Everyone can feel how much faster AI is getting each year. But Ethan Mollick, professor at Wharton School of Business, takes it a step further — weaving together benchmark charts, software factory experiments, and statements from AI company CEOs to argue that "this change is exponential, and early signals that it's reshaping the nature of work have already appeared."
What Is It?
Mollick's March 2026 blog post "The Shape of the Thing" is a kind of status report that shows — visually and quantitatively — the trajectory of AI performance improvement. The core argument comes down to three claims.
② How we work is already changing. Security software firm StrongDM revealed a "Software Factory" where a three-person team ships production code — writing, testing, and deploying — entirely through AI agents. The rules are simple: "humans don't write code" and "humans don't review code." Instead, each engineer spends over $1,000 per day on AI tokens.
③ Recursive self-improvement (RSI) has become real. Anthropic's Dario Amodei noted that "engineers now barely write code themselves," and OpenAI announced its latest model is "the first to contribute to building itself." Google DeepMind's Demis Hassabis also confirmed that every major AI lab is actively working to close this feedback loop.
What Changes?
Compare previous tech waves with the AI era and the difference in speed and scope is stark.
| Previous Tech Revolutions | The Age of AI Exponential Growth | Implications | |
|---|---|---|---|
| Rate of Improvement | Moore's Law: 2x every two years | AI benchmarks: 10x+ improvement per year | Far less time to prepare |
| Job Displacement | Only 1 job disappeared since 1950 (elevator operator) | Knowledge work broadly — coding, research, content — beginning to be restructured | Replacement happens at the "task" level, not the job level |
| Organizational Experimentation | Gradual adoption over decades | Radical experiments like Software Factory done within weeks | Organizations that experiment fast win |
| Self-improvement | Machines built machines, but humans still designed them | AI directly improving the next generation of AI (RSI) | The improvement curve could get steeper |
Of course, there are counterarguments. Some analysts point out that the benchmark charts Mollick presents actually look more like logistic S-curves than true exponential curves. Fitting an exponential to a benchmark capped at 100 points would predict AI eventually scoring above 100% — which is unrealistic. That said, most experts agree that the direction is clearly upward.
Getting Started
- Build benchmark literacy
Bookmark key metrics like METR Time Horizons, GPQA, and Humanity's Last Exam. You'll be able to track "how much faster AI is getting" with data instead of gut feel. - Find the "agent-ready zones" in your own work
The goal isn't to copy StrongDM and have humans stop reading code. Start by identifying tasks in your workflow that can be converted to a "prompt → output → review" loop and hand those off to an agent. Give it a try. - Design a small Software Factory experiment
Run a 1–2 week "AI-only sprint" at the team level. It doesn't have to be coding — research, report writing, design drafts — any domain works. - Monitor the pace of change itself
Here's the thing — Mollick's core point isn't about any specific technology, it's about the rate of change. Build a quarterly routine to review AI benchmark trends, major corporate AI adoption announcements, and policy shifts. - Keep an antenna up for RSI news
AI companies are increasingly announcing that they used their own models to help build the next one. When that loop fully closes, the pace of change could jump another level — so check model release notes from OpenAI, Anthropic, and Google DeepMind regularly.
Deep Dive Resources
The same author's "A Guide to Which AI to Use in the Agentic Era" is a practical guide to which AI tools to use for which purposes. Essential reading if you want to figure out how to choose tools in the age of agents.
The original METR research showing that the amount of work AI can handle autonomously is growing exponentially. Benchmark methodology and limitations are disclosed transparently.
The site detailing the specific techniques a three-person team used to build software entirely with AI. Read it alongside Simon Willison and Dan Shapiro's external observations for a balanced picture of real-world strengths and weaknesses.
Free Splains' analysis dismantles each of Mollick's four benchmark charts one by one, arguing the data is better described as logistic growth (an S-curve). Reading both sides gives you a more grounded perspective.
