The Experience Bottleneck

Language models learned to write from text that already existed. Nobody had to produce it for them. A robot learning to fold a shirt has no comparable library, and that one asymmetry explains more about the state of physical AI than any argument about model architecture does.

The popular framing treats embodiment as the next scaling problem: take the recipe that worked for language, point it at hardware, wait. But scaling laws describe what happens when you add data. They are silent on whether the data can be obtained at all. For text, supply was a stock that human civilization had already written down. Llama 3 was pretrained on more than 15 trillion tokens from public sources, gathered at almost no cost per token. For robots, supply is a flow. Physical Intelligence trained its π0 model on over 10,000 hours of data across seven robot configurations, and the Open X-Embodiment collaboration pooled 60 datasets from 34 laboratories to reach roughly a million real trajectories. Every hour inside those numbers was a physical machine moving in real time.

Text was a stock that already existed. Embodied experience is a flow you buy in wall-clock time, on hardware you own and maintain. Scaling laws say nothing about the purchase rate.

Two kinds of data, two kinds of cost

Put the two training regimes on the same axis and the distance is easier to accept than to argue with. It is roughly four orders of magnitude, and it is not closing through effort alone.

This is why the binding constraint in physical AI is throughput of practice rather than parameter count. You can rent ten thousand accelerators this afternoon. You cannot rent ten thousand robot-years of contact experience, because nobody has it to lend, and generating it consumes calendar time at a fixed exchange rate. A fleet that collects data eight hours a day produces eight hours of data a day, and no amount of capital changes that ratio without more machines and more operators.

What Moravec got right, and where the story is thinner than it looks

The usual explanation reaches for Moravec's paradox. In Mind Children (1988), Hans Moravec observed that it was comparatively easy to get computers to adult-level performance on intelligence tests or checkers, and difficult or impossible to give them the perception and mobility of a one-year-old. The lesson drawn from it is that human-easy implies machine-hard.

That lesson deserves more scrutiny than it usually gets. Arvind Narayanan has argued the paradox has never really been fact-checked, and that its apparent support comes from selection bias: the anecdotes always compare chess against laundry and quietly skip the tasks that are easy for both humans and machines, and the tasks that are hard for both. Open-domain reasoning in law or science is difficult for current systems, and Moravec's scheme would have filed it under easy. Once the ignored cells return to the picture, the clean inverse relationship stops being clean.

The gap you can randomize away, and the gap you cannot

If real experience is the expensive input, simulation looks like the way out. Simulated robots run faster than real time, in parallel, without maintenance. The standard technique is domain randomization: perturb the simulator's physical parameters during training so the policy learns to work across a distribution of dynamics, and the real world becomes one more sample from that distribution. It works. Legged locomotion and gross manipulation have both crossed from simulation into hardware this way.

The limit is specific and worth stating precisely. Domain randomization can only close gaps the simulator represents. Randomizing friction, mass, latency, and lighting covers the parameters someone thought to expose. Sim-to-real surveys point out that it does nothing for the effects the simulator omits entirely: contact and deformation, aerodynamic detail, wear, and response delays that were never in the model. You cannot randomize over a variable that does not exist in your equations.

Where the opportunity actually sits

The field's most productive response has been to stop paying for what it can inherit. Google DeepMind's RT-2 showed that a vision-language model trained on web data transfers that knowledge into robotic control, so the robot arrives already knowing what a banana is and what "put it in the bowl" means. Physical Intelligence built π0 on top of PaliGemma, a 3-billion-parameter vision-language model, and added a flow-matching action head. Open X-Embodiment went further and pooled trajectories across 22 different robot bodies, then found that policies trained on the pool beat policies trained on any single embodiment, including in domains that already had plenty of data.

Read together, those results describe a division of labour. Semantics, object knowledge, and language understanding come from data that is already abundant and still growing for free. Scarce robot data gets spent only on the part nothing else can supply: the mapping from intent to actuation on a specific body. That is the opportunity, and it is an engineering strategy rather than a hope.

What this says about AGI

Two camps disagree about whether any of this is on the critical path. One holds that general intelligence requires grounding: that a system with no ability to act on the world and observe the consequences cannot acquire causal understanding, only correlational summaries of someone else's. The other holds that competent physical action is a useful application of general intelligence rather than a precondition for it, and that reasoning can be learned from description alone. The argument is unresolved and will not be settled by assertion.

What the data economics settle is narrower and more useful. If grounding is required, then the arrival of general intelligence is governed by an acquisition rate rather than a compute budget, because the necessary experience has to be generated by machines acting in real time. That reframes the forecast from an algorithms question into a question about fleets, operators, and calendar time, which behaves very differently: it is slower, more capital-intensive, and much harder to accelerate with a breakthrough. Anyone quoting an AGI date on the strength of a training-compute curve is answering a question about text and presenting it as an answer about the world.

The strategic asymmetry follows from the same arithmetic. Deployment is data collection, so whoever fields machines earliest compounds fastest, and the advantage is cumulative in a way model weights are not. A competitor can match your architecture in a quarter. They cannot retroactively collect the robot hours you started gathering two years ago.

What leaders should do

1. Instrument every deployed machine for data capture before you optimize its autonomy. Deployment is the only faucet you control, and unrecorded operating hours are permanently lost.
2. Select automation targets by contact complexity rather than by how hard the job looks to a person. Work whose physics your simulator already models is cheap to train. Contact-rich dexterity is not, and no vendor roadmap changes that.
3. Put teleoperation on the balance sheet with an hourly rate, and track cost per successful trajectory as an operating metric. Teams that do not measure it discover their data cost after committing to it.
4. Join or fund a cross-embodiment data pool instead of assuming your own fleet will reach useful scale. The Open X-Embodiment result is that pooled data beats private data, including for organizations that already had plenty.