AI unlocks startup ideas that previously failed on economics or timing

Executive overview

Many startup ideas from the 2010s failed not because the concept was wrong, but because the technology wasn't ready — evaluating humans at scale, personalised tutoring, full-stack services with thin margins. LLMs change the underlying economics and feasibility of these ideas. The right approach now is to explore the technology first, follow curiosity, and trust that good ideas surface from direct contact with what's newly possible.

The best startup ideas today are old ideas whose time has finally arrived.

Old marketplace ideas now viable with AI

• Three-sided marketplaces collapse to two-sided when AI handles the intermediary layer — recruiting being the clearest example
• Mercor rebuilt the Triple Byte thesis (curated engineer marketplace) but used LLMs for evaluation from day one, skipping years of label-data accumulation
• Apriora runs AI-led technical screening interviews; companies now apply it to senior engineers, not just junior filtering
• The pattern applies broadly: identify any marketplace with a human-in-the-middle and ask what AI replaces
• Founders must push through investor cynicism — capital lost on prior attempts (hundreds of millions in recruiting alone) creates lasting bias against the category
• Instacart's unlock was the smartphone; many AI-era companies share the same structure — same idea, new infrastructure makes it work

Personalised learning as a breakout consumer category

• Truly personalised tutoring has been an internet promise for 25 years; LLMs are the first technology that can actually deliver it
• Revision Dojo (YC-funded): exam prep with personalisation, high DAU and power-user retention
• Aidexia: AI grading assistant for teachers — grading is the leading cause of teacher attrition
• Speak: language learning startup that bet on personalisation before LLMs existed; doubled down on GPT-3.5 early, now growing strongly
• Consumer AI unit economics are still marginal, but inference costs are falling fast — freemium may become viable again at scale
• If an app matches the quality of a human tutor, parents pay tutor prices, not app prices — the business model changes entirely

Full-stack startups: gross margin was the real problem

• Full-stack startups (tech-enabled services) were popular in the 2010s — the thesis was: own the whole stack, capture 100% of value
• Triple Byte, Atrium, Zenefits-adjacent models all hit the same wall: ops complexity scales headcount, gross margins stay thin
• The real constraint: running ops consumes founder attention that should go to product and distribution
• WeWork is the extreme case — no tech margins, just creative accounting
• The bull case now: AI agents replace the ops layer, so full-stack companies can have software-level margins for the first time
• Legora (YC) is building legal AI tools that will extend into doing the legal work — effectively an AI law firm in formation
• Justin Kan noted Atrium failed partly because the AI wasn't good enough yet; it is now

Platform neutrality and the AI assistant bottleneck

• Siri remains poor because Apple has no competitive pressure to improve it — a structural platform problem, not a capability one
• Historical precedent: government intervention forced Windows browser choice; that move prevented IE from locking out Google
• The same logic applies to voice AI on phones — users should be able to set a default AI assistant
• Gemini 2.5 Pro benchmarks as well as or better than O3 on many tasks, but consumer usage is a fraction of ChatGPT's — brand and first-mover inertia are real moats
• Google is shipping the org: two separate Gemini APIs (DeepMind vs GCP) signal internal fragmentation
• Copilot in Windows and Gemini in Gmail are widely seen as inferior integrations — being embedded doesn't guarantee adoption
• TPUs give Google a structural cost advantage for large context windows; other labs find it expensive to match

Infrastructure still to build

• The tooling layer around agents — evals, deployment, orchestration — is still underdeveloped
• ML ops companies in 2019 were ahead of their customers: more applications to build tooling than to use ML
• Replicate (YC W20): nearly abandoned the company during COVID; image diffusion models arrived and it exploded overnight
• Ollama: similar story — quiet years, then Llama's release created instant demand for local model deployment
• Deepgram: two physics PhDs spent years on speech-to-text before it worked; now powers most voice agent startups
• Moral: being on top of the well before oil comes up is the strategy, but it requires conviction without external validation

How to find startup ideas now

• The lean startup / customer discovery model was correct for an era when the idea space had been picked over for 20 years
• That advice is now outdated — in an AI-first world, the right move is to use the technology, follow curiosity, and bump into ideas
• Most incumbent companies (100–1,000-person startups with strong revenues) are not running skunk-works AI projects; the gap between what's possible and what's deployed is still enormous
• Vibe-code a blog platform with interactive AI prompts embedded — free idea, YC will fund it
• The unlock: apply the right prompts, the right data, ingenuity, evals, and taste — the output is still surprising even to people building the models