AI summary by TwinMind
Startup Ideas You Can Now Build With AI
Y Combinator
40m 46s
AI Infrastructure and Tooling Opportunities
- Significant tooling and infrastructure still needs building for AI deployment and agents
- Living at edge of future technology leads to discovering startup ideas naturally
- Apply right prompts, data sets, ingenuity, evals, and taste to get magical output
- Million token context window in Gemini 2.5 Pro creating new capabilities
Recruiting Startups Transformed by AI
- Triple Bite experience (2015): recruiting marketplace requiring 3-sided model
- Companies hiring engineers, engineers seeking jobs, contracted interviewers
- Spent years building software for technical interviews to create labeled dataset
- Took 3-4 years before could run ML models
- Meror: modern AI-powered recruiting marketplace
- Can do evaluation on day one using LLMs
- No need to build large labeled dataset
- Easily expand to other knowledge work types
- Hundreds of millions invested in recruiting marketplaces previously with poor overall results
- Psychological challenge: investor skepticism from past failures (Triple Bite raised $50M, main competitor $100M+)
Technical Interview Screening
- Apriora (Winter 24 YC company): AI agents for technical interview screening
- Solves pre-screening workload problem for engineers
- Large companies using it successfully
- Can do sophisticated evaluations vs simple filtering
- Expands market beyond international/college applicants to senior engineers
Education and Personalized Learning
- Personalized learning: long-standing internet dream now possible with AI
- Revision Dojo: AI-powered exam prep with high DAUs and power users
- Tailored flashcards that students actually enjoy using
- Adexia: AI grading tools for teachers
- Addresses major teacher churn cause (hate grading assignments)
- Private schools more nimble adopters than public schools
- Speak: language learning company that started pre-LLM
- Early believers in personalization before AI capabilities existed
- Doubled down when GPT-3/3.5 emerged
- Strong growth trajectory with many MAUs
Consumer AI Economics and Distribution
- Intelligence costs still require charging users but decreasing rapidly
- Distillation from larger to smaller models working effectively
- Consumer AI may finally be viable soon with pennies/10-15 cents per user cost
- Could return to freemium model from web 2.0 era
- Better AI products don't automatically get more distribution
- Parents will pay significantly more for AI tutors that match human performance vs basic learning apps
Platform Neutrality and Big Tech Competition
- Need for platform neutrality similar to historical net neutrality fights
- Users should choose voice assistants (not forced Siri/Google Assistant)
- Government historically intervened for free markets (Windows browser choice, search engine selection)
- OpenAI hired Instacart CEO, showing focus on application layer
- Gemini 2.5 Pro as good/better than O3 for various tasks but low consumer adoption
- Google's organizational issues: two different Gemini APIs (DeepMind vs GCP)
- Classic innovators dilemma: Google can't replace google.com with Gemini Pro without losing 80% revenue
Meta AI Integration Issues
- Meta AI embedded in WhatsApp feels invasive
- Can't access user's Facebook data ("Sorry, as an AI, I don't have access")
- Reminds of original Facebook newsfeed launch controversy
- Lacks design taste in implementation
Full-Stack Startups 2.0
- 2010s tech-enabled services wave (Triple Bite, Atrium) generally failed
- Poor gross margins, complex operations, scaling difficulties
- Required constant fundraising vs sustainable profitability
- Zenefits example: relied on hiring more people vs building software for gross margin
- Rippling lesson: force engineers to do customer support to build better software
- WeWork extreme example: no tech margins, "community adjusted EBITDA"
AI-Enabled Full-Stack Revival
- AI agents can now handle knowledge work previously requiring human ops teams
- Legora: fastest growing YC company building AI tools for lawyers
- Eventually agents will do all legal work, becoming largest law firm
- Atrium 2.0 potential: Justin Kan noted AI now good enough for legal automation
- Virtual assistant marketplaces could work with AI + ability to hire real people
- Full-stack companies can now look like software companies under the hood
ML Ops Evolution
- 2019: More ML tooling applications than actual ML product applications at YC
- Problem: building tools for non-existent market (ML didn't work yet)
- Replicate (Winter 20): nearly stopped working during pandemic, exploded when diffusion models launched
- Ollama: quietly worked on open source model deployment, took off with Llama release
- Deepgram: physics PhDs working on speech-to-text since 2016, exploded with voice agents
Startup Advice Shift
- Pre-AI era: lean startup methodology, customer discovery, fail fast
- AI era: follow curiosity, use interesting technology, live at edge of future
- College students still receiving outdated lean startup advice
- Many unicorns (100-1000 person companies) not yet transforming internally with AI
- Bob McGrew (OpenAI) surprised by slow enterprise adoption despite releasing advanced models
