Harj compares Mercor with Triplebyte, the recruiting startup he ran. Triplebyte spent years building a labeled dataset and relied on human interviewers, while LLMs let Mercor evaluate candidates from day one and expand beyond software engineers. He suggests looking for marketplaces where AI can take over a human evaluation role or simplify the number of sides involved.
Startup Ideas You Can Now Build With AI
Speaker
Y Combinator partners and guests discuss AI-enabled startup ideas, recruiting, education, and tech-enabled services on The Light Cone podcast.
The speakers describe personalized tutoring and grading as work AI can make more accessible. Examples include RevisionDojo’s tailored exam prep and Adexia’s assignment grading agent. They note that when an AI learning product approaches the value of a human tutor, parents may pay more than they would for a low-completion self-study app.
The speakers revisit full-stack startups that combine software with a service, such as Triplebyte’s recruiting operation and Atrium’s law firm. Earlier versions struggled with low gross margins, staffing needs, and operational complexity. They argue that agents could now handle much of the service work, making some full-stack companies operate more like software businesses.
AI agents still need deployment tools, evaluations, and other infrastructure. The speakers recall that many MLOps startups struggled before useful models had enough demand, while Replicate gained traction with image diffusion models and Ollama with Llama. Their examples suggest that infrastructure ideas can be early for years, then become valuable when the underlying technology catches up.
The speakers question whether customer discovery and selling before building should remain the default advice for founders. They propose exploring new AI capabilities directly, using prompts, data, evaluations, ingenuity, and taste to discover useful products. They also observe that many established companies have yet to change their roadmaps or start internal AI projects.
Takeaways
- Look for marketplaces where AI can automate evaluation or remove a costly human intermediary.
- Test whether AI can deliver tutor-level value, since that may support a higher price than self-study software.
- Include gross margins and operational complexity in the plan for any tech-enabled service.
- Infrastructure bets may need patience until models and customer demand are ready.
- Explore new AI capabilities hands-on; useful startup ideas may emerge from what the technology can newly do.
Mentioned in this talk
Search inside this talk
What was claimed 20
- feature
Google: Gemini 2.5 Pro has a million-token context window“a million-token context window in Gemini 2.5 Pro”
YouTube Unverified1 source· Talk
- foundedHarj Taggar founded Triplebyte
“I ran a recruiting startup, Triplebyte, for almost five years.”
YouTube Unverified1 source· Talk
- featureTriplebyte: Curated marketplace that evaluates engineers
“build a really curated marketplace that evaluates engineers”
YouTube Unverified1 source· Talk
- feature
Mercor: Uses LLMs to evaluate candidates from day one“they could do the evaluation piece on day one using LLMs”
YouTube Unverified1 source· Talk
- metricTriplebyte: Raised something like $50 million
“Triplebyte raised something like $50 million.”
YouTube Unverified1 source· Talk
- previously atHarj Taggar previously at Nianticran engineering teams
“When I used to run engineering teams at Niantic”
YouTube Unverified1 source· Talk
- featureApriora: AI agents run screening for technical interviews
“AI agents that run screening for technical interviews”
YouTube Unverified1 source· Talk
- featureRevisionDojo: Exam preparation tailored to students’ learning journeys
“helps students do exam prep... tailored to their learning journey”
YouTube Unverified1 source· Talk
- featureAdexia: AI agent helps teachers grade assignments
“Adexia is an agent that’s very good at helping teachers grade assignments.”
YouTube Unverified1 source· Talk
- feature
OpenAI: Consumer AI product with paid monthly subscriptions“5% or 10% of users want something enough to buy a $5, $10, or $20 monthly subscription. That's basically what OpenAI is doing.”
YouTube Unverified1 source· Talk
- feature
Speak: Personalizes language learning“They focused on personalizing language learning”
YouTube Unverified1 source· Talk
- uses
“When GPT-3 and GPT-3.5 started coming out, they were early adopters.”
YouTube Unverified1 source· Talk
- featureGoogle: Gemini 2.5 Pro performs as well as or better than o3 on some tasks
“Gemini 2.5 Pro is as good, and in some cases better, than o3”
YouTube Unverified1 source· Talk
- featureGoogle: Gemini can summarize YouTube videos and handle multimodal tasks well
“If I need to summarize YouTube videos, it’s very good at that; it’s really good at multimodal tasks.”
YouTube Unverified1 source· Talk
- integrates with
“You now have an AI system in all of your chats”
YouTube Unverified1 source· Talk
- metricTriplebyte: $20 million annual run rate, reaching $24 million within a few years
“we got to a $20 million annual run rate—$24 million within a few years.”
YouTube Unverified1 source· Talk
- metric
WeWork: Margins were not there; it did not have tech margins“The margins weren't there; it didn't have tech margins.”
YouTube Unverified1 source· Talk
- feature
Replicate: Model deployment tooling“They were building this thing in obscurity for two years until the image diffusion models came out”
YouTube Unverified1 source· Talk
- feature
Ollama: Helps developers run open source models locally“the easiest way for any developer to run open-source models locally”
YouTube Unverified1 source· Talk
- feature
Deepgram: Speech-to-text technology“built speech-to-text technology”
YouTube Unverified1 source· Talk
