hawkes.aiKey ideas
- The one idea — Embodied AI is the hard mode: Moravec's paradox, the reality gap and sim-to-real, dexterity far harder than locomotion, and a missing internet-scale corpus of physical action — with vision-language-action models a promising but early answer.
- How it is shown — A glass hand fumbling a simple object while the same AI wins at chess; a clean sim world versus a messy real one; a backflip beside a failed grasp; a data-ocean of text beside a data-puddle of action.
- The trap to avoid — Reading viral humanoid clips as reliable, general-purpose home robots — impressive demos are not the same thing, and they're often curated.
This branch looks like it should be the easy one. It's actually the hardest.
The one idea
Embodied AI is the hard mode: Moravec's paradox, the reality gap and sim-to-real, dexterity far harder than locomotion, and a missing internet-scale corpus of physical action — with vision-language-action models a promising but early answer.
The same AI that plays grandmaster chess still struggles to pick a strange object off a cluttered table. To see why, you have to meet a fifty-year-old paradox. It's called Moravec's paradox. The things a toddler does without thinking — grab a cup, cross a messy room — are brutally hard for machines. The things we find hard, like chess and arithmetic, are easy for them. Evolution spent a billion years on movement and perception — our deepest, least conscious skill. Take the reality gap. Training a real robot is slow and expensive, so much learning happens in simulation. But simulated friction, light, and contact never match the real world. A policy that looks flawless in the sim can fail the instant it touches reality. Closing that gap, sim-to-real, is its own discipline.
How it works — the demo
A glass hand fumbling a simple object while the same AI wins at chess; a clean sim world versus a messy real one; a backflip beside a failed grasp; a data-ocean of text beside a data-puddle of action.
And not all motion is equal. Legged locomotion has come a long way — the walking, running, backflipping robots in every viral clip. But manipulation — hands, fingers, gripping a soft or slippery object without crushing or dropping it — is far harder. Walking impresses the camera. Dexterity is the real wall. Then there's the data problem. Language models had the internet — trillions of words to learn from. There is no internet of physical action. No giant archive of grasping, balancing, and manipulating. A robot has to gather its own experience, slowly, one real attempt at a time. That scarcity is the bottleneck.
The trap to avoid
Reading viral humanoid clips as reliable, general-purpose home robots — impressive demos are not the same thing, and they're often curated.
Why it matters — and what’s next
The newest idea borrows from this course's own models: vision-language-action systems. Feed in camera pixels and an instruction, and let one big model output the motion — the recipe that scaled language, aimed at robots. The humanoid demos come from here. Promising, genuinely. But early, and the clips you see are carefully chosen. So be honest about where this stands. In controlled factories, robots already deliver real value. A reliable, general-purpose robot for your home is not here yet — and the demos, however stunning, are not the same thing. Which sets up the pattern behind our final branch: AI wins fastest wherever the score is clean.
This is a supplement in AI: Zero → Frontier — a side-trip that deepens the act it sits beside, one file and one loop at a time.
Full transcript 2:45 of narration
This branch looks like it should be the easy one. It's actually the hardest. The same AI that plays grandmaster chess still struggles to pick a strange object off a cluttered table.
To see why, you have to meet a fifty-year-old paradox. It's called Moravec's paradox. The things a toddler does without thinking — grab a cup, cross a messy room — are brutally hard for machines.
The things we find hard, like chess and arithmetic, are easy for them. Evolution spent a billion years on movement and perception — our deepest, least conscious skill. Take the reality gap.
Training a real robot is slow and expensive, so much learning happens in simulation. But simulated friction, light, and contact never match the real world. A policy that looks flawless in the sim can fail the instant it touches reality.
Closing that gap, sim-to-real, is its own discipline. And not all motion is equal. Legged locomotion has come a long way — the walking, running, backflipping robots in every viral clip.
But manipulation — hands, fingers, gripping a soft or slippery object without crushing or dropping it — is far harder. Walking impresses the camera. Dexterity is the real wall.
Then there's the data problem. Language models had the internet — trillions of words to learn from. There is no internet of physical action.
No giant archive of grasping, balancing, and manipulating. A robot has to gather its own experience, slowly, one real attempt at a time. That scarcity is the bottleneck.
The newest idea borrows from this course's own models: vision-language-action systems. Feed in camera pixels and an instruction, and let one big model output the motion — the recipe that scaled language, aimed at robots. The humanoid demos come from here.
Promising, genuinely. But early, and the clips you see are carefully chosen. So be honest about where this stands.
In controlled factories, robots already deliver real value. A reliable, general-purpose robot for your home is not here yet — and the demos, however stunning, are not the same thing. Which sets up the pattern behind our final branch: AI wins fastest wherever the score is clean.