Y CombinatorHow Internships Seed Most of YC's Billion-Dollar Startups
Internships at Cohere and going undercover as a medical biller surface real problems; hackathon ideas miss these, but outsourced jobs signal the next wave.
CHAPTERS
- 0:00 – 1:45
YC AI Startup School announcement + why this episode matters
Garry opens with an announcement for YC’s AI Startup School and transitions into the episode’s goal: helping technical builders find strong AI startup ideas. The hosts frame this moment as unusually promising for ambitious new companies.
- •YC AI Startup School details: date, speakers, who can attend, travel support
- •Transition from announcement to the core topic
- •Framing: now is a special time to start AI companies
- •Episode promise: share internal YC idea-finding “tricks” publicly
- 1:45 – 3:24
Stop chasing hackathon wrappers: aim for harder, more ambitious first versions
The group explains a common failure mode: founders gravitate toward ideas that are easy to build in a weekend. YC’s advice is to lean toward ideas that are meaningfully harder to ship, because that difficulty often correlates with defensibility and real value.
- •Default bad idea: bandwagon/hackathon-style “wrapper” thinking
- •Founders’ subconscious bias toward easy builds
- •Best ideas often require a harder first version
- •Ambition and execution difficulty can be a positive signal
- 3:24 – 4:30
Two reliable paths to ideas: deep introspection or getting ‘out of the house’
Garry outlines a blueprint: either mine your unique background for unfair advantages, or go immerse yourself in unfamiliar industries to discover real problems. Great AI ideas often come from first-principles understanding at the edge of what people do today.
- •Idea generation requires either internal depth or external exploration
- •“Get out of the house” to find underserved problems
- •Look for edge-of-world knowledge from research, jobs, or lived experience
- •First-principles problem understanding as a generator of AI opportunities
- 4:30 – 6:53
Founder-market fit from prior work: Salient (auto loan processing) + Diode (PCB copilot)
Diana shares examples where founders’ prior roles revealed niche, high-value workflows ready for AI automation. These work because the founders had rare domain exposure or an unusual intersection of skills.
- •Salient: Tesla finance ops insight → AI voice agents for loan processing/debt collection
- •Manual, outsourced ops as an AI-automation wedge
- •Diode Computer: software+hardware skill intersection enables circuit-board copilot
- •LLMs parsing datasheets/verification as a hardware-engineering unlock
- 6:53 – 8:03
Operate like a product-focused researcher at the edge: ‘If not us, then who?’
Garry argues that successful founders resemble PhD-level explorers: they push to the frontier of understanding, then build a product instead of publishing a paper. The chapter emphasizes the commitment test—why your team is uniquely positioned to solve it.
- •Startups as applied frontier research: edge understanding → productization
- •Founder-market fit can be N=1 unique timing and background
- •Commitment heuristic: “If not us, then who?”
- •Turning deep domain pain into something people want
- 8:03 – 10:45
More inside-edge examples: Spur (AI QA) + DataCurve’s pivots and the internship insight
Jared and Garry show how concrete work pain (like complex frontend testing) leads directly to viable products, while shallow personas lead to dead ends. DataCurve illustrates how pivots often resolve by returning to genuine expertise—sometimes from an internship.
- •Spur: founder’s Figma experience → AI agent that writes/maintains tests
- •DataCurve: started as “Uncle GPT” wrapper → no willingness to pay
- •Second attempt: ‘AI for product managers’ lacked founder experience
- •Breakthrough by revisiting Cohere internship: build what prior team truly needed
- 10:45 – 14:00
Founders underestimate their own expertise; internships and picking ‘edge’ companies matter
The hosts discuss how founders often dismiss the domains they know best as “boring,” then chase shiny ideas without depth. They highlight internships and deliberately choosing cutting-edge employers as a repeatable way to gain access to future-defining problems.
- •Pivot coaching heuristic: start by identifying founder expertise
- •Common trap: boredom with hard-won domain knowledge
- •Internships frequently seed billion-dollar YC ideas
- •Be picky: work at ‘bleeding edge’ orgs (e.g., Cohere, Scale) to see the future early
- 14:00 – 18:43
Ideas that capture imagination + removing ‘blinders’ (Can of Soup, Happenstance, EasyDubs)
Jared shares a story about choosing ideas that inspire long-term motivation, not manufactured B2B sameness. They connect this to Paul Graham’s ‘blinders’ concept and give examples of ambitious AI-native products enabled by new search and translation capabilities.
- •Can of Soup: advice—build what captures the human imagination, not generic SaaS
- •Happenstance: smarter people-search via LLMs + vector search + SQL
- •PG essay: ‘blinders’ hide ambitious ideas from consciousness
- •EasyDubs: real-time universal translation as a once-too-ambitious idea now feasible
- 18:43 – 22:24
When you lack expertise: build it by shadowing, family access, and on-site immersion (Egress Health)
Harj explains how to generate ideas when you don’t have a strong domain—by embedding yourself in a real workflow. Egress Health found its wedge by spending a day inside a dental office and spotting insurance/admin tasks ripe for LLM automation.
- •Shift mindset: act like researchers; go acquire domain understanding
- •Egress Health: dental office shadowing → automate insurance/admin back office
- •Use family/friend connections to get access others can’t
- •AI agents raise the ROI: replacing labor is bigger than niche SaaS
- 22:24 – 26:12
Going undercover for extreme access: logistics + medical billing automation playbook
Garry describes ‘undercover’ approaches to learn closed-door industries, from friendly networking to literally taking a job inside the workflow. Examples include trucking logistics coordination and a founder who became a medical biller to automate the role safely with local tooling.
- •Happy Robot: personable founders gained insider access to trucking logistics pain
- •Undercover tactic: take the job you want to automate to learn the workflow
- •Medical billing example: founder employed as biller, built automation on own machines
- •Why it works now: capable LLMs, strong laptops, smaller models, rapid prototyping
- 26:12 – 28:34
Undercover to the edge of society: police paperwork as an AI opportunity + job-board ‘alpha’
The conversation extends the undercover idea to civic workflows: ride-alongs revealed police time lost to paperwork. Jared then offers a practical method: scan job boards for repetitive remote ‘clerk/analyst’ roles as targets for AI automation.
- •Able Police: crime-driven investigation → police paperwork reduction with LLMs+CV
- •Ride-alongs and first-principles observation reveal root causes
- •Paperwork burdens are widespread, not just local
- •Indeed.com tactic: search for remote clerk/analyst roles to find automatable jobs
- 28:34 – 32:51
More idea sources: shadow friends, follow outsourcing signals, and replace consultant-heavy products
Harj and Diana share additional heuristics: shadow friends in boring roles, look for work outsourced to low-wage countries, and target markets where expensive consultants are required because tools don’t work. These are reliable signals of high-value automation opportunities.
- •Sweet Spot: automating government contract monitoring and bid generation
- •Outsourcing signal: Lilac Labs automates drive-thru order taking
- •Consultant signal: Automat aims to replace brittle RPA deployments like UiPath consulting
- •Temporary/outsourced/consultant-heavy work indicates high automation leverage
- 32:51 – 37:21
Live at the edge + build anything: friends, technical communities, and shipping create new idea surface area
The hosts emphasize being early users/builders of new tech and surrounding yourself with smart builders who surface problems. They also note that simply shipping products—even if the first idea fails—creates user exposure and new idea pathways.
- •Meta lesson: ‘Live at the edge and notice what’s missing’
- •PreyDB: friend need → Postgres↔vector workflow → pushing pgvector to replace vector DB/Elastic
- •Reducto: YC/founder network revealed RAG chunking/PDF extraction pain
- •Juicebox → pivot to PeopleGPT: shipping creates domain learning and new opportunities
- 37:21 – 43:48
Fundraising, competition fear, and persistence: trust reality, not toilet-scroll narratives
Garry contrasts real user traction with investors’ pattern-matching and doom-scrolling; founders should trust what they’ve validated in the field. They close by discussing crowded markets, technical differentiation (GigaML), and why longer pivot cycles are now normal in fast-moving AI.
- •If users pay, that’s real—don’t be derailed by early investor ‘no’s
- •Crowded-space anxiety can hide great opportunities; execution quality matters
- •GigaML: deep technical capability beat ‘crowded customer support AI’ skepticism
- •AI’s rapid change supports longer searching/pivoting; keep going—luck breaks often come soon
