Worldbuilders: Why Most AI Startups Won't Survive | The Model Economy by Sumeet Singh

What It Covers

Sumeet Singh of Worldbuild presents the Model Economy framework, arguing that AI scaling laws will eliminate most specialist SaaS-style AI apps, and that durable venture value accrues to model infrastructure and post-skeuomorphic applications instead.

Key Questions Answered

• •The Bitter Lesson: Frontier model task completion length has doubled every six months since GPT-2, growing from two seconds to 6.6 hours of autonomous operation. Specialist AI apps built around teaching models domain rules—accounting, marketing—will be outperformed by base models within months.
• •Model Economy Infrastructure: Rather than building AI applications, target infrastructure that feeds model growth: compute marketplaces that trade GPU capacity like commodity futures, smooth supply volatility between shortage and glut cycles, and data businesses that sell experiential physical-world data directly to model developers as a new revenue stream.
• •Offensive AI Security: Security in the model economy shifts from defensive firewalls to active red-teaming. As models enter cars, robotics, and critical systems, the venture opportunity lies in dedicated teams that systematically jailbreak and stress-test models before bad actors can exploit those same vulnerabilities.
• •Post-Skeuomorphic Applications: Target workflows impossible without AI, not existing workflows made faster. Two concrete examples: multi-agent swarms where distinct models write, review, test, and deploy code collaboratively, and self-healing observability systems that autonomously diagnose, patch, and deploy fixes without human intervention at 3AM.