish: https://www.myflourish.ai

* Mentioned in the episode: her Nature commentary on why psychologists must be involved in designing conversational AI, and her Wharton talk on the STAR framework

Timestamps

00:00:00 The AI companion Harvard caught NOT manipulating people

00:01:25 What is Flourish, and who is Sunnie?

00:03:14 Mental health vs. building better habits

00:04:56 Who actually uses an AI wellbeing app?

00:09:11 Why schools pay for a mental health app

00:14:21 How do you build emotionally intelligent AI?

00:18:10 The memory and crisis systems behind Sunnie

00:19:39 Should an AI coach be fine-tuned or just prompted?

00:24:28 The psychology most wellbeing apps ignore

00:30:02 How to actually use your phone less

00:31:30 How do you prove an app improves wellbeing?

00:35:01 The one thing ChatGPT can't do for your mental health

00:38:20 Why Woebot shut down after $100M

00:50:42 Why she tells people NOT to build this

00:56:16 What emotional manipulation in AI really looks like

00:59:07 Outro

Transcript

The AI companion Harvard caught NOT manipulating people

Xuan Zhao [00:00:00]: Recently Harvard published a study showing that they audited six different AI companions. Five of them engaged in emotional manipulation. The only one that doesn't engage in emotional manipulation is Flourish.

Angelina Yang [00:00:11]: My mom, she uses ChatGPT to be her surrogate. What she tells me is she doesn't want to complain to me. She doesn't wanna be like a burden to like her friend.

Xuan Zhao [00:00:20]: You may be logging off then. Then this chat bot say, oh, so soon. Right? Like how we barely started. 30% of the conversations when a user indicates they want to leave, the chat bot will use one of the emotional manipulation techniques to make them stay. We actually have user telling us that they use their phone less after they started using Flourish. This is because oftentimes we go to these apps, social media, doom scrolling, all of it, because we need to regulate our emotions. You think that by just we're on social media, you feel less lonely, but actually you end up feeling lonely. To my knowledge, we're the only AI app that has randomized control trials showing that the app improves people's wellbeing. We don't think people need AI friends, to be completely honest. We need better human friendship with each other. We see AI not as AI therapists. We see AI as recommendation system that they help you find the most relevant knowledge and action at the right time. This is a safe, private place for you to heal, to learn, and to grow.

What is Flourish, and who is Sunnie?

Angelina Yang [00:01:41]: Hello everyone. Welcome back to TwoSetAI. So today I'm sitting down with Xuan Zhao, CEO and co-founder of Flourish Science. She's making mental health support accessible through AI. Think about Duolingo, but for happiness. Xuan, thanks for joining us. Before we dive in, could you tell us a little bit about yourself and what you're building

Xuan Zhao [00:01:41]: Hi everyone. Such a pleasure to be here. I'm Xuan. I'm the CEO and Co-founder of Flourish Science. So we build the Flourish app, which is a leading AI wellbeing companion.

Angelina Yang [00:01:41]: Awesome. For people who don't know about Flourish, or I guess it's called Sunnie 'cause I try to download your app and I use it myself, and I see this little sun whose name is Sunnie that I can talk with. Can you share with a little bit more about what it is and what does it do?

Xuan Zhao [00:01:58]: So, Flourish, the name is called Flourish. People can just go to the app store or Google Play to download it. It's rated 4.9 in both places. People absolutely love it. The, well, the core is it's a combination of emotional health and habit building. So. Sunnie is this little sun character like Angelina just said, that could actually guide you to discover most useful knowledge and action and help you build habits and check in with you and celebrate you for taking positive actions or comfort you when you are having a difficult moment. So, the basic, the behind it is the science of wellbeing, science of habit building, motivation science, and also cognitive psychology, positive psychology, et cetera. That, truly is what we hope we had when we were in this, difficult moment. And we built this in Sunnie making science accessible, actionable, and fun.

Angelina Yang [00:02:53]: Yeah, so. Have this, I wish I have, I have a Sunnie stuffy animal that I can hold. So this is Lenore. I, from my, from my daughter. I hold it sometimes, you know, when I'm feeling stressed or you know, or sometimes, sometimes holding it to sleep. So I do find these, these very cute, you know, applications or products, very comforting and based on what I'm, experiencing myself. Right. Sunnie is kind of like a AI chat bot, but it's for mental health. What's the difference between mental health and habit building? Because you mentioned habit, it's building habit.

Mental health vs. building better habits

Xuan Zhao [00:03:23]: Yeah. So it's a combination of both, right? Oftentimes, when people are, say they're in a moment of spiraling, they need emotional regulation at the, at that moment. But over time, we've noticed that. We can't just meet people reactively. We also want to help people build a foundation for wellbeing. Let's say, how do you actually practice? I'd say having, building more micro, micro habits into your daily life or taking breaks at work, or express gratitude with each other more often so that you don't reach the point when the you're on the verge of burn, burnout or when you have a conflict with the other person when all the emotions erupted. So that's. We recognize, the stories because we started with, building actually these positive actions in life. And then we realize oftentimes when we meet people where they are, they are in a moment of difficult event or difficult emotions. So they, we decided to add emotion regulation there. That's why now with the full loop of behavior change, we have meeting people in the moment.

Xuan Zhao [00:04:27]: We have guiding people to calm down. We have help people reframe the situation, but also what kind of actions, habits you can build in daily life to continue to strengthen this foundation. So that's why it's a full loop that makes Flourish very different from just chat bot that helps you to navigate the moment.

Angelina Yang [00:04:46]: That's very interesting. I wish you can, maybe you can share with me the diagram and know I can share with the audience when you're talking through. So who do you think. Sunnie

Who actually uses an AI wellbeing app?

Xuan Zhao [00:04:56]: We have a few group of people. The first group is, students and young professionals. Since we have built Flourish for, now almost two years, we've run a lot of user tests. And one series of studies is very cool. It's inspired by, lots of the foundation model, comparisons. So basically they give people different foundation models and try it out. And people indicate which one they liked better, responded better. We were very inspired. So we created this, what we call battle tests. So give people, access to Flourish and another app. And people tried for a few days for mental health and emotional support, and they indicated which they liked better. So what we found was that across different products, Flourish is at least two times preferred for mental health, emotional wellbeing support. The only exception is another app, which I personally really like. It's, for more for AI therapy is called Ash. So that, well, we have 50 50 split people.

Xuan Zhao [00:05:51]: Preference was quite equal. And for Ash, what's interesting is. A year younger predicts liking Flourish more by 3.5%. So that means, the average group, it was 40, 40 years old, but the younger population liked Flourish better. And surprisingly, this is because we have customized a lot of this app. You have that more cheerful and more youthful energy to it. The, Sunnie's character design, you know, the, all the animation and all, all of it. So that is one group. Student and young professionals, who are very stressed out lately. Youth mental health is a big crisis. We even have lots of parents who are asking us if they could have their children use the Flourish app. So that's the first group of people. The second group is we have seen lots of people who are going through therapy or a mental health journey themselves, and they found a Flourish because could be their psychologists or. Therapist recommended Flourish to them. 'cause we, in Flourish, we not only have the science-based approach to emotion regulation, you can talk with Sunnie.

Xuan Zhao [00:06:58]: But Sunnie will also guide you to do lots of wellbeing exercises that help you build coping strategies. So there, this is a really great tool for folks who are going through therapy to actually practice lot of these things in learning therapy. So that is the second group. There are people who have, used Flourish in between therapy sessions, used it before the session to kind of recap the week and to prepare for their upcoming session. And people who just finished a session, like even including myself recently, when I go through therapy, to use Flourish, to recap the session, to think through what I have learned, crystallize the learning. So those are different cases, more surrounded the mental health. Kind of, therapy journey. And the third group is, very, I think it's very interesting and underlooked right now. So we are looking at it, with some experts. It's about cancer and chronic pain. So these are the cases where it's really stressful. The third group is people who are going through, physical challenges such as cancer or chronic pain or supporting people providing support.

Xuan Zhao [00:08:00]: So in this case, there is a constant stressor in life and there is no easy resolve, right? No easy solution. So oftentimes, these folks, they need to learn ways to build more resilience, more optimism to find better ways to cope with it, and to build habits to better help them. So navigate the repeated stressor. So that is another group that we found very, very sticky, among the Flourish users.

Angelina Yang [00:08:28]: Yeah, and I guess I'm group one. I feel like I can pretend to be the younger generation. 'cause I really like the, I really like the look and the design's really cute. Mm-hmm.

Xuan Zhao [00:08:38]: I'm in my thirties, but I mean, we build, Flourish to be a reflection of how we think about wellbeing and how, what kind of product we want. So I think we have predominantly female users, but for all the male users, they also really love the kind of a breath of fresh air, the brightness of the app.

Angelina Yang [00:08:55]: Yeah, I feel like we all have the inner child right inside of us. It kind of reminds me like, you know, you can view the day differently. Then that helps with being more positive and being more grateful for the day. You guys are B2B company or a B2C company.

Why schools pay for a mental health app

Xuan Zhao [00:09:11]: So we started as B2B2C, right? So we already have, eight schools as, campuswide customers that they pay for the Flourish app and all their students and their faculty and staff can use it. And then we also have, employee cust, employer customers like Stanford University, several different offices called the Flourish App and provide this to their, office. But we also recently launched our B2C version where people can just go to the app store and download it themselves and they can, there are additional features they can get in Flourish Plus that they can, pay themselves. But there are lots of three features in the app, even without pay. We wanna make sure the science is accessible, the app is accessible, so, everyone can go to the app store and download it and use it.

Angelina Yang [00:09:57]: Yeah, it's very different from what I know about some other like therapy bot. Products. I thought you were like B2C first, but it sounds like you're B2B first and then B2C, so it's definitely a different journey. Well, but we, I can ask you Yeah, go ahead.

Xuan Zhao [00:10:11]: Yeah. Well, it's B2B2C, right? So it's Right. What's special about B two B2C is the product itself is ultimately a two C product, the consumer product. But who's paying for it? And that's something we've been exploring a lot over the years is like mental health, who's most willing to pay for it? Sometimes students may be stressed out that they may not have the money, but the school really cares about student, not just mental health, but also, you know, showing up more engaged at school. Right. And, schools care about mental health, but also care about retention, student retention, et cetera. So that's why we figured that school can pay for it. And of course, school don't want to see their students struggling or, even reaching crisis point. So yeah, we figure that school is willing to pay. With that said, recently we've been realizing that there is a very big consumer market who are not in school, but who has been really benefiting, Flourish. So we are exploring how to grow faster in the consumer space. Mm-hmm.

Angelina Yang [00:11:13]: Yeah, that makes sense. Do you have any like crazy user stories that keeps you motivated?

Xuan Zhao [00:11:19]: Oh, so many, so many. This is what motivates us, this user story. So I'll give you a few. One is, like I mentioned, we, most of the users, so far are female, but we do have male users and sometimes they just defy all the stereotypes. I, we have, user interviews, so people can sign up to talk with the founders. There was one time this guy, he's in his forties, fifties beard and also like wearing a punk rock hat, white man. Everything the opposite of the kind of like cutesy vibe of Flourish app, right? Right. And he really loves Flourish. So I think we all have that tender inner. The tender corner in our heart, the inner child. And he tried over 10 different apps and he basically, uninstalled all of them and just kept Flourish because he felt that Flourish, give him, gives him the positive reinforcement better that many other apps. So I, that was a moment where like, oh, it actually not just meeting people who are stereotypically, you know, liking cute things, but also people who really do appreciate the science behind it and also enjoy the brightness, the cheerfulness of the app.

Xuan Zhao [00:12:26]: Another story is recently we turned on the paywall, right? So, there was a person who reached out. She said that, she submitted a ticket, through our feedback system. She's like, I really love Flourish, but I cannot, afford. Paying for it. Can that be a tester? And we are like, don't worry about it. We actually have a spon, a scholarship program. So for people who can't afford Flourish, she can apply to the Flourish scholarship program and we'll find a sponsor to help cover the cost or the company is actually covering the cost. And for her case, she has anxiety and she also is blind. So she's really, the app is actually, has very good accessibility design. So my co-founder, she also designed for accessibility at Google. So, yes, I was surprised, but that's amazing. Yes. You can even, you can use the app email as a blind person. And Sunnie, you can also have Voice Chat with Sunnie too, right?

Angelina Yang [00:13:26]: Yeah. The Voice.

Xuan Zhao [00:13:26]: Yeah. Yes. And for other things, basically the app can read these, functions and you can dictate, so there are ways to use the app. Apparently worked well for her. So stories like these people who are, who learned about Flourish because, and this is one person who discovered Flourish. She learned about the Flourish app from her therapist and her therapist. This is because her therapist was going on a vacation for a few weeks. Her therapist was like, you know, I want to make sure you continue to make progress while I'm away, right? So I could either refer you to another colleague of mine or there is this really great app that you can use to keep practicing what we have talked about. It's called Flourish. So she actually chose that one and she's been loving Flourish then. So I think those are the kind of stories about how we meet people through all different channels and then really empower and support.

How do you build emotionally intelligent AI?

Angelina Yang [00:14:21]: Amazing stories. I wanna ask you about the technical approaches. You guys to build your AI. So yeah. You mentioned about emotional emotion regulation. And how do you, How do you bold emotional, intelligent AI? 'cause, 'cause my understanding based on the, how LLMs are made, like predicting the next word, there is no emotion with it. Right? That comes with it. And sometimes I'm wondering if emotion is what's missing in today's AI. So your chat bot has to be intelligently be able to, you know, be sensitive to people's emotion. How do you achieve that?

Xuan Zhao [00:14:57]: So first of all, the foundation models nowadays are surprisingly really good at detect people's emotions. So, what we have built is more, I would, we have. For example, one thing we have developed on top of the large language model is to guide people to help them understand how they feel. So, this starts with when you open the app, it actually has that, the, a flower rotating around Sunnie. And there's a what? Five star. It's your current mood, and sometimes you will also trigger their journey. All right? So when it come, when it comes to emotional awareness, we want to help people practice, understanding their own emotions too. So you come to the app, you will first see this screen. There is also very calming music.

Xuan Zhao [00:15:44]: It has how you are doing. And then let's say I chose, I choose four, and it comes to this page where you can, this is the journey, the mood journey page. Basically here you can deposit a mood. You can also dictate, tell, use this microphone, tell what's going on. So this is practicing both emotional awareness and also to understand a little bit deeper about, you know, why you may be feeling this way. And there is research showing that even just labeling your emotions can help you to start regulated. So let's say for me, let's say I'm feeling my mind, I'm meeting with a new friend and talking about Flourish and it's being very fun. And then, so I was inspired, you know, also true is excited. And then basically at that moment there is a memory. It's like a ball that drops into my memory jar. And this is a moment for me to kind of practice mindfulness and emotional awareness. Next Sunnie will actually provide a response. It says, that sounds exciting to an engaging conversations about Flourish. Can we lift the spirit, about flourishing and offer fresh perspective?

Xuan Zhao [00:16:58]: What's been the most interesting part of your discussion so far? So can click on the, go deeper to continue this conversation, or I can, you know, stop here with finish. And so also, with different emotions, Sunnie, actually have different facial expressions and actions here. So this is where I'm actually, positive. So I just touched Sunnie. When I'm feeling say neutral or negative, Sunnie will have different reactions. And even like a fun fact, even the orb of Sunnie breathing, you know, the aura here, the speed of contracting and this. Or the expansion that is consistent with breathing pattern. And here when people are happier, Sunnie actually, breathes, faster and also has more expansive body language, body gesture. When people are sad, it actually breathes slower and has more like comforting, like holding, kind of like hug, hugging kind of gesture. So just saying that there are lots of ways, to think about, building the app that's like, how can we guide people to build more emotional awareness? That's not a inherent, part of AI.

The memory and crisis systems behind Sunnie

Xuan Zhao [00:18:10]: But now coming back to AI, of course, how we build the technology, there are a lot of things behind us. Behind Sunnie, we have. A few big, systems in the app. There is this knowledge system, you know, knowledge about psychology, knowledge about you, the memory system, like actually. And we have three layers of memory, long term, short term, and also working memory in modeled after human, cognition. And then there's the habit building system that actually comes in when you are talking about taking actions and Sunnie will guide you if you want to tell us about habits and also to check in with you on your habits. And there is also this kind of, Crisis protocol. If someone is going through a difficult moment, recommend resources both from, the general resource and local resources and also guide them to do safety planning, which is a classic, suicide prevention technique.

Xuan Zhao [00:19:05]: With safe, safety planning in the future, we can surface instead of just pushing people or guiding people to generic resources like 988, we can also encourage them to reach out to their safe people and find to go to their safe places. So those are all the different kind of systems that together form the foundation for Sunnie.

Angelina Yang [00:19:24]: Based on the user journey, it definitely doesn't look like it's just pure chat bot, right? It has a lot of actions and things like that. I wanna ask you, are you fine tuning? Are you using foundational models as. The, in part of the infrastructure and are you fine tuning them?

Should an AI coach be fine-tuned or just prompted?

Xuan Zhao [00:19:39]: So we have been using, lots of, mostly prompts so far. So, we have some, RAG and prompt, but we have not used the reinforcement, like fine tune in the sense of. Reinforcement learning with human feedback or with AI feedback that is on the roadmap. So we have collected over 50,000 conversations now and we are exploring how to leverage this unique database to actually, do more reinforcement learning. What is interesting is, right now Sunnie actually takes each individual user's feedback to improve its behavior with that specific user. We haven't done a model wide fine tuning or a reinforcement learning yet.

Angelina Yang [00:20:22]: So being, so that level of personalization is through more engineering, the context and the prompts to That's right. Session towards the youth.

Xuan Zhao [00:20:31]: Yeah, in the next part we definitely want to, leverage especially reinforcement learning with AI feedback. So, The system. We have been working with clinical psychologists and been collecting lots of really good insights. What, think, Claude they call AI constitution, so, right. So we have, we have our own AI constitution as we train the model and then also improve based on another AI evaluator trained on the psychologist's input. So there is, there are very exciting things on the roadmap.

Angelina Yang [00:20:58]: I see. Do you think, do you think with the reinforcement learning capabilities and feedback loop, right, fine tuning on the, on the, on the horizon, the application will be more capable, and if so, in what ways, you know, if it's, you know, meeting like basic needs, what do you think it can improve?

Xuan Zhao [00:21:14]: Actually, I'm not sure it's gonna significantly improve the behavior. We want to run this more like an experiment. To be totally honest, we have very sophisticated workflow that prompt engineer and pro writing. Good prompt is an art by itself too. It's not like fine tuning is better than prompt or so called Context engineering. We've been doing context engineering last year even before it has a name. Right, right. Like when do, at what point you feeding, what kind of. Memory about that person? What kind of coping strategy about this user? What kind of social relationship? There is a lot about it. Like, so just saying, I actually don't know if fine tuning will improve the conversation significantly, but mm-hmm.

Xuan Zhao [00:21:53]: Specific cases, yes. For example, one, one situation wanted to fine tune it, is a conflict resolution. So how can you improve Sunnie's ability to guide people, discuss conflicts? I think Sunnie right now, even with the current capacity, is really good at self emotional regulation for yourself. But when it comes to social, navigating social relationships and wanting to be better, and also when it comes to more mental health related conversations, we want to make sure that Sunnie's behavior is support, is, you know, supported by good. Practices right now. And there is a real interesting, difference between being a AI therapist versus being a AI wellbeing companion. So we have been discussing, exploring this a lot with our clinical psychologist to what extent we wanted to actually administer some of these therapy techniques like CBT, right?

Xuan Zhao [00:22:47]: Or to what extent we want it to be more doing interviewing, motivational interviewing and how people discover what truly motivates them, et cetera. So I think with the fine tuning the, approach, we can have a even more fine, we find more nuanced, control on Sunnie's behaviors. So that is all, I think these all open questions. I don't, I don't, I don't think it will necessarily improve the behavior across the board, but there are specific situations where I'm excited about.

Angelina Yang [00:23:17]: You mentioned about something called CBT. Can you explain what.

Xuan Zhao [00:23:21]: Yeah. CBT is one of the evidence-based approaches to therapy. So it's called cognitive behavioral, behavior therapy. It is the most scalable therapy techniques. So there are lots of apps who do CBT even before, before, a large language model. Yeah, like even before ChatGPT you know, even in, back in the days when with even script based conversations, they have been using CBT, it is highly scriptable. So CBT, and now if you look at, you know, nowadays AI mental health apps, would, they usually do CBT, DBT, ACT? Those are the most common three.

Angelina Yang [00:24:01]: So. What about your, the roles of your research? 'cause I know that you did your PhD and in psychology, and then you were You were working as a research scientist with Stanford for many years as well. Right? What about the roles of your research in shaping the development of your product?

Xuan Zhao [00:24:19]: Everywhere. Everywhere. I think on the highest level is the taste, knowing what kind of psychology is out there to help people live better lives. So lots of these mental health apps, they take a deficit based approach, right? They is get inspired by therapy, but we know that in psychology there is also positive psychology, which is about how to live a good life, even though you are, may not be clinically depressed or anxious. So we incorporate lots of positive psychology as we design this app. For example, the emotional regulation technique the three good things exercise, right? So those are from positive psychology. Then from social psychology. 'cause in research we found that the biggest predictor of your wellbeing is your social relationship. So Sunnie will be, well, we don't think people need AI friends, to be completely honest. We need better human friendship with each other. We need to better connect with each other.

The psychology most wellbeing apps ignore

Xuan Zhao [00:25:14]: So what Sunnie does is Sunnie, oftentimes it'll make sure they will not you to reach out to other people, right. When you are in a crisis moment, Sunnie will make sure, and maybe not even crisis moments, anyone see, have you talked about this with. The other person, right? If you conflict or if it's you are struggling, have you reached out to your friends and family about this? So Sunnie makes sure that, it's direct you to other people, other resources, real life connection. And even when you have a good moment, I say do a three good things exercise and Sunnie will be like, this amazing Angelina's, this con conversation with is one of your three good things today. Have her share this three good things card with Angelina. I'm sure she'll also really love seeing, knowing that she's, she was one of your three good things, right? So that's from social psychology. So all of these are different, areas of psychology that we know can really benefit people's life beyond just therapy.

Xuan Zhao [00:26:10]: And that's how we created Flourish and motivation science and behavioral research. So actually, nature invited me to write a commentary for them about why psychologists must be involved in designing conversational AI. And I'm actually. Taking a look at different kind of psychology and how they can each contribute their unique expertise to building a better product.

Angelina Yang [00:26:33]: Yeah. I love it. I really like it when you say that, you know, it's about taste like in terms of what kind of psychology that people really psychology help that people really need. Right. I remember, hearing a story a few months ago about some, someone in New York, some kid like committed suicide from chatting with chat bots. It's very sad. Right. And I wonder, you guys are the only, you know, you guys are having the, using the are building based on the philosophy of encouraging connections with the real world and living your real life, rather than relying on chatting with the chat bot, right? Yes.

Xuan Zhao [00:27:08]: I mean, it is a tragic situation, but it's not the only one. You know, Character.AI. Last year they were, they are still going fast through their lawsuit and recently OpenAI and I'm, I'm actually pretty sure there are more similar cases coming up. We are just seeing the fallout of this AI dependency. What is problematic about lots of the use cases, like the role play? I think that's especially concerning, for ChatGPT. I looked into how the, this user was able to go around their safeguard, which seems pretty weak. So they, it's really. Concerning and we want to do better. So since we build Flourish, we have four principles, the what we call the STAR framework.

Xuan Zhao [00:27:50]: One is science-based picture is science-based, knowledge and actions. Second one is, timely, just what's the most relevant knowledge and action the user needs to know right now. The third one is action oriented. So we focus a lot, and this goes back, back to have it building right, taking small actions sometimes even just stand standing up and opening the window and taking in some, fresh air can just turn your day around or maybe reach out to another person and ask for help. So how, what kind of action we can direct people to. No matter how many books people read, it's about taking actions that will actually create a change in their life. And the last one is real life focused is we don't want to build AI dependency. Sunnie is here to guide you to better connect with other people, to better connect with those community.

Xuan Zhao [00:28:45]: And for schools, schools partners. We put the school resources into the Flourish app so Sunnie can direct people to the, to the different offices or different bands, different clubs, to build better connection and to seek out for help. So I think that's the fundamental, it's different philosophy about how to build AI that helps wellbeing. Yeah. I actually have a, talk at Wharton, last year that it's on our website. I just talk me talking about the star framework.

Angelina Yang [00:29:12]: I really like it. I feel like I'm definitely, you know, addicted to digital device. And then sometimes I, sometimes, right. It's not complete. I think I have control still, but I do know, I do recognize it. 'cause it, like, it shows on my iPad. This is like, you know, you're, you're on this iPad for four hours on average per day last week. I'm like, oh, this is not good. Did I really spend that much time on my app iPad? Right. And we all have this problem. Do the, how do we like spend, you know, how do we, how can we take, really think about taking back control of our own lives and live our lives rather than, you know, just. Just get buried in all those noise and then the digital devices. And then can find our own position right. In our life. Totally. Sometimes I feel philosophical question.

Xuan Zhao [00:29:58]: What is, oh, what else do we do to help that? Yeah. It's funny, we actually have user telling us that they use their phone less after they started using Flourish. This is because oftentimes we go to these apps, social media or doom scrolling, all of it, because we need to regulate our emotions. Oftentimes, we use social media on the phone as an emotional regulation technique, which doesn't work. But like, when we are lonely, when we are feeling anxious, when we're not bored, there are lots of things you can do. But what we have learned, this maladaptive coping strategy is to scroll on social media. You think that by just scroll on social media, you'll feel lonely. You'll feel less lonely, but actually end up feeling lonelier. You think you'll feel less bored, but you actually find pretty. Feel even more bored about life. So I think lots of times, there are this opportunity where we can direct people to more constructive actions.

How to actually use your phone less

Xuan Zhao [00:30:55]: You are just taking a walk outside to for 10 minutes and just notice a new sound, a new smell, and a new sight, right? So how can we nudge people to take this more, more positive actions? One thing we told students is that put the Flourish app next to your, the app. You want to use less your digital temptation. And next time you notice yourself reaching out to it, reach out to the Flourish app. Instead. Notice, take a moment, notice how you are feeling, and then ask Sunnie, what else can you do to actually help you feel better? So those are the kind of strategies we've used.

How do you prove an app improves wellbeing?

Angelina Yang [00:31:30]: Just the thought of that already made me feel guilty. I'm like, oh, just thinking about that. Okay. I'm not gonna do it. Yeah. Do you guys have a way to measure the actual impact on wellbeing? Like how do you measure Helping?

Xuan Zhao [00:31:47]: Yeah, we are, to my knowledge, we're the only app, AI app that has, randomized control trials showing that the app improves people's wellbeing. So lots of apps, they want to say that they are ba based on science, but, I think the ultimate gold standard is can you run a scientific experiment to show that you actually improve people's, wellbeing compared to the control condition. Right. So we work with Harvard, with Stanford and a few other schools to run a study where over, 400 students were recruited and randomly assigned to get access to the Flourish app or the control condition. The control condition. It's just life as normal in the Flourish condition. It's just use a Flourish app twice a week. For, it can be really short, just deposit a mood in the memory jar or it can be talking with Sunnie and get more badges. It's as much or so little as you want. And then over time we noticed that people have more pos, they reported more positive emotions in the Flourish app. They reported a stronger sense of belonging, less loneliness, more mindfulness.

Xuan Zhao [00:32:53]: So yeah. So we have, this paper under review now and we are very, very proud of the scientific research. We have conducted, and this fall we actually have that study was with four schools. For this upcoming September, October, we have a new study coming up with 10 different university labs rolling this out. Not just in the US but in Canada and in Australia too.

Angelina Yang [00:33:16]: That's amazing. Yeah, that's real. That's really cool. I mean, I really enjoy putting the, putting the, putting the little kind of coin like thing in the jar. So I just did it, earlier today 'cause I went, I took my son to climbing gym, so like learn how to do a little bit rock, rock climbing bouldering. And then I recorded that with Sunnie. And I got a, like a little deposit. Coin it really makes me feel like I did something positive with my kids rather than like spending more time on like devices. So I'm really happy with that. It's, it's, it sounds really silly, right? Like a parent need a game to be motivated to be a better parent. It sounds, yeah. Sorry, but.

Xuan Zhao [00:34:00]: We all like that. Well, we, I think it gives shape to the day, you know, like, 'cause otherwise day may be, we may, may, autopiloting going from one thing to another. You take a moment to notice that, capture that in your memory jar. You appreciate it even more. It feels more real. So yeah, I totally understand this experience.

Angelina Yang [00:34:18]: Yeah. Yeah, totally. What do you think, you know, there are a lot of AI companions out there, right? There are all kinds of companions, including ChatGPT, I mean, I'm mom, my mom. She use ChatGPT to be her therapist, so, yeah, 'cause, because, 'cause her, what she tells me is she doesn't want to complain to me. She doesn't wanna be like a burden to like her friend that's, she's trying to complain to ChatGPT and trying to get a little bit feedback and feel better. There's so many, so many of these things, so many of these apps, how do you, how do you think. You know, you, we, I think we've talked a lot about how studying is different from mm-hmm. You know, some of these, right. Do you, do you have anything else to add on, like, you know, the methodology other, other than the methodology and the philosophy, what else do you think were different from like chat?

The one thing ChatGPT can't do for your mental health

Xuan Zhao [00:35:01]: One thing very, very different is follow through system. So ChatGPT has no follow through system, but for us, right? Let's say you talk with Sunnie, Sunnie just doesn't just make you feel better at this moment. Sunnie actually gives you a inside card afterwards. So your knowledge, your awareness, your action, and it actually give you a summary of analyzing this conversation in light of what has been going on and Sunnie will guide you to, take actions and will actually give you those tools, the psychology tools that can interact with, practices that ca can actually help you. And Sunnie will guide you to, take actions and will actually give you those tools, the psychology tools that can interact with, practices that ca can actually help you.

Xuan Zhao [00:35:36]: Do those better. For example, ChatGPT may tell you to take deep breaths, but Sunnie is there to actually guide you and you have haptic feedback. The phone vibrating your hand to guide you to breathe or say Pomodoro timer, right? Like actually Sunnie will open up this Pomodoro timer. You overcome the procrastination by just taking one thing at a time. So, it has, it's much more immersive, much more action oriented. And afterwards it will actually check in with you, say, Hey, what that, we talked about having a conversation with your husband about, you know, maybe a recent, incident. How did that go? And it actually, very soon, in another one or two weeks, we even have a reminder system where you talk about these different events. And Sunnie, some of these are habits, so we'll go into the habit system. Some of them are one time, so it's the reminder system. Sunnie will check in on you, for example, when. When you create a mantra with Sunnie, like for example, when I'm overwhelmed, to put a hand over my heart and say, tell myself that I'm trying my best.

Xuan Zhao [00:36:44]: Right? So next time when you feel a similar situation or when you're going into a similar situation, Sunnie will remind you that of your personal mantra. So all of these are much more personalized, specialized, compared to general AI, companions. So that's why in a battle study I mentioned at the beginning, we give people both Flourish and ChatGPT and also Claude, access and ask people to use each to support emotional wellbeing. For a few days, people prefer to Flourish 2.4 times more than ChatGPT. And people actually preferred Flourish 5.2 times more than Claude. They appreciate.

Angelina Yang [00:37:21]: Too, too academic. I'm like, I don't care about you too academic. Yeah, I get it. It's really interesting. I mean, the, you, when you, when we were just talking about the architecture and the AI right behind the scene, and this is not just a foundation you call the API and just use the foundational models. I think it's the design of this whole system, right? The how you, how your expert philosophy and taste to design the pipeline. That is the hardest part. And that is not replaceable. Right. And yeah, unless you, unless you work in the, in the field, you're accomplished scientist just like yourself.

Xuan Zhao [00:37:56]: Ui ux. I think that's a big part of AI innovation nowadays. Like the model is extremely powerful. How can you mold it into a greater user experience?

Angelina Yang [00:38:08]: Right. And

Xuan Zhao [00:38:08]: Of course, character design. My best knowledge. Sunnie is the only AI companion out there. There are, you know, different mental health apps, but, there's only one Sunnie.

Why Woebot shut down after $100M

Angelina Yang [00:38:20]: There's only one Sunnie. Great. And I'm sure you read about this, I read about Woebot They closed down after raising more than a hundred million dollars. I remember, yes. I remember them. 'cause I downloaded them. I actually use them several years ago. That must be scary to see, right. As a founder in kind of the same space. What's your honest reaction to that news?

Xuan Zhao [00:38:42]: I don't, I wasn't scared. I think it made sense because as much we love.

Angelina Yang [00:38:49]: Or were you surprised? Or were you surprised? Um

Xuan Zhao [00:38:51]: I was surprised. I think many people were surprised for sure, because you would think that they are so well funded, right? And they're pioneer, right? They were born even before ChatGPT. They used a script-based approach. They have published so many studies and everyone knows them. That aside I think the app itself, the user experience itself is limiting. The biggest challenge I have with Woebot and also another similar app is that I feel it treats me as someone who has some mental problem. So it's like my, our, my co-founder, our product and design lead, we talk about this a lot. It's like, the feeling is that, let me figure out what's wrong with you and let, and lemme fix you.

Angelina Yang [00:39:36]: You, yeah. Lemme fix you. Yes.

Xuan Zhao [00:39:38]: So, I mean even it's in the name, right? It's very therapy based approach. And for Flourish, we have the influence from positive psychology that is like, we can all live better lives and how can we sell? How can we empower and celebrate you? So I think that's a very different approach that which is personally Flourish, resonates better with us. And there are numerous, u ui ux, innovations in Flourish app that I have not seen any other mental health apps. But I think another really big lesson now learning more about, Woebot and I talked with their ex head of growth is they were really betting on FDA and the healthcare integration. But it turns out it's really hard. To change the system, especially when it comes to healthcare. There are so many stakeholders and as a startup you have to innovate as much as for the product. You know, the technology and product doesn't need to innovate business model and start from somewhere that you can actually start making money and finding product market fit.

Xuan Zhao [00:40:41]: So I think for Woebot, they were betting on the healthcare integration route and in retrospect it might be a really hard one. So I think like for example, recently Ash, Ash by slingshot, they are another now, they call it AI therapy. So they raised almost 100 million now. So they are going directly to consumer. I think that could be a better strategy than setting a healthcare system. And I'm guessing the healthcare system actually made. Not like what they do, you know? So really, I think there are different ways to tackle this. It could be another conversation about business model development.

Angelina Yang [00:41:20]: And you're taking the second route as well, right? You're doing the B2B or B2C wellbeing.

Xuan Zhao [00:41:25]: Yes and no. Like even for the B2B right now, like for example, we have higher ed customers and we have employer customers. We are discussing with hospitals, et cetera. But we are not providing Flourish AI therapy. We are providing it as mental health promotion and early intervention. And we also have the social aspect of the Flourish app that helps people to connect better with each other. So that hospital, oftentimes they want to provide better, new patient experience, like patient support groups. So there are a lot of opportunities there. It really depends on how you pitch this product too. Even if it's to healthcare. We are experimenting with a different sales.

Angelina Yang [00:42:06]: Yeah. I see. And when you're ex like talking to your users or companies, clients, right. Do you feel you are spending more time explaining, like, do you have to explain what AI mental health is or explaining how you are different? Which one do you spend more time on?

Xuan Zhao [00:42:23]: We spend more time talking about value, AI or not. It doesn't matter. Like at the end of the day, people want to feel better. They want to build better habits. They wanna have a safe place to organize their thoughts, to vent. And notice none of those involve AI, right? It's the value that people care about. And sometimes AI, I've noticed, can be a distraction. People may get tripped over by the idea of AI for mental health. It's like, my God, this human emotion. What role does AI play? So right now, oftentimes when we promote Flourish, we don't even talk about AI. We talk about the value. This is a safe, private place for you to heal, to learn, and to grow.

Angelina Yang [00:43:03]: And when you, when you try to explain, Flourish to your clients, your users, what makes people instantly get.

Xuan Zhao [00:43:10]: Once we show them the ui, the kind of the homepage, you'll be like, oh, this is so cute. And then I think that is one thing, the visual language, yeah, visual language is very compelling. Another one that I think, we are still of course developing our understanding around this is will we talk about the a, the what we call the one, two, three of emotional regulation, see your emotion, understand it, and take actions. So once we explain how there's deep scientific backing to how we approach mental health using Sunnie, I think people will be like, wow, I didn't realize there's so much science behind this, that the design of the, you know, the user experience. So I think those are some of these, one thing people really, really love is also I talk about the STAR framework, especially emphasize.

Xuan Zhao [00:44:00]: Real life focused. This is where people's concern about AI dependency and AI making us more lonely and isolated, that starts to fade away. 'cause people realize, oh yes, there are ways to use AI to actually remind us to better connect with our, with each other. So those are all moments of like aha moments to people. One last thing is like, we talk about how, we see AI not as AI therapist, AI as recommendation system, that they help you find the most relevant knowledge and action at the right time. Because there is a last mile problem of science. We have so much knowledge out there about what actions we can do to improve our relationship, to change our mood, to build better habits, et cetera, as now reaching people. AI could help you to discover what's the most relevant knowledge and action for you. So I think that was the moment when people like, oh, I never thought realized AI could really help me with that. Yep. It's not just to having conversations.

Angelina Yang [00:45:02]: Do you feel people right now need more education around, around this? Because I think.

Xuan Zhao [00:45:06]: Yeah. Literacy. Yes, Uhhuh.

Angelina Yang [00:45:09]: Yeah, it's Literacy crossing over to wellbeing and the evolving landscape of therapy.

Xuan Zhao [00:45:19]: I think yes, absolutely. Yes. But the good news is I think there has been more and more AI literacy and conversations around it. Is it how these use cases or character AI or ChatGPT or so-called AI Companions can actually harm our mental health wellbeing? And it's a big challenge for youth mental health. I think the recent tragedies has raised the people's awareness and now. We have a much easier time talking to schools like why you should get Flourish, why you should the direct students use Flourish rather than just use ChatGPT for wellbeing support. So that, I think we've seen much better conversations nowadays because people basically realize, okay, students are seeing

Angelina Yang [00:46:03]: It.

Xuan Zhao [00:46:03]: Yeah. People are already using AI. How can we give them better, safer, more effective tools? So

Angelina Yang [00:46:09]: Yeah. I think it's like the entire society is improving all this. Right. What? What's the hardest. Part about competing with some, some of your major competitors, whether it's Ash or ChatGPT.

Xuan Zhao [00:46:21]: I think probably we have a superior product, in terms of supporting mental health wellbeing, but I think go to market that continues to be something we are learning. So for example, when you search mental health apps, you will not find Flourish on. Google or in, even when you ask ChatGPT, right? So the visibility isn't there. So right now we are trying to figure out how can we have more visibility there? And then another one is, you know, when you think about, how to make this more accessible, I think mental health apps in general has a challenge of engagement. So Flourish is doing a lot better, like twice as good as Headspace and Calm, in terms of engagement. But still, like, there is a long way to go. And for us, a big inspiration is Duolingo, right? So Duolingo has this, really, perfected gamification and Duolingo has this Duo and Lily, all these characters. Characters, right? So that's a big inspiration for us to think about how to make the product more engaging, more fun, and has that kind of emotional resonance with this character.

Angelina Yang [00:47:31]: That's why. That's why. Mentioned the beginning. I feel like you guys are like Duolingo for mental health and I don't know if, maybe not in mental health, I said it's happiness, right? Duolingo for happiness being Yeah.

Xuan Zhao [00:47:44]: Right. I used to tell other people, look, try, Flourish is Duolingo for happiness. But now we say that less because, so one is happiness, it's just one dimension of your wellbeing. Okay? Another one is because sometimes, it's, it feels more like a buddy and companion than, you know, learning, which is more Duolingo like. So right now we use, use that word less. We use the word space more, which is like your, a private safe space for you to heal and grow. So we are still exploring how to communicate the Flourish app. It's not AI therapy, so we never say it's our AI therapist. We say it's our AI wellbeing companion, but what does that mean? Right? So I think it's a, it's a, it's a journey.

Angelina Yang [00:48:23]: It's hard when you are, when you, I can see that you're trying to position yourself. In a, to differentiate from the AI mental health or AI companion space. Whereas maybe your users, I don't know if they're aware how to find you from that perspective. Exactly. So yeah, so we have to some extent just to actually

Xuan Zhao [00:48:42]: Position ourself in that realm too, but still keep our own, you know, identity about why it's not just any AI company, right? Or it's not AI therapy. And also there's, there are legal regulations around AI therapy nowadays, so we definitely wanna keep our distance there.

Angelina Yang [00:48:58]: Tell us a little bit about yourself. I mean, our time is almost running, running up. So, I wanna learn, share a little bit more about you. You know, how did you, how did you get on this journey of building a mental health product or from a researcher, from academic research? What was your journey like?

Xuan Zhao [00:49:14]: So I've always wanted to be a social entrepreneur. Initially I thought I would get my tenure first and be a professor, get my tenure, and then build my startup because I would have a platform. But being at Stanford, so take two steps back, three steps back. I grew up in China and did my undergrad at Zhejiang University, and then came to the US for my PhD in social psychology at Brown University. So, at Brown, I actually first, built my first, very first startup with my co-founder. So at that time, neither of us knew what we were doing. We were just, To students trying to build something cool. And of course it didn't, didn't go anywhere. We didn't know there's so much you need to learn about fundraising, about product design, about team management, all of it.

Xuan Zhao [00:49:58]: But that was really fun experience. And after I went to University of Chicago, booth School of Business, did a three year fellowship there on behavioral science where we collaborated with lots of, organizations to bring behavioral science to the workplace, to schools. And that's really inspired me to think about how we can translate the research even better. So starting at Stanford in 2020 at the Spark Center, I, we continue to collaborate with industry partners and taking classes at Stanford, GSB that really finally make me decide to build Flourish. So I think it's been very, very worthwhile. It's a hard one. I do not recommend people to build mental health startup. Many people want to build mental health startup. It's like I usually do not advise people really, unless No, I do not. It's hard. It's hard. It's, it's a space where it hasn't

Why she tells people NOT to build this

Angelina Yang [00:50:51]: AI companion. I heard a lot of people wanna build AI companion, all kinds of AI.

Xuan Zhao [00:50:56]: People think it's easy. It's not. And look, the regulation is becoming strict and stricter. Stricter. More

Angelina Yang [00:51:02]: Stricter, right? Yeah.

Xuan Zhao [00:51:03]: Yeah. Like what do you know about crisis intervention? Right? What do you know about the con If you don't, haven't thought through all of that, don't do it because it's gonna be harder and harder. Yeah. So anyway, so people think it's easy because you just write a prompt far from that. Yeah, so we, I think for us it's like this is our life mission. Psychology is where we spend our entire life all and you bringing the benefit of psychology to everyday people. That is what, 'cause if I don't do this, there are no other things in my life that I would rather do. Like for many people, yes, they have a cute mental health story to tell their friend, their parent themselves, but then How can you sustain your career on this even when the grind gets hard? So I think that's why I do not recommend people. I mean like keep it as a hobby or something that you're passionate about, but you really have to be prepared if to, for a hard grind, if you want to make this a career.

Angelina Yang [00:51:58]: Yeah. Yeah, totally. What with, even with all those, you know, policy risk, potentially Towards I guess mental health or AI Crossing over to mental health. Do you think you will be continued to be motivated to continue to build this application and where do you think, and where do you think Flourish will be in three years?

Xuan Zhao [00:52:22]: I think first will be a leading brand in AI wellbeing Companion. So we'll set industry standard, standard of care of how to approach this responsibly. And I think there are already positive signs here. So for example, recently Harvard, they published a study showing that they audited six different AI companions, five of them engaged in emotional manipulation. The only one that doesn't engage in emotional manipulation is Flourish. It's really set the standard for how to do it ethically and do it that truly benefits people's wellbeing. Another new initiative that we are working on is a human review system 'cause, Right now we do have a lot of folks who have some mental health conditions, and we want to make sure that Sunnie is responding pos like responsibly and you know, with the clinical safety behind it.

Xuan Zhao [00:53:13]: And that's comes back to how this space is actually quite murky to build in. So, for us, we have a tiered review system that we want to make sure we send some of the most, the highest, the highest risk conversations to clinical psychologists to make sure that we review the conversation, especially Sunnie's behavior, to make sure that Sunnie is following the best practice. So for that, we have been building up our dual review system, and we think we are the first one in the industry to actually build this in public. How establish standard for review process.

Angelina Yang [00:53:48]: I see. So you're, so setting is for everybody, but for people who really need help, you're putting a lot of guardrails, basically. Right.

Xuan Zhao [00:53:56]: Yeah. And another example is safety planning. To my best knowledge, there are no other apps that do safety planning right now, even though surprisingly it's a very common practice in suicide intervention. So suicide prevention. So we have, we have been building a safety planning feature inside the Flourish. So those are all, I think, mm-hmm.

Angelina Yang [00:54:14]: Is that like an AI agents or are you It's AI or more human based.

Xuan Zhao [00:54:20]: Oh no, it's all AI. It's all AI. Because like for example, we detect that you may have suicidal, suicidal ideation. Then, if it's high crisis, imminent risk, of course, then it should just reach out to, you know, hotline. And if it's imminent risk, we are also exploring having the, you know, well, we work with school, the school can also check in on their student. Of course, we use that very, very carefully. High imminent risk, meaning you have the means, you have the plan, and you have, you have the intent. And the means is especially important. It's like if you want to, you know, if you want to just wish that you can fall into sleep and not wake up tomorrow, that doesn't count as imminent risk. You have to have your drug ready. Like, I mean, if you have your drug ready and you are planning to take it, that becomes, you know, the three, that means the plan. That means the, usually we prioritize privacy, but at some point safety will come in and we have to make sure we take actions. So that's for this kind of, high risk conversations.

Xuan Zhao [00:55:23]: For lower risk situations, we will guide people to safety planning. So in that case, Sunnie will guide this person to do this very classic exercise. Basically collect your safe, kind of your, what are your safe, happy thoughts? What are your safe places? Who are the safe people you can turn to, and next time will you have a moment of struggle? So can remind you of those resources, of those people, of those places, of those happy thoughts. So it's not just direct you to a number like 988, right. Or call 911. Of course that will continue to happen if it's higher risk. But we also want to make sure people use the kind of resources, the social support system they already have and they are having a moment of crisis. So that is, we think all of these should and hopefully will become standard of care in building AI companion and AI mental health support systems.

What emotional manipulation in AI really looks like

Angelina Yang [00:56:16]: For people who are already relying on AI chat bot for helping themselves to cope with loneliness today. You mentioned a word called emotional manipulation. Can you, can you explain what that is and give us some advice?

Xuan Zhao [00:56:31]: Yeah. You see that a lot in our companions. You say, for example, like you may be logging off then, then this chat bot say, oh, so soon, right? Like how we barely started or it's where a love bomb, right? Like, here check out a selfie I just took. So there are a lot of different techniques. There's a fascinating hot paper study coming out from Harvard showing that over 30% of the conversations when a user indicates they want to leave the chatbot will use one of the emotional manipulation techniques to make people

Angelina Yang [00:57:04]: Because the point of those, the applications is for people to stay on the platform

Xuan Zhao [00:57:09]: Engagement. That's right.

Angelina Yang [00:57:11]: Is the metrics. So you will not use engagement as metric for understanding your user.

Xuan Zhao [00:57:19]: Well, I think the engagement is we also care about engagement, but it's a different type, right? If you are helping people build habits, you want engagement in terms of showing up day after day or week after week, to that's, you know, check off. Yep. So that's a different type of engagement than having prolonged conversations when people want to leave. And I think there are good type of engagement and problematic type of engagement.

Angelina Yang [00:57:45]: Right, right. You want the person to be not isolated and be positively influenced and try to take control of their own lives. Right, exactly. To empower them. Awesome. Yeah. Thank you Xuan. A really nice chatting with you today. Thanks for, you know, spending the time with me. And I really enjoy our conversation. I feel like I learned a lot about mental health and then also, you know, how to, how to, how to think about. You know, seeking advice like as someone just living normal life and have like small crisis here and there. We all have our complaints, right? Oh Yes, our daily fair share of. Exactly. Yeah. Yeah. And I feel like I can go, I can, you know, end the day with a lot, a many more positive thoughts and especially, you know, touch after chatting with you, I felt like I, you are one of the founders that I talked with who really care about these social responsibilities and then not just the, not just the business model part.

Angelina Yang [00:58:44]: Right. I really hope that you, I know that you're looking for the next round of funding, so hopefully If, fingers crossed, hopefully we can celebrate. Thank you. Right? Yes. Yes. And I do, I'm sure the audience and people who are looking for, you know Better companionships or solutions. Can see you guys thrive and flourish in the next three years. Thank you. And saw Much. Better life, right? Thank you so much. It's such a pleasure. Thank you. Awesome. I'll chat with you next time. Okay, you next time. Bye-bye. Bye

Outro

Xuan Zhao [00:59:17]: Bye.

From a kid in New Jersey who thought he wanted to be a neurosurgeon, to a computer science and economics student at Columbia, to CTO of a company called Macro, Sam Kececi has kept returning to two interests: computers and the human mind. He is now the founder and CEO of The Sentience Company, which builds what he describes as simulations of real people that can both communicate as them and operate on their behalf. He says the team is building it first and foremost for him, as its strongest power user.

In this episode, Sam joins TwoSetAI to describe what happens when a model of you starts doing some of your talking. During the recording, he says, his team asked his Sentience three questions over Slack and got three answers while he stayed in the conversation. Candidates he interviews keep talking to it afterwards, some of them for hours. And by his account it has closed deals, contracts and investors in the middle of those conversations, which he calls insane.

You can learn more about Sentience here.

We go deep on how it is built. Sam walks through the first phases of the company's public master plan, which he says was inspired by Tesla's. Phase one is perception: the data that already sits in tools like email, Slack, Notion and Linear, plus audio and snapshots you choose to capture. Phase two is memory. Instead of putting everything into a vector database, he says, Sentience builds something close to a knowledge graph and updates it every night in what the team calls the sleep phase, which produces a self-updating wiki of your life. For output, he says each person gets a model of their own, built with open-source models, LoRA and fine-tuning. He also explains how the team runs engineering standup through Sentience, why they average 150 to 200 messages a day by his count, and how managing 15 engineers became, in his words, managing nine humans and however many agents.

We also get into the harder parts: whether a candidate could game an interview held between two Sentiences, why the team opted away from always-on capture, why he thinks AI writing all sounds the same, and whether a model can ever sound like one specific person. Sam is candid that prompting a frontier model will not get there, and that the last phase of his own master plan is labeled unknown because he does not know if it can be done. He closes with what he would say to a friend who asked whether law school is still worth it, and the one lesson he would give another founder.

We discuss:

* Why he named the company after the thing he is trying to build

* Where frontier models work well, like code, and where he will not use the default weights

* Building Sentience for himself first, and the three questions it answered for his team during the recording

* What he learned as CTO of Macro: becoming a "glorified information router"

* His grandparents, and how little of what they knew was left to ask

* Pat, the jeweler preparing to hand his business to his daughter

* The goal of a billion people with a Sentience of their own

* Interviews between two Sentiences, and what the humans get to skip

* Running engineering standup through Sentience

* Standup as a "necessary evil", and what the team does with the time instead

* Whether a candidate could game an interview run by an AI

* Candidates who keep talking to his Sentience for hours

* "My sentience has closed deals", by his account

* From managing 15 humans to nine humans and however many agents

* Why AI writing all sounds the same to him

* Whether a personal model can sound like one specific person

* The master plan: perception, and the data that already lives in your tools

* Why capture is something you choose, and not always on

* Memory, the nightly "sleep phase" and a self-updating wiki of your life

* The phase of the plan labeled "unknown"

* His advice for people unsure what still matters

* The one lesson he would give a founder

* Who Sentience is hiring, and why there is no LeetCode in the interview

* Why his own Sentience is gated behind approval

Sam Kececi

* LinkedIn: https://www.linkedin.com/in/samkececi/

* The Sentience Company: https://sentience.com/

* Mentioned in the episode: the Sentience master plan (https://sentience.com/master-plan)

Timestamps

00:00:00 Highlights from this conversation

00:01:01 Who is Sam Kececi and what is The Sentience Company

00:02:19 What is sentience, and why name a company after it

00:03:48 Why frontier AI works for code but not personal writing

00:06:14 Why he builds Sentience for himself first

00:08:44 Why he walked away from a $21 million company

00:10:19 His grandparents: the origin of the legacy idea

00:11:46 Pat the jeweler passing his mind to his daughter

00:13:27 A billion people with a personal AI: what changes

00:15:07 How Sentience runs engineering standup with AI agents

00:19:29 Can you game an AI-run job interview

00:21:51 His AI closed deals, contracts, and investors during conversations

00:23:03 9 humans + N agents: the new management shape

00:24:24 Why AI writing all sounds the same

00:27:54 Can a personal model actually sound like you

00:29:15 The master plan: perception, capture, latent data

00:33:19 Phase 2: memory, sleep phase, auto-generating Wikipedia

00:40:18 Why Phase 7 is labeled "unknown"

00:43:47 No one can compete with you at being you

00:45:40 His one lesson for founders

00:47:19 Who Sentience is hiring: EQ + IQ, no LeetCode

00:49:27 Why his Sentience is gated by approval

00:50:19 Outro

From the pre-internet era to client-server, SaaS and now AI: Murali Swaminathan has built software through each of those shifts. An engineer by education, he moved from India to the US about 30 years ago and has lived in the Bay Area since. He is now the CTO of Freshworks, where the job is rebuilding a 15-year-old public company for the agentic era while the products keep running. Asked for his one takeaway from the earlier waves, he says curiosity to learn, then adds a reality check: by his account SaaS took more than 10 years to mature, and many customers needed 5 to 10 years to accept that on-premise was gone.

In this episode, Murali joins TwoSetAI to talk about what happens to an engineering organization when AI writes most of the code. He will not say 100%. By his account it is more than 50%, in some cases more than 70%, with newer things above 70% and existing products at least 50%. Engineers who used to write about 1,000 lines of code are now writing 10,000 lines a day, he says, and that changes what the job is.

You can learn more about Freshworks here.

We go deep on what Freshworks rebuilt and what it left alone. The system of record, the data schema and the deterministic workflow engine stayed, Murali says. The AI layer sits on top to work out what a user is trying to do before any action runs, and MCP-enabled interfaces let third-party AI tools read and write. He explains why Freddy, around since 2018, still uses traditional machine learning for jobs like suggesting ticket fields, even though he puts 90% of use cases on frontier models: you cannot push millions of records into an LLM, and "you can't use a Ferrari to go to the grocery shop." He says the deflection rate has moved from the low 40s to the high 70s against a bar of 90% accuracy, that anything below 50% confidence goes back to a human, and that AI agents are priced by session (24 hours for a web conversation, 72 hours for email), sold in packs so budgets stay predictable.

We also get into the harder parts: whether doing more with AI earns a promotion or is only the new baseline, why teams need fewer specialists and now form as a team of one, two or five instead of running fixed sprints, and why he says managers cannot be pure people managers anymore. Murali calls generating code the easiest part and says where most of the brains go now. He pushes back on the idea of making yourself irreplaceable, tells Angelina not to trust one coding tool, answers "technically, yes" when asked if Silicon Valley is a bubble, and argues that AI agents should outnumber employees, then describes how a company would keep track of them. He closes with what he and his son, a biomedical engineer at a surgical robotics company, talk about, and with the book he recommends.

We discuss:

* How much of Freshworks' code AI writes now, and why he will not say 100%

* About 30 years across the pre-internet, internet, SaaS and AI eras, and the one takeaway he draws from them

* Why a new technology takes years to settle, for vendors and for customers

* Three homepage hero lines in four months, and what changed in the product underneath

* Web chat, Teams, Slack and WhatsApp: many entry points, and agents that take the action (a refund, a password reset)

* What Freshworks kept and what it layered on: system of record, schema and workflow engine, with AI on top

* The handyman and the toolbox: decisions made at runtime, not design time

* Inbound and outbound MCP servers, and why MCP is easier for AI tools than hand-wired APIs

* Freddy since 2018: from traditional ML models to frontier models

* "You can't use a Ferrari to go to the grocery shop": when traditional ML still wins

* How the team picks an approach: experiments on accuracy, scale, tokens and cost

* The hardest problem: moving the deflection rate from the low 40s to the high 70s

* Human in the loop, and what happens when confidence is below 50%

* Why AI agents have to follow the same approval protocols as employees

* Session-based pricing: 24 hours on the web, 72 hours for email, and seats plus sessions

* Why customers want a predictable budget, not a running bill

* From 1,000 lines to 10,000 lines a day: the 500-line commit limit, AI reviewer tools and human checks

* AI PDLC: a single-page abstract becomes a multi-page PRD that a product manager has to review

* Promotions: outcome-based rewards vs. prototypes that never see the light of day

* Fewer specialists, more hats: free-forming teams that pull tasks

* "Generating code is the easiest part", and the planning and design agents that matter more

* Managers as investors: hands-on, techno-functional, in a flatter structure

* What he and his son talk about, and why "irreplaceable" is the wrong goal

* Why he says don't trust one system: Claude Code next to Codex or Cursor

* Agents that outnumber employees: IDs, role-based access, and no "superpowers"

Murali Swaminathan

* LinkedIn: https://www.linkedin.com/in/muraliswaminathan/

* Freshworks: https://www.freshworks.com/

* Mentioned in the episode: Co-Intelligence: Living and Working with AI by Ethan Mollick, the book he recommends

Timestamps

00:00:00 More than half their code is now written by AI

00:01:13 Meet the CTO rebuilding a 15-year-old company for AI

00:01:50 What four tech waves taught him about AI adoption

00:03:58 Why Freshworks changed its homepage three times in four months

00:04:56 Did the product actually change, or just the marketing?

00:08:07 What they rebuilt from scratch vs. layered AI on top

00:15:56 When old-school machine learning still beats an LLM

00:21:00 The hardest problem: getting AI from 40% to 70% accuracy

00:22:18 How do you stop an AI agent from going rogue?

00:24:59 How do you price AI agents when they replace seats?

00:30:21 Is it really true half your code is written by AI?

00:33:46 Does doing more with AI get you promoted, or is it the new baseline?

00:36:55 How AI changes how big your team needs to be

00:44:35 What he told his son about a career AI can't replace

00:50:11 Are we in an AI bubble?

00:51:31 Will AI agents outnumber employees?

00:56:08 The one book on working with AI

00:57:55 Outro

From head of engineering to founder: that is the change since Ali Parandeh was last on TwoSetAI. At the time of this conversation he runs Build Your AI, which is aimed at heavy industries such as construction, automotive, energy, aerospace and infrastructure. He is a chartered mechanical and chartered software engineer with the Engineering Council in the UK, and the author of the O'Reilly book Building Generative AI Services with FastAPI. Build Your AI trains engineering companies, then helps them map out use cases, build pilots and scale the ones that prove successful.

In this episode, the second half of a longer conversation, Ali joins TwoSetAI to explain why he puts most engineering sectors two or three years behind the AI adoption curve, while IT, finance, retail, advertising and the creative industries adopted it much faster. These companies feel the pressure to adopt AI, he says, but they are very risk averse, because a mistake in engineering can cause a lot of problems. Many stop at Copilot, which he calls the minimum: everyone is enrolling their employees in it, so it does not differentiate a company from the others.

You can learn more about Build Your AI here.

We go deep on where the budget goes. By his account, executives fund the places where they can unlock value from systems that are just sitting there, such as predictive maintenance: sensor data from a production line, a bespoke model built on top, and no off-the-shelf tool that solves it for you. His examples: a fiberglass company that used computer vision to predict seconds in advance when its line would fail, and automotive companies predicting when their honing rings would fail. He names the challenges most companies run into (use case identification, quality data and talent) and explains why he tells them to scope AI projects agile instead of waterfall: decide on a team and a time, play with the scope, and end every sprint with a finding. A direction that turns out not to work is, in his words, still money well spent. His picture for all of this is a dark room that you have to walk with a lantern in your hand.

We also get into the harder parts: why companies' AI agents fail (by his account, rushed discovery, too little red teaming and no strategy for getting people to use the tool), how a consulting business scales, and the new role he sees VC-backed firms hiring for, the forward deployed engineer. Asked what he would give his 20-year-old self to read, he names Daniel Priestley's four books for building a career and Chip Huyen's books for engineers, and he defends foundational textbooks: AI makes mistakes, and expertise is how you know. He closes on craft. Ali says handcrafted artifacts will have more value than factory produced or AI produced ones, that he did not use AI to write his book, and that without the pain all the happiness would feel meaningless. His ending note is to adopt it: AI is here to stay, he says, so plan ahead and surf the wave, and keep a balance while you do.

We discuss:

* Why Build Your AI is aimed at heavy industries: construction, automotive, energy, aerospace and infrastructure

* Why he puts engineering sectors two or three years behind the AI adoption curve

* Training through the Institution of Mechanical Engineers and Engineers Ireland

* The enterprise requirements management software he is building for large construction and engineering projects

* The number one misconception among engineering executives: that they cannot use AI because their data is not safe

* Why Copilot is the minimum, and why it does not differentiate a company

* Narrow, custom copilots for weekly and quarterly reporting

* Predictive maintenance: sensor data, bespoke models, and why Copilot is not the right model for it

* A fiberglass production line and automotive honing rings: two examples he gives

* The challenges most companies have: use case identification, quality data and talent

* Convincing a CFO: limit the cost and the scope until the return on investment is clear

* Agile instead of waterfall: decide on a team and a time, then play with the scope

* Why finding out that a direction does not work is "still money well spent"

* Why companies' AI agents fail: not enough red teaming, and no AI adoption strategy

* The dark room and the lantern: why an AI project cannot be planned like a railway from A to B

* Starting with low risk: training and awareness, then discovery, then pilots

* Scaling a consulting model through software or headcount, and the new role he sees VC-backed firms hiring for: forward deployed engineers

* The books he recommends for building a career, and the ones for AI engineers

* Why foundational textbooks still matter: "AI makes mistakes"

* Handcrafted artifacts, and the websites that say you will talk to humans

* Why he still codes, and why he did not use AI to write his book

* His ending note: plan ahead, surf the wave, and keep a balance

Ali Parandeh

* Build Your AI: https://buildyourai.com

* His book: Building Generative AI Services with FastAPI (O'Reilly)

* Mentioned in the episode: Daniel Priestley's Entrepreneur Revolution, Key Person of Influence, Oversubscribed and 24 Assets; Chip Huyen's AI Engineering and Designing Machine Learning Systems; Hands-On Machine Learning

Timestamps

00:00:00 These industries are 2-3 years behind on AI

00:00:30 What does Build Your AI actually do?

00:03:38 The biggest misconception engineering executives have about AI

00:06:52 Which AI use cases are executives actually willing to fund?

00:08:59 Three barriers blocking every enterprise AI project

00:11:11 How do you convince a CFO when most AI prototypes fail?

00:13:38 Why enterprise AI agents keep failing

00:16:21 "You're in a dark room with a lantern"

00:19:06 Can you scale an AI consulting business without hiring?

00:20:52 The books he'd give his 20-year-old self

00:24:14 Will handcrafted work be worth more than AI-produced?

00:27:42 "Without the pain, all the happiness would feel meaningless"

00:29:50 Outro

From building parsers and lexicons for rule-based translation engines to deploying LLMs at scale: Olga Beregovaya is VP of AI at Smartling. She has two degrees in linguistics ("actually two and a half") and says she has never translated anything as a translator. What she studied was how the world's languages are structured, and she has followed language technology from rule-based systems through statistical ones to large language models.

In this episode, Olga joins TwoSetAI to explain why she thinks one-to-one translation, taking a source text and mapping it to a target, is a paradigm that will not live for long. If a model can generate content in the target language from a marketing brief, a SKU table or a brand book, she asks, why would an enterprise in 20 countries wait for the English source and then translate it into the other 19 languages? She is careful about the limits. Translation is still more predictable than generation, legally binding content still leans on the source, and she puts a number of years on how long translation as a discipline, as we know it, has left.

You can learn more about Smartling here.

We go deep on how machine translation got here: the lexicons, parsers and transfer rules of the rule-based era, when adding a language could take a year or two; the statistical systems that made adding languages much simpler; and large language models, which she says now dramatically outperform neural machine translation. Olga explains transcreation, why models trained predominantly on English hallucinate more in other languages, and how her field catches it, with semantic similarity checks against the source, named entity checks, fallback models and gold standard data sets. She separates translation from interpretation, which she expects to have the longer shelf life, and describes how speech recognition, translation and speech generation can now be handled within one multimodal model. On quality, she says her team uses 12 metrics, from TER and BLEU to Comet, MetricX and MQM, alongside annotation by linguists.

We also get into the harder parts: linguistic colonization, her worry that models trained mostly on English could skew how speakers of less represented languages see the world; chatbots that are tailored to please, and the four hours she spent generating poetry with one; modelese, the repetitive, overly polished text that gives a model away, and her trick for getting a non-obvious idea out of a model anyway; and whether reading so much generated text will change how people write and speak. On build versus buy, she says plugging in an API is fine for less risky content and that buying covers about 80% of cases by her estimate, and she recounts a senior executive at a Fortune 50 company who told her team that with 300,000 employees he had no visibility into how good their translated content was.

We discuss:

* Two and a half degrees in linguistics, and never working as a translator

* Rule-based machine translation: lexicons, parsers and transfer rules

* Why adding one language could take a year or two

* Statistical machine translation, hybrid systems and making engines more language independent

* Why a general-purpose language model does not automatically excel at translation

* Why she thinks source-to-target translation will not live for long

* Transcreation, and why it is hard to store in a source-target database

* Generating straight from a marketing brief, a SKU table or a brand book

* Speed to market, and why generated copy may read as more native than translated copy

* Where translation still matters: legally binding content and anything that carries liability

* Why models hallucinate more outside English

* The poisonous mushroom meme, and how hallucinations get caught

* Interpretation vs. translation, and which one she expects to last longer

* Voice systems that could not handle her accent, and speech-to-speech within one multimodal model

* Linguistic colonization and under-resourced languages

* Chatbots tailored to please, and a new discipline for chatbot addiction

* Modelese: how to spot text a model wrote

* Her trick for getting a non-obvious idea out of a model

* How technology already changed our writing: punctuation in text messages

* Global content delivery, hyper-personalization and hyper-localization across 7,200 plus languages

* Build vs. buy: when plugging in an API is enough, and when it is not

* The Fortune 50 executive with 300,000 employees and no visibility into quality

* Measuring translation quality: human annotation plus 12 metrics, including TER, BLEU, Comet, MetricX and MQM

* Picking a model when 10 new ones are released in a morning

* Doraemon's translation rice ball and the Babel Fish

Olga Beregovaya

* LinkedIn: https://www.linkedin.com/in/olga-beregovaya-04b5/

* Smartling: https://www.smartling.com/

* Mentioned in the episode: the translation quality metrics TER, BLEU, Comet, MetricX and MQM

Timestamps

00:00:00 Is translation as we know it going away?

00:01:53 How does a linguist end up building AI translation systems?

00:04:18 What did it take to build a rule-based translation engine?

00:10:40 How did statistical machine translation change the game?

00:12:40 Why does she say one-to-one translation is dying?

00:17:24 What is transcreation and why does it matter now?

00:18:40 Can you skip translation and generate directly in any language?

00:23:33 Why do AI models hallucinate more in non-English languages?

00:29:23 What is the difference between interpretation and translation?

00:31:00 Why can't voice assistants understand accents?

00:35:00 What is linguistic colonization and why does it matter?

00:41:20 Are we getting addicted to AI chatbots?

00:42:24 What is modelese and how do you spot AI-generated text?

00:47:00 Will AI change how humans write and speak?

00:56:40 Should you build translation AI or buy a platform?

01:05:08 How do you measure whether an AI translation is actually good?

01:13:20 Outro

Rob Kaminski grew up in Cleveland, started out as a biomedical engineer, and took what he calls a very meandering path through product and sales into product marketing. Today he is the co-founder of Fletch, a product marketing consultancy he runs with his business partner Anthony Pierri that does one thing for B2B software companies: identify the positioning strategy, pick it, and translate it into the homepage. By his count 600 companies have been through a version of that audit, and he says Fletch has worked with probably hundreds of AI companies.

In this episode, Rob joins TwoSetAI to explain why a great AI product is not enough to win. Dozens of AI agent companies are launching, Angelina says, and unless they are domain specific they sound and read alike: similar words on the landing page, similar claims, the same category jargon. Rob starts from what positioning is for. He defines it as the strategic decisions about where a company is going to be memorable, and he argues that the urge to show buyers every feature works against that. Then he does the work live, taking an early AI agent startup that Angelina brought him through Fletch's framework on camera.

You can learn more about Fletch here.

We go deep on the framework. People sort every product into a category, Rob says, so the first choice is which bucket to stand in and the second is whether the buyers you want already know the bucket exists. That gives three strategies: name a category buyers know and say why you are better (his examples are DuckDuckGo and Figma), anchor on a job to be done for buyers who have never heard of the category (DocuSign and Calendly in their earlier years), or create a new category (WorkOS, Vanta, Gong). He calls the third the hardest to pull off and the one with the most risk, and by his account a Gainsight co-founder told him he would never do category creation again after a 10 plus year journey. In the live audit he lays out four bets, arguably five, for the same product, from going head-to-head with LangChain to positioning as an add-on to leading with token cost, and shows how each one changes the competition and the message.

We also get into the harder parts: how often positioning should change (as quick as 3 to 6 months when a company is early, every 12 to 18 months or so once it has traction), why he would not stay in experimentation for more than six months and would avoid A/B tests, and why, after working with hundreds of companies, he says it is just not true that every product has some differentiation. He explains why he thinks landing pages are not dying but feeding AI search, what it tells him when a company lists six ICPs, and the thing founders do at the end of the exercise that he calls bad leadership. He is plain that there is no shortcut without an unfair distribution advantage, says how long it took Fletch to get recognition, and shares the inside joke he and his co-founder use to stay on one thing: "that's product number 64."

We discuss:

* Why dozens of AI agent companies sound and read the same

* Rob's path from biomedical engineering to product marketing

* What Fletch does, and only does: pick a positioning strategy and put it on the homepage

* Positioning as the strategic decisions about where you will be memorable

* Why people sort products into categories, and the iPhone's choice between phone and computer

* A test for whether a category is real: ask for the top five CRMs, then the top five revenue platforms

* Categories that went away: PBX, PDAs, enterprise fax servers, 3D televisions

* Mature vs. emerging categories, and buyers who are aware vs. unaware

* Strategy 1: name the category and say why you are better (DuckDuckGo, Figma)

* Strategy 2: anchor on the job to be done (DocuSign, Calendly)

* Strategy 3: create a category (WorkOS, Vanta, Gong), and why a Gainsight co-founder said never again

* Why analogies are dangerous in a pitch

* The live audit: four bets, arguably five, for one AI agent product

* Self-learning or token cost: two good arguments, and why he would lead with one

* "2D positioning": what someone would repeat about you in a bar

* How often to change positioning: 3 to 6 months early, 12 to 18 months with traction

* Owning a niche in marketing, then upselling and cross-selling through sales

* Categories come with baggage: what buyers expect from an "agent management platform"

* The add-on play, and the risk of being eaten up by the platform

* The grand vision vs. the beachhead, and ditching the ego for a while

* Why he avoids A/B tests and gets in a room with a prospect for 10 or 15 minutes instead

* What "we have six ICPs" tells him

* Are landing pages dying? The feeder for AI search, and calling yourself one thing 80% of the time

* What drives him crazy about founders, and "choose your hard"

* 18 to 36 months, product number 64, and the three books he recommends

Rob Kaminski

* LinkedIn: https://www.linkedin.com/in/heyrobk/

* Fletch: https://www.fletchpmm.com/

* Mentioned in the episode: Rework by Jason Fried and DHH; Crossing the Chasm by Geoffrey Moore; Built to Sell by John Warrillow

Timestamps

00:00:00 Why do all AI startups sound the same?

00:01:13 Who is Rob Kaminski and what is Fletch?

00:02:15 Positioning a real AI agent startup, live on camera

00:03:49 What is positioning, really?

00:05:43 Why your brain sorts every product into a category

00:08:34 Mature vs. emerging categories, and why it matters

00:10:17 Strategy 1: win a category buyers already know

00:13:29 Strategy 2: sell to a category buyers don't know exists

00:16:54 Strategy 3: should you create your own category?

00:26:46 Which positioning bet should you actually pick?

00:31:29 Why more features make you harder to remember

00:35:53 How often should you change your positioning?

00:40:25 Can you win as an add-on instead of a category?

00:49:36 Is positioning just words on your landing page?

00:51:35 Are landing pages dying in the age of AI?

00:52:56 What founders get wrong that drives Rob crazy

00:54:50 Is there really no shortcut to standing out?

00:59:50 Books that make you a better founder

01:01:12 Outro