Physical AI Lab // Episode 0.2 // Quiz
Why now?
Data, models, compute — three curves rising together
01
Why did robots seem to learn to use their hands only in the last couple of years? What is the episode's core answer?
A brand-new kind of motor suddenly made robots far more dexterous
Robot algorithms suddenly got smart and no longer need much data
Data, models and compute all matured during the same period
Training compute grows 4–5x a year, and compute alone is enough
02
Before 2023, each lab collected only a small set of data from its own robots. Which problem did the Open X-Embodiment paper tackle directly?
Robot data had no action labels, so models could not learn from it
Data was split across labs and robot types, each set too small
Real-robot data was too risky, so it had to come from simulation
There was too much human video, and irrelevant parts had to go
03
Someone wants a large model trained on huge amounts of images and text to control a robot arm directly. Following RT-2, what is the key step?
Write actions as “words” the model can output from an image
Fit the arm with more cameras so the model can see more clearly
Drop web images and text and retrain from scratch on simulation
Let a person teleoperate while the model narrates what happens
04
Someone writes: “Training compute for frontier models grows 4–5x a year, so the chips on robots also get 4–5x faster each year.” What is wrong?
Nothing; the two are simply the same curve
The number is off; the statistic says 7.5x a year
The statistic covers only humanoids and leaves out other robots
It counts frontier-model training, not robot chips
05
A vendor says its new robot computer has 7.5x the peak compute of the last generation at a specific precision. To learn how much faster the robot actually works, what should come next?
Divide the task completion time by 7.5 directly
Measure latency and results on the same task and setup
Convert it to total data-centre compute, then compare
Use the yearly growth in training compute to estimate it
Check