The success of large language models has given a lot of people the wrong intuition: AI went from text to images to video, flattened by scaling laws every step of the way — so surely crossing from the digital world into the physical one (robots) is just “stacking more compute and data” one more time.

That analogy is the most expensive illusion in this whole humanoid narrative. Embodied AI isn’t AI’s next step. It runs headfirst into four walls that digital AI never had to face.

Four walls — and only one yields to scale

Break down the obstacles to mass deployment of robots, and you get four walls of completely different natures:

• The data wall. Robots need real-world physical-interaction data, and unlike text, you can’t scrape it infinitely off the internet — you collect it, one real machine and one real scene at a time. This wall, capital and scale genuinely can grind down.
• The safety wall. A 200-pound machine moving near humans turns a single failure into bodily harm. That’s not an accuracy problem — it’s a liability and legal problem, and you can’t “train away” a legal liability.
• The reliability wall. A chatbot that answers wrong just retries; a physical action that’s wrong is a shattered cup or a halted production line. Industrial settings demand 99.9%+ uptime — a physical-world certainty requirement that scaling laws can’t guarantee.
• The economics wall. For a robot to make commercial sense, its total lifecycle cost has to beat the labor it replaces. That’s an accounting problem — bill of materials, maintenance, utilization, depreciation — and no model, however strong, changes arithmetic.

See the difference in nature, and you see the problem: the LLM playbook only works on the first wall. Safety, reliability, and economics are problems of physics, liability, and accounting — and you don’t flatten those by “scaling once more.”

Moravec’s paradox: easy for humans, hard for machines

There’s a deeper, counterintuitive law underneath — Moravec’s paradox: the more “instinctive” a task is for humans, the harder it is for machines.

Chess, coding, math — hard for people, long since surpassed by machines. Folding a shirt, twisting an odd-shaped cap, picking up an egg in a messy kitchen — the moves of a three-year-old — are exactly where robots still fail. And what humanoids are trying to sell is precisely that second category: “easy for humans, brutal for machines.”

None of this says humanoids have no future. It says: when you price this space, you’re not facing a software problem that “scales in two years.” You’re facing four walls of different natures, on a timeline that can’t be compressed. Anyone extrapolating this as “the next LLM” is paying up for a badly overstated timetable.

Next time you hear “scaling solves everything,” ask one question: which of the four walls are you talking about?

Of the four walls, which do you think falls first — and which drags longest?

(Independent industry and educational research, not investment advice. Companies named are illustrative examples, not recommendations. Data is from public sources and may change. The author and affiliates may hold positions in securities mentioned.)

— Adapted from Embodied Intelligence Investing, Ch. 5: The Four Walls

#EmbodiedAI #Humanoids #PhysicalAI