TwoSetAI: Decoding AI Builders

Angelina Yang

Behind every AI company is a founder who almost quit, made a bet no one believed in, or figured something out that changed everything. TwoSetAI tells those stories through long-form interviews with the people building the future. Hosted by Angelina Yang. angelinayang.substack.com

  1. Oct 15

    Why His Creative Kids Didn't Want This Book | Joseph Enochs of Enterprise Vision Technologies

    Got a real industry problem? Post it on Painpoints, the problem board Angelina and Joseph announce at the end of this episode. From applied machine learning and COVID tracking in Los Angeles to an O'Reilly book on generative video: Joseph Enochs is Chief AI Officer at Enterprise Vision Technologies (EVT), which works with large enterprises from Fortune 1 down to Fortune 300. He joined when EVT had nine employees and was around $30 million; now, he says, it is "approaching a billion." O'Reilly has accepted his second book, on physical AI. In this episode, Joseph joins TwoSetAI to talk about the book his kids did not want him to write. They are creatives, and their reaction sent him into a lot of soul-searching about the place for creative people in the world of AI. "I always wanna try to build a bridge versus try to build a wall," he says. His son put it as a request: "If you could create a digital paintbrush that would allow me to have the agency of my creativity while not taking that away." Joseph's book, Video Generation with AI, is on O'Reilly, and more of his writing is here. We go deep on how the book is built: free video generation in Google Colab, the provenance of training data, diffusion transformers built from scratch, vision language models, audio, and safe products with ComfyUI, LTX Desktop and InvokeAI. For commercial work, he says, use models that were not built from copyrighted material and key frames that are your own or that the artist gave permission for; for open source, he names LTX. On why AI video still looks fake, his answer is consistency: everything inside the 3D "scene box" around a shot, from the lighting to the shape of your glasses, has to hold from frame to frame. He holds up a small AMD box with 128 gigs of shared memory that he says can run pretty much any video generation model, and he explains why he sees video as the on-ramp to tokenizing proteins, chemistry and, one day, a digital twin of a mouse model for clinical trials. We also get into the harder parts. Asked what is left for creators that AI cannot replace, he answers, "Everything. There's no fear. There should be no fear." On agents, his test is how fast a mistake could be undone: two minutes or five minutes is probably fine, a day or so to remediate means reconsider, and anything critical keeps a human in the loop. Angelina says that as a solo founder she would not go after enterprise to start. Joseph argues the other side, pointing to patch management across 5,000 machines, to Opus, which by his account was acquired about 10 weeks in, and to the standards buyers will hold you to, such as ISO 42001. On jobs, by his account Microsoft saved $500 million with AI in customer service, and he argues the lesson is to empower the people you have rather than reduce headcount, especially where regulations such as GxP and FFIEC require controls and separation of duties. The episode closes with the two of them announcing a problem bounty board on TwoSetAI. We discuss: * Why his creative kids did not want him to write a book on AI video * Automated attackers, and his estimate of around 43 days to remediate some vulnerabilities * Bounded autonomy: what to hand an AI agent, judged by how fast you could undo it * Making your first AI videos for free in Google Colab * Inside the book, from data provenance and diffusion transformers to safe products * Which models are IP safe, and why your key frames need permission too * A digital paintbrush, and the book's first title, Making Machines Dream * Tokenizing the physical world, and his second O'Reilly book, on physical AI * A digital twin of a mouse model, and cutting the pharmaceutical cycle in half * Why AI video still struggles with consistency * The 3D scene box, Gaussian splatting and open world models * What is left for creators that AI cannot replace * John Lasseter's first job on the Jungle Cruise ride, and the empathy he asks of executives * The minimum stack: a small AMD box with 128 gigs of shared memory * Follow your passion, find your community, then fill a need in it * Should a solo founder sell to the enterprise? * ISO 42001, the NIST AI Risk Management Framework, and enterprise governing bodies * Value streams from W. Edwards Deming to The Goal and The Phoenix Project * A problem bounty board for builders * Long-term thinking, relative to the speed of the foundation models * Is AI cutting jobs? His reading of Microsoft's $500 million, and why regulated roles only compress so far * Stick to your ethics: his advice to be a better founder Joseph Enochs * LinkedIn: https://www.linkedin.com/in/josephenochs/ * Enterprise Vision Technologies: https://www.evtcorp.com/ * Video Generation with AI (O'Reilly): https://learning.oreilly.com/library/view/video-generation-with/9798341653337/ * His writing: https://itrevolution.com/author/joseph-enoch/ * Mentioned in the episode: LTX, ComfyUI, LTX Desktop and InvokeAI; Cal AI; ISO 42001 and the NIST AI Risk Management Framework; The Phoenix Project by Gene Kim and The Goal by Eli Goldratt; the HDF Group, where he sits on the board Timestamps 00:00:00 Why his creative kids didn't want this book 00:01:02 Meet Joseph Enochs 00:01:47 Why do companies take 43 days to patch a threat? 00:03:51 How much should you let an AI agent do for you? 00:06:26 Can you make AI videos for free? 00:08:09 Which AI video models are safe to use commercially? 00:10:08 Why a chief AI officer wrote a book for his kids 00:13:24 Could a digital twin of a mouse speed up drug trials? 00:18:33 Why do AI videos still look fake? 00:22:59 What's left for creators that AI can't replace? 00:29:21 What's the cheapest setup to ship an AI video product? 00:30:37 What should a solo builder build first? 00:32:05 Should a solo founder sell to the enterprise? 00:35:49 How do you find a problem worth solving? 00:38:27 What if enterprises posted their problems for builders? 00:39:48 Does long-term thinking still matter for founders? 00:42:58 Is AI actually cutting jobs? 00:48:02 One piece of advice to be a better founder 00:49:07 Outro Get full access to Decoding AI Builders By Angelina Yang at angelinayang.substack.com/subscribe

  2. Oct 12

    How Fast AI Learned Math | Phillip Kerger of UC Berkeley

    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 Get full access to Decoding AI Builders By Angelina Yang at angelinayang.substack.com/subscribe

  3. Oct 8

    MrBeast's 10% Rule, Applied to Podcasts | Rox Codes of Flightcast

    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 Get full access to Decoding AI Builders By Angelina Yang at angelinayang.substack.com/subscribe

  4. Oct 5

    The One Question to Ask About Your Company's AI Agents | Hector Hernandez of Speakeasy

    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 Get full access to Decoding AI Builders By Angelina Yang at angelinayang.substack.com/subscribe

  5. Oct 1

    The AI Bubble Has an Upside | Rachel Chalmers of Generationship

    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 Get full access to Decoding AI Builders By Angelina Yang at angelinayang.substack.com/subscribe

  6. Sep 28

    You Don't Need AI Friends | Xuan Zhao of Flourish Science

    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 peopl

  7. Sep 27

    When Your AI Starts Acting as You | Sam Kececi of The Sentience Company

    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 Get full access to Decoding AI Builders By Angelina Yang at angelinayang.substack.com/subscribe

  8. Sep 21

    What Engineers Do Now That AI Writes Most of the Code | Murali Swaminathan of Freshworks

    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 Get full access to Decoding AI Builders By Angelina Yang at angelinayang.substack.com/subscribe

See All (57)

About

Behind every AI company is a founder who almost quit, made a bet no one believed in, or figured something out that changed everything. TwoSetAI tells those stories through long-form interviews with the people building the future. Hosted by Angelina Yang. angelinayang.substack.com

Information