You've probably watched frontier models fail at the now-famous "strawberry" prompt, or the equally famous "car wash" question. Now ask the very same models to tackle an eighty-year-old geometry conjecture from Paul Erdős, and one just disproved it.

Back in the 1980s, Hans Moravec noticed the same asymmetry: expert-level chess was easy to build, a one-year-old's perception nearly impossible—an observation that became known as Moravec's paradox.

Stanford's newly released 2026 AI Index documents this phenomenon, and gives it a name: the "jagged frontier."

The same generation of models that can disprove a decades-old conjecture still misreads an analog clock about half the time, while humans get it right about 90% of the time. "Frontier" has become the dominant metaphor in AI, but here I wonder whether "topography" better captures what we're actually discovering.

For decades, we treated intelligence as one-dimensional—a single ladder to climb. LLMs completely overturn that intuition. Foundation models are not becoming uniformly smarter. I increasingly think of them as a "landscape" we're only beginning to map.

⸻

Here's the real paradox: we call them "foundation" models, yet nothing about them behaves like a foundation. Foundations are supposed to be stable. These ones shift beneath our feet every few months. Their capabilities evolve, their failure modes move, and many of their emergent abilities remain only partially understood (Anthropic's recent note on J-space is a fascinating example.)

Across this shifting landscape, some mountains are disappearing almost overnight. Others barely move. Coding, once one of the highest peaks, is already becoming a plateau. Common sense—long assumed to be flat ground—turns out to be one of the deepest valleys.

Benchmarks once expected to last for years now saturate in months. Leadership changes hands every few model releases. Yesterday's breakthrough quickly becomes today's baseline. We're no longer building on stable ground. We're building on tectonic plates.

For enterprises like Michelin, the challenge is no longer simply adopting AI. It's learning how to inhabit—and gradually urbanize—this moving continent: deciding where to build, where to automate, where to keep a human guide, and how to redraw the map every time the landscape shifts. And here the challenge is not about finding the bedrock, but rather learning to build without it!

That's how I increasingly think about Harness Engineering. Not just as an orchestration layer, but as the discipline of continuously making this moving landscape habitable.

For decades, we treated software as the ground itself, with AI as just another layer built on top of it. It seems to me that an inversion is already underway.

Perhaps we've misunderstood the metaphor all along. Foundation models aren't the foundations. They're the geology. Everything we build above them has to adapt to a terrain that never stops moving.