WhoBuilt

How To Get AI Startup Ideas

Feb 7, 2025· 44 min· 293K views·Watch on YouTube
#Product Discovery#AI Agents#Workflow Automation#Product-Market Fit

Speakers

Garry TanPresident and CEO at Y Combinator
Jared Friedmangroup partner at Y Combinator
Harj
Diana HuManaging Partner at Y Combinator

Y Combinator president Gary Tan and partners Jared Friedman, Diana Hu, and Harj discuss how founders can find AI startup ideas on The Lightcone.

Look beyond easy ideas

The hosts share advice for technical people who see promise in AI but have not found an idea they want to pursue. Jared Friedman warns that founders often gravitate toward ideas they can build in a weekend, while promising ideas may take more effort. Gary Tan recommends looking deeply into your own experience or going out to learn how other people and industries work.

Start with hard-won experience

Diana Hu describes founders who found problems through specialized experience: Salient's founder saw manual loan collection work at Tesla, and Diode Computer's founders combined hardware and software knowledge to build a circuit board copilot. The hosts also discuss Spur, whose founder saw front-end testing challenges at Figma, and Datacurve, whose founder drew on an internship at Cohere to build data tools for AI. They recommend looking for problems close to expertise built through jobs, research, or internships.

Get close to real work

The hosts suggest visiting workplaces, shadowing people, or taking a job to understand tasks that software could improve. Examples include S&S Health, which built an AI back office for dental practices after a founder observed insurance administration, and Sweetspot, which automated government contract discovery after a founder's friend described the manual process. They also point to outsourced and temporary remote jobs as places to investigate repetitive computer work.

Build knowledge by trying things

The hosts encourage founders to use emerging technology and talk with people who build at the frontier, so they can notice problems others have not solved. PreDB found a database and search opportunity through conversations with technical founders, while Reducto learned about document extraction needs from builders working on retrieval applications. Juicebox's founders also developed user knowledge while building a freelancer marketplace, which led them to PeopleGPT, a search product for recruiters.

Give ambitious ideas time

Founders should not rule out an idea just because competitors exist, the hosts argue. Giga moved from fine-tuning models to customer support and won an enterprise opportunity by delivering on a difficult technical problem. The hosts note that it took about a year to find the right idea, and say changing AI capabilities can create new opportunities over time.

Takeaways

  • List problems you understand from work, research, or internships before chasing a trend.
  • Spend time with users and observe their daily workflows to uncover costly manual tasks.
  • Use your existing relationships to get access to industries and underserved problems.
  • Keep building with new tools and discuss real problems with people at the technical frontier.
  • Do not reject a promising idea solely because competitors exist or early attempts take time.

Mentioned in this talk

Y CombinatorSalientTeslaDiode ComputerSpurFigmaDatacurveCohereScale AIDavid AICan of SoupSubstackLinkedInHappenstanceEasyDubsS&S HealthHappy RobotAbel PoliceSweetspotLilac LabsUiPathAutomatePineconePreDBReductoJuiceboxGigaMLZeptoThe Light ConePeopleGPTlarge language modelsAI agentssemantic searchpgvectorVector RAGfounder-market fitstartup pivotingcustomer discoveryAriGabrielChris

Search inside this talk

What was claimed 25

Related talks