AI learned everything it knows by reading. You learned by living. That's not a small distinction.
The gap most AI discourse glosses over
That's the core tension Sebastian Mallaby unpacks in a recent Hard Fork episode, drawing on three years embedded with Demis Hassabis and the team at Google DeepMind. There's a fundamental difference between crystallized knowledge: language, text, the recorded residue of human experience — and the kind of grounded, embodied learning that happens when you actually do something in the world.
Bodies learn differently than models
Neuroscience calls it action/perception coupling. You don't truly learn to ride a bike by reading about balance. You fall. You adjust. You build a model of the world through interaction, trial and error, feedback loops. That's reinforcement learning in its purest form — not abstracted into data, but lived.
The bet LLMs were built on
LLMs trained in 2018 and 2019 were built on a specific bet: that language encodes enough of how the world works to simulate understanding. And to a remarkable degree, it does. Language is extraordinarily rich — it carries logic, emotion, cultural context, cause and effect. But it's still a record of the world, not the world itself. It's the fossil, not the organism.
Why this matters for creative work
The outputs you get are deeply pattern-matched to what has been said before, not what has been experienced. AI can't describe grief. It has never lost anything. AI knows what resilience looks like in language, but It has never had to find any.
That gap is not a bug. It's the job description for human creatives.
The frame that actually works
Use AI to compress and accelerate the parts of the process that are crystallized: research, testing, execution at scale. Reserve human judgment for the parts that require being in the world: reading the culture, sensing the moment, knowing when something is actually true versus just plausible.
Designers who understand this boundary will outperform those who either fear AI or blindly defer to it. The tool accelerates. The person decides.
That's not a limitation of the technology. That's the whole point.
🎙️ Hard Fork, Episode 188 — worth your commute.