TwoSetAI: Decoding AI Builders

From a data platform for enterprises to an AI agent that answers questions about your data in plain language: that is the road that led to Wren AI. Howard Chi, its CEO and co-founder at the time of this conversation, says his team helped enterprises connect their data silos and still heard the same complaint day to day: people had to learn SQL. William, the CTO and co-founder, has spent his career in data infrastructure and open source, working with big frameworks such as Hadoop and Spark.

In this episode, Howard and William join TwoSetAI to explain why, in their view, the hard part of text-to-SQL is not writing SQL. AI is good at generating queries and fixing bugs, Howard says. What goes wrong is the interpretation of what the user meant, because every company has its own definitions. His example: in this month's total payment for each customer, does a payment include refunds or cancelled orders? Wren AI's answer is a semantic layer that a data team fills in much as it would onboard a new team member, with documentation, SQL examples and feedback when the AI misreads a question.

You can learn more about Wren AI here.

We go deep on how that works in practice. Howard describes the three things new clients ask first (which data sources connect, whether the output can be trusted, and how to embed it in their own tools) and the feature he calls AI augmented semantics: connect a data source, upload documentation, and let the AI work out the semantics for a person to review. By his account, an early version in which his team filled in descriptions by hand took two or three months, and a later one that generates them from documentation performed better. William explains the step-by-step reasoning users see, why the product itself uses no AI judge, which he says would add latency, and why row level and column level access control should be deterministic and should not rely on prompts or model behavior.

We also get into the harder parts. Angelina, who used to lead a data team, brings the pushback she has heard: nobody ever wrote the data catalog, analysts already use Tableau, Sigma, Power BI or Looker, and senior analysts doubt any AI can write their queries. Howard's reply is that Wren AI is not trying to replace the BI tools or the data team. It takes the repetitive questions, which he puts at 40, 60 or even 80% of the work, while messy data, pipelines, compliance and governance still need people. William adds that people he meets at meetups in Taiwan complain that they are just SQL compilers. The conversation then jumps to 2030: Howard expects BI to become proactive, to be embedded in the tools people already use, and to move from charts to actions, and he predicts that chat will replace the BI interface. Asked whether the dashboard is dead, he says it is still needed, though its format may change. His advice for teams that are just starting is in the title.

We discuss:

* Where Wren AI came from: an enterprise data platform, and people who still had to learn SQL

* Who gets the most value: data teams hit by repetitive requests, then product, marketing and sales teams

* Whether data teams resist a tool like this, and the work Howard says still needs people

* "SQL compilers": the complaint William hears at meetups in Taiwan

* What new clients ask first: data connections, trusting the output, and embedding it in their own tools

* The semantic layer, and what counts as a payment

* Teaching the AI like a new team member: documentation, SQL examples, catalogs and feedback

* What users see: the reasoning steps behind an answer, in a chat interface

* Pushback: nobody wrote the data catalog. AI augmented semantics, with a person reviewing the result

* Two or three months of filling in descriptions by hand, and the version that replaced it

* Pushback: one more tool next to Tableau, Sigma, Power BI and Looker

* Can the results be trusted? William on the first time he used Cursor

* Retraining or prompts: prompt based for now, with fine tuning under experiment

* Senior analysts and very complex queries: teaching the logic in smaller pieces

* Hallucination as default assumptions: how a model defines revenue or active users

* Why there is no AI judge in the product, and what the dry run does instead

* How good text-to-SQL is today, and why Howard says benchmarks are not the real world

* What generative BI should look like, and what happened when the chart guardrails came off

* A good first use case, and how much data maturity you need

* Security that does not rely on prompts: row level and column level access in the semantic layer

* 2030: an always-on analyst, BI inside every tool, and charts that turn into actions

* Is the dashboard dead? Howard's bold prediction about chat

* The advice: start before the data is perfect

Howard Chi and William

* Wren AI: https://getwren.ai/

* Mentioned in the episode: Cursor, which William compares the experience to; and Tableau, Sigma, Power BI and Looker, the BI tools Angelina says her team used

Timestamps

00:00:00 Intro

00:02:06 How did Wren AI start

00:03:08 Who are the teams getting the most value from the tool you're building

00:05:42 Are data teams motivated to have a text-to-SQL tool

00:08:52 What's the most common question clients ask you

00:10:11 How does your semantic layer work

00:12:07 How do new data team members receive semantic information

00:12:51 Do users see reasoning steps and give feedback

00:13:39 How do you handle a lack of context or pushback

00:16:58 How do you respond to "it's the same thing"

00:18:23 Can the results really be trusted

00:20:47 Does feedback come from multiple users

00:21:24 Are you retraining models or using prompt engineering

00:22:30 Do clients give similar feedback

00:25:14 How do you handle domain-specific needs like legal or healthcare

00:25:48 Is hallucination still a problem

00:26:46 Is there a fallback when the model struggles

00:27:58 Are you using AI agents to evaluate reasoning

00:28:56 How good is text-to-SQL today

00:30:29 What's the biggest technical challenge you've faced

00:35:23 What's a good first use case for generative BI

00:35:58 Is there a data maturity requirement to use Wren

00:36:33 Is a customer success layer still needed

00:36:52 How do you position Wren among competitors

00:38:48 What does your semantic layer architecture look like

00:40:09 Are knowledge graphs part of your system

00:41:25 How will Wren AI evolve by 2030

00:43:26 Are dashboard apps dead in the future

00:44:27 Thoughts on the time travel question

00:47:21 Any question we missed

00:47:36 Advice for companies exploring text-to-SQL

Yuzheng Sun sees the job market from both sides. At the time of this conversation, in the first quarter of 2025, he has a YouTube channel and a community of data scientists, teaches an AI course on Maven, and works at Statsig, a startup he describes as a testing platform. Online he is known as Kedaibiao. From his community he hears the horror stories and the success cases, like the member who landed a job after half a year of searching. From his company, which he says is doubling headcount, he sees the other side: over 3,000 applicants for each position it posts.

In this episode, Yuzheng joins TwoSetAI to explain why the traditional way of applying for a job has stopped working. He passes on what a hiring manager at Databricks told him: because everyone today applies to everything, the precision and recall of the process have changed and it no longer carries signal. His answer is not to give up on the resume, which he says still matters. It is to change the order of the work and where a job seeker's time goes.

You can find Yuzheng's AI course on Maven here.

We go deep on what the hiring side looks like. By his account the company screens with high precision filters (are you local to an on-site company, do you have the relevant experience, did you read the job description), fewer than 100 applicants get a serious look, and each of those gets two to five minutes. He compares resumes to an online dating app where most profiles look the same and about 1% are very popular, explains what makes an AI project credible (not a class project; he suggests looking at the latest YC batch), and points out that nobody has 10 years of large language model experience, so everyone starts at about the same time. For a company like his, he says, a side project that applies AI to A/B testing in a useful way would get a lot of attention.

We also get into the harder parts: his view that data scientist is a made-up role whose value is the delta between the decisions a company makes with one and without one, why he thinks big tech hired too many people and will not be where the growth is, and why tools that automate applications make a candidate the same as everyone else. He walks through the framework he credits to Naval (build, sell, get lucky) and four kinds of luck, including the one that brings opportunities to you, and argues that technical people were trained by exams not to broadcast their work. He closes on the builder's mindset that he and his co-teacher teach in their Maven course, where he calls himself the seller of the pair.

We discuss:

* Over 3,000 applicants per posted role, up from several hundred a year earlier, by his account

* What a Databricks hiring manager told him about precision, recall and signal

* Why job openings in analytics now ask for large language model experience

* "It's fake": why he calls data scientist a made-up role

* The construction worker who says "I build houses," and what the result of a data scientist is

* Empire building, the pendulum and why he does not expect big tech to hire aggressively

* Where he sees the hiring: new companies like OpenAI and Anthropic

* Why the old way of applying fails when every resume looks the same

* Product market fit for job seekers: "applicant and market fit"

* The multipliers he puts on a tailored resume and on talking to insiders

* How 3,000 resumes become fewer than 100 serious looks

* The online dating app analogy and the popular 1%

* AI projects on GitHub, and why class projects do not count

* How he would split a job seeker's time, and what most people do instead

* Why he does not like automating applications

* Same format, different substance: the one page resume

* Why advice that worked three years ago does not apply today

* Titles are made up: build, sell, get lucky, and the 3x3 matrix

* Four kinds of luck, from blind luck to luck by attraction

* Getting inbound: publishing projects and sharing the building process

* Why exams teach technical people not to broadcast their work

* Builder plus seller: teaching a Maven course with a partner

* The builder's mindset, and acting as the manager of AI

Yuzheng Sun

* AI course on Maven: https://maven.com/kedaibiao/genai

* Mentioned in the episode: his interview with a hiring manager at Databricks; the build, sell, get lucky framework and the four kinds of luck he credits to Naval; and the latest YC batch as a source of project ideas

Timestamps

00:00:00 Importance of resumes for data scientists

00:03:17 Data science and engineering roles in the coming years

00:09:43 Big tech companies slowing down hiring?

00:11:10 What should new grads or laid-off engineers do in the market?

00:15:03 Number of resumes being received per role at Statsig

00:16:04 Resume filtering

00:16:25 Do resumes still matter?

00:17:11 Standing your resume out

00:18:28 Adding AI projects and GitHub work in resumes

00:19:19 Showcasing work experience and side projects

00:20:14 Suggestions for someone currently out of work

00:21:44 Resume automation according to job description

00:22:31 Finding unique value as a data scientist?

00:24:37 Do job titles still matter?

00:26:11 Developing the skill of getting lucky

00:27:32 How to get inbound?

00:30:07 What skills matter most in the AI era?

00:31:38 Mindset in the Maven AI course

00:34:08 Summary of Maven AI course

From a PhD at Purdue to a hiring startup that had launched three months earlier: that is where Jerry Wang was at the time of this conversation. He is a co-founder of OpenJobsAI and the one responsible for the tech side, next to a business lead and a product designer. The company is building agents for both sides of hiring, so that applicants no longer have to hunt for jobs and companies no longer have to hunt for candidates. Its slogan, in his words, is "where dream jobs find you."

In this episode, Jerry joins TwoSetAI to look at the engineering job market from the hiring side. His main point is that software and AI roles are converging. It is not a replacement of traditional software engineering by AI roles, he says, but a fusion of skill sets, and many companies are looking for hybrid people who can do the engineering part and the AI part. He also shares the question he thinks everyone in tech should ask themselves every day.

We go deep on what hiring teams ask for now. Front-end and back-end titles still exist, Jerry says, but the job descriptions add AI knowledge, from writing code with tools like Cursor up to running a large language model on multiple GPUs. Copilots are welcome on the job and forbidden in interviews on his platform, and he explains how a hiring manager can still test whether a candidate knows the tools. He names the two technical skills he sees most in demand, model inference and vector databases, and argues that a general embedding model may fool you in a specific area such as hiring unless you fine-tune your own. He says where he thinks end-to-end AI engineers should be trained, what he would pay one in Silicon Valley, and cites a stat he saw that AI-related positions doubled compared to two years ago. On the company side, he says agents could save maybe 90% of the regular work of HR.

We also get into the harder parts: technical debt from engineers who push copilot code without checking it, which he treats as a matter of code review and training; why a team made up only of AI experts could run into problems; and a job market he calls a little better than the previous two years but still not comparable to before the pandemic. His advice to new grads is plain: do open source projects or start something small, look at startups and traditional industries when big tech is the most competitive door, and find a job first. He gives his predictions for the next three to five years, including which roles he expects to decrease, calls communication the most important thing in a team, and tells AI engineers to read the Bitter Lesson. At the end he runs through the three startups he tried during his PhD and why he gave each one up.

We discuss:

* What OpenJobsAI is building: agents for both sides of hiring, and "where dream jobs find you"

* How much of HR's regular work he thinks agents can take over

* The founding team, three months after launch

* Software and AI roles: a fusion of skill sets, not a replacement

* What an entry-level engineer needs to know about AI, and what an expert needs

* Front-end and back-end titles with AI added to the job description

* Copilots on the job vs. copilots in the interview

* How a hiring manager can test whether you know tools like Cursor

* Technical debt from copilot code, and why he puts reliability before speed

* Startups that hire only AI engineers, and the problems he sees with that

* What he would pay an end-to-end AI engineer in Silicon Valley

* Which degree program he thinks should produce full-stack AI engineers

* The job market now, compared with the previous two years and with before the pandemic

* The new grad paradox: no job without experience, no experience without a job

* Startups vs. big tech: more to learn, more stress

* "Find a job first": his advice in a hard market

* His predictions for the next three to five years

* What non-technical people need to learn about AI

* "Maybe in the future everyone is the product manager"

* Hiring budgets moving toward AI roles, and the value of domain knowledge in healthcare or finance

* How to stand out when companies screen resumes with an ATS

* The two technical skills he sees most in demand, and why a general embedding model may fool you

* Soft skills: explaining your code, not only leaving comments

* The Bitter Lesson, and what happened to the three startups he tried during his PhD

Jerry Wang

* Mentioned in the episode: the Bitter Lesson, which he tells AI engineers to read; Cursor and Claude, the coding tools he names

Timestamps

00:00:00 Intro

00:04:43 Where do you see these engineering roles heading in the next one to two years?

00:06:11 Do you have a rough estimate of what percentage of those engineer roles are requiring some level of AI capabilities?

00:06:58 Are these roles still distinguished in a way?

00:07:45 Do companies expect candidates use copilots?

00:09:54 Are using copilot causing more tech debt?

00:12:49 Is there any issue with hiring only AI eng roles?

00:14:51 Salary trend

00:17:03 How is AI impacting the job market

00:21:16 What's the difference between working for a startup vs. big companies?

00:22:03 Advice to new grads

00:25:56 Perhaps in the future, everyone is the product manager.

00:26:23 Hiring budgets shifts

00:28:03 How to stand out in job applications?

00:29:30 What tech skills are more in demand?

00:32:52 Soft skills still in?

00:34:58 What is one thing that you wish every AI engineer knew before applying for jobs?

From sales to founder to coach: Yinan Gu is originally from Shanghai, where, by her account, she and her husband founded the city's first cloud computing company in 2008. At the time of this conversation she is a full-time coach in the Bay Area, working with CEOs, founders and executive teams, and a part-time entrepreneur mentor for students at Stanford. She brings more than 20 years of global sales and marketing work and 15 years as an entrepreneur into her coaching sessions.

In this episode, Yinan joins TwoSetAI to explain what a coach is for. A therapist, she says, looks for the root cause in your history. A coach assumes you are okay and works on now and the future, acting as a mirror: a third party who sees you clearly, without judgment. She tells the story of the question her own coach asked when she kept running out of energy, and how the answer changed which clients she takes on.

You can find Yinan on LinkedIn.

We go deep on the eight questions she uses to turn a dream into a practical career goal: who you are, what you have, what you lack, who can help, what you can give them, what you can put aside, what keeps you motivated and, last, is that real? Angelina tries it live with a made-up goal, an AI sales product that reaches 5 million in ARR in two years, and gets sent back to the first question. Along the way Yinan explains why she asks clients to find five people who can help them and to think about what they can give those five, and what she tells students at Stanford: fall in love with the problem, not with your product or your solution.

We also get into the harder parts. When she asks the last question, quite a few people say the goal is not what they want. By her estimate, 50% of the people she sees, highly educated people, are living in other people's expectations, and when Angelina asks how many dare to burst their own bubble, she again says about half want to change. She is plain that living up to someone else's expectations is a choice with its own responsibility, not something for a coach to judge. She also talks about what AI cannot replace, why she counts more than salary in your market value, why she asks clients how much time goes to work and study, health, relationships and play, and why sharing your struggles matters more than sharing only your happiness.

We discuss:

* Why she became a coach: the support that got her through the darkest moment of her entrepreneur journey

* A job with no limit on age, and why she did not start another company

* The question her own coach asked: you are a coach, not a therapist, and the people around you matter

* Choosing clients who are good and want to be better

* What she means by energy

* Therapist or coach: healing a wound, or working on now and the future

* The coach as a mirror, and a third party who sees you without judgment

* Why a coach "only tries to feed the angel"

* A ladder of small goals when the dream is too far away

* The eight questions, starting with who you are

* What you have, what you lack, and who can help

* Finding five people, and why a no should not end the connection

* A live example: Angelina's made-up AI product and its 5 million ARR goal

* What she tells Stanford students: fall in love with the problem

* What you can give first: "Silicon Valley is the culture of giving"

* What you can put aside, and the monkeys on your back

* Consistent motivation, and how it differs from a goal

* The last question: is that real?

* How many highly educated people, by her estimate, live in other people's expectations

* What to do when the goal turns out not to be yours

* Her one tip, and how to know you are ready for a coach

* What AI cannot replace, and what she counts in your market value

* Work and study, health, relationships and play: where your time goes

* Sharing your struggles and not only your happiness: "the power of vulnerability"

Yinan Gu

* LinkedIn: https://www.linkedin.com/in/yinan-gu/

* Mentioned in the episode: the design thinking framework she uses with clients

Timestamps

00:00:00 Why even elite performers need coaches

00:15:28 The 8-Question Framework to design your life

00:31:30 Now what?

00:33:31 How many people admit that they are "lying" to themselves?

00:36:45 Am I ready for a change?

00:38:08 Do I need a coach

00:38:32 AI FOMO

00:41:35 Your market value

00:42:26 The first step to get unstuck

00:43:20 Vulnerability

From a habit of taking notes to two books on building with AI: Chip Huyen says her first book, Designing Machine Learning Systems, was an accident. She was teaching a course on machine learning systems design at Stanford, was conscious of her accent, and wrote a lot of notes so that students who did not follow her in lectures had something to go back to. At the time of this conversation her second book, AI Engineering, had just come out, and she was done with the startup she ran for a little over two years and deciding what to build next.

In this episode, Chip joins TwoSetAI to talk about keeping your footing in a field that changes this fast. Her approach is not to chase it. She does not read a lot of news, she says, because news gets old quickly. She follows a few problems she cares about, looks something up when three smart friends mention it, and leans on Lindy's law: the longer an idea or a technology has been around, the longer it can be expected to last. Researching AI Engineering, she found that much of what sits under the new applications is old. Language modeling goes back to the 1950s, and RAG rests on retrieval technology that has long powered search and recommendation systems.

You can find Chip on LinkedIn and GitHub.

We go deep on how the second book was written and where AI is working. When ChatGPT came out, she and her friends had what she calls a bit of an existential crisis, and as both a writer and an engineer she set out to learn what could be automated, interviewing a ton of people along the way. While writing she had about 3,000 conversations with ChatGPT: to understand papers better, to compare a new technique with the one before it, and to brainstorm examples and definitions. On adoption, she says the most successful companies she has seen do not talk about it much, and that the most popular use cases are the ones where the return can be evaluated clearly, such as recommendation systems, fraud detection and coding. She calls evaluation one of the biggest bottlenecks for AI adoption, if not the biggest, and walks through the crawl, walk, run framework she credits to Microsoft: a human in the loop first, then internal use, then external users.

We also get into the harder parts. Chip describes the question she took to her mentors while running her company, whether a dark stretch is a tunnel or a cave, and the lesson she thinks founders are taught too little. She names governance and compliance as one of the biggest blocks to adoption, by the account of people she spoke with, next to the fear of losing jobs and the safety of people who do not yet know what AI can do. Her example is her parents in a village in Vietnam, who could be targets for scams built on fake videos. She argues that schools should not ban AI but work it into the curriculum, gives her predictions for 2025, answers Angelina's question about how any of us adapt, and closes with a cynical note about productivity and the 40 hour week that she asks to have fact-checked.

We discuss:

* What she is working on after her last company: her interests, not a list of to-dos

* Why launching a book is not like launching a startup

* Running a company vs. building a product, and what she wants more of

* Tunnel or cave: the question from her hardest moments as a founder

* The accidental first book: notes written for her Stanford students

* Writing as a way to organize her thinking

* ChatGPT, an existential crisis, and the questions behind AI Engineering

* About 3,000 ChatGPT conversations during the writing, and what she used them for

* Will AI replace human writers? What it is good at, and what stays hard

* Why she does not try to keep up with AI news

* Three smart friends, and Lindy's law

* RAG, language modeling and other ideas that have been around for a long time

* Finding core problems through personal interest and domain expertise

* The gap between demos and good products

* Governance, compliance and trust as blocks to adoption

* Scams, fake videos and the people who are new to AI

* Use cases that work: coding, data analysis, chatbots and companionship

* AI in education: essays to fact-check, debate partners and adaptive learning

* How executives judge the return, and why evaluation is a bottleneck

* Three angles for executives: how much to care, which use cases, and how to upskill their people

* Crawl, walk, run

* Her predictions for 2025

* Optimism, anxiety and how to adapt

* Time to think, and why the working week has not become shorter

Chip Huyen

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

* GitHub: https://github.com/chiphuyen

* Mentioned in the episode: her books AI Engineering and Designing Machine Learning Systems; Lindy's law; and the crawl, walk, run framework she credits to Microsoft

Timestamps

00:00:00 Interview with Chip Huyen

00:05:54 The most important lesson from a startup founder

00:07:27 What is writing for Chip?

00:12:26 How to keep updated in this fast moving space?

00:13:59 Lindy's law

00:15:07 RAG

00:21:03 What AI use cases have been really successful in companies?

00:21:40 Classic use case

00:23:08 Chatbot

00:23:52 Gaming and Education Use Cases

00:25:42 How education is gonna change?

00:29:14 The framework of Crawl, Walk, and Run

One main client, four to five advisory clients and a collective of 18: that is how Holly Chen works at the time of this conversation, as an interim CMO and growth advisor. At any given time she runs marketing for one company as a full-time executive and advises a handful of others, typically Series B to pre-IPO companies building SaaS products for a professional audience. She also runs an advisory collective of 18 former CMOs and heads of marketing, most of them from B2B SaaS.

In this episode, Holly joins TwoSetAI for a go-to-market session with a founder who does not exist. Angelina plays a pre-seed founder whose AI product, text to shorts, launched last week and has four customers, two of them paying, and asks what to do now. Holly's answer starts with two questions to ask up front and with a rule about where to aim: she would rather solve 80% of the pain for 1% of the people than 20% of it for 80% of the people.

You can find Holly on LinkedIn.

We go deep on how a founder finds a first segment and reaches it. Holly explains why a product should start with something as narrow as agencies with fewer than 10 people that repurpose long videos for e-commerce clients, where she would go first to find them, and why founders have to do inbound and outbound at the same time: SEO and content can take five or six months to show anything, and outbound, she says, should start with first-degree and strong second-degree connections. She gets specific about channels for B2C and for B2B, how often to post, what a founder should establish themselves as in public, and the three types of content she would mix, with product posts at maybe 20%.

We also get into the harder parts: why she is cautious about any formula for going viral, and what she calls the retention problem of AI startups, where AI tourists register, find no sustained need for the product and never pay. She says a landing page is for customers and the grand vision belongs in the investor deck, argues for describing use cases and not only features, and sorts competitors into three buckets: direct competitors, incumbents and replacements, such as Excel or not using anything. Her advice is to stay aware of the market without becoming reactive to it. DoorDash, she points out, was not the first one to go into food delivery.

We discuss:

* What an interim CMO and growth advisor does, and how she selects clients

* The advisory collective of 18 former CMOs and heads of marketing

* The role play: a pre-seed AI product with four customers, two of them paying

* How much to customize for the first customers

* The two questions to ask up front

* Why a feature is just a means to an end

* Who to talk to besides the customers who already pay

* Michael Skok's minimum viable segment, which Angelina brings up

* Solving 80% of the pain for 1% of the people, and expanding once that segment is won

* A vision for later, a very specific pain point for now

* Where she would go first to find the segment

* Inbound and outbound: why she says you have to do both

* Why SEO takes time, and what founder-led sales looks like

* First-degree and strong second-degree connections

* The channels she suggests for B2C and for B2B

* How often to post

* What to establish yourself as, and the three types of content

* Going viral, AI tourists and the retention problem

* A landing page for customers, an investor deck for the grand vision

* Use cases and features

* Three buckets of competitors: direct, incumbents and replacements

* Why she says not to be reactive, and the DoorDash example

Holly Chen

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

* ExponentialX: https://exponentialx.com

* YouTube: https://www.youtube.com/@ExponentialX-Collective

* Mentioned in the episode: Michael Skok's minimum viable segment; LinkedIn Sales Navigator; Wordware, the AI tool Angelina brings up

Timestamps

00:00:00 Intro

00:05:20 What audience am I building it for? And what pain point am I solving for them?

00:06:00 "A feature is just a means to an end, xxx is more important"

00:06:48 "It's the pain point, not ..."

00:09:46 "the first place I will go is just..."

00:12:09 "for outbound, I would say you want to be methodical..."

00:17:06 "You want to establish yourself as a..."

00:17:28 "there are 3 types of content. one is..."

00:19:16 How to go viral?

00:26:12 Use Cases vs Features

00:27:38 How to think about competitors?