Raw Content

Two Ways to Win in the Post-software Era

Author: Sumeet Singh

Publication: Every

Date: December 8, 2025

Core Argument

Singh contends that most AI founders are making a critical strategic error by building specialist tools atop foundation models using outdated SaaS frameworks. He references Richard Sutton’s “bitter lesson”—the principle that simple systems with massive scale and compute consistently outperform specialized, hand-crafted solutions.

The Central Problem

“The length of tasks generalist models can achieve is doubling every seven months,” Singh notes, making specialized AI wrappers vulnerable to obsolescence within 18 months as base models grow more capable.

Two Viable Paths Forward

Path 1: The Model Economy

Companies building infrastructure that models require to improve, rather than applications using current models:

Compute as commodity: Trading compute resources and energy via exchanges like San Francisco Compute Company
On-device AI: Hardware-software combinations enabling local model deployment (examples: Meta wearables, Truffle)
Data exchanges: Facilitating licensing of specialized data to model providers
Security: Offensive security services identifying AI system vulnerabilities

Path 2: Post-Skeuomorphic Apps

Applications discovering entirely new workflows impossible without AI—avoiding the trap of digitizing old processes:

Multi-agent coordination: Using multiple models simultaneously to eliminate individual model limitations
Simulations at scale: Running thousands of parallel experiments with continuous feedback
Continuous feedback loops: Self-healing software that autonomously diagnoses and fixes problems without human intervention

The Critical Question

“Is this company in line to be absorbed by the bitter lesson, or can it thrive by building for scale or true novelty?”