AI Just Gave You Superpowers — Now What?

What It Covers

Christian Catalini, co-founder of LightSpark and creator of MIT's Cryptoeconomics Lab, joins Eddie Lazarin on the a16z podcast to unpack Catalini's 100-page paper "Some Simple Economics of AGI," examining how AI reshapes labor markets, startup formation, verification costs, and the complementary role blockchain infrastructure plays in an automated economy.

Key Questions Answered

• •The Automation-Verification Split: Every job contains two categories of tasks: automatable work (anything measurable, with existing data) and verification work (judgment calls drawing on unique human experience). As AI absorbs the first category rapidly, the economic value of the second category rises proportionally. Workers should audit their current roles and deliberately shift time toward verification tasks — the ones requiring out-of-distribution judgment no training dataset fully captures.
• •The AI Sandwich Org Structure: Catalini's framework for future firms has three layers: one human "director" steering intent and course-correcting drift at the top; a swarm of AI agents executing in the middle; and a small team of domain-expert verifiers at the bottom reviewing agentic output with specialized tooling. Startups building toward this structure today — rather than traditional headcount scaling — position themselves for the one-person billion-dollar company model that AI now makes structurally achievable.
• •The Codifier's Curse: Top domain experts hired to evaluate and label AI outputs — writing evals, training data, and verification benchmarks — are simultaneously creating the datasets that will automate their own peers. This self-displacing loop means verifiers must continuously move up the knowledge stack, staying one step ahead of improving models. The practical response is hyper-specialization: own the thinnest, highest-leverage slice of a domain where data remains sparse and judgment remains irreplaceable.
• •Systemic Risk from Unverified AI Output: When 60% or more of code ships machine-generated and human review becomes physically impossible at that throughput, organizations accumulate hidden technical debt and latent security vulnerabilities. Catalini draws a parallel to Long-Term Capital Management's collapse — rational short-term optimization masking systemic fragility. The emerging response is AI liability insurance, exemplified by ElevenLabs insuring their audio agents, signaling that financialization of AI risk is a near-term structural shift, not a distant concept.
• •Verification-Grade Network Effects as the New Moat: Traditional two-sided marketplace network effects are increasingly vulnerable to AI, which can bootstrap both sides of a market at low cost. The durable competitive advantage instead comes from proprietary failure data — years of logged edge cases, anomalies, and out-of-distribution events — that trains better verification systems. Companies that build feedback loops converting every human expert decision into labeled training data will underwrite risk more accurately and deliver safer products at lower cost than competitors.
• •Blockchain as Verification Infrastructure: As AI agents proliferate and single-person companies multiply, coordination across fragmented economic actors requires credibly neutral rails for identity, provenance, payments, and insurance. On-chain transaction flows give agents richer, real-time context versus opaque legacy APIs — one founder switching to stablecoin payments found agent reliability improved because all signals were visible on-chain. Crypto primitives — smart contracts, cryptographic provenance, prediction markets — become foundational verification tools precisely when trust in digital information becomes scarce.