Robot data estimate is off by 100–500 times, a user argues
The user says the cited figures—64TB of robot data per day versus a 120TB language-model training corpus—imply roughly half as much data, not 200 times as much.
TLDR
A user disputes a headline they describe as claiming that a robot may generate 200 times the data in an entire frontier language-model training corpus in one day. They say the cited figures are 64TB per day for the robot and 120TB for the corpus, contradicting that comparison. The user also challenges the underlying data estimates, saying the cited research uses the DROID database, whose 350 hours of training data total 1.7TB, with the raw footage totaling 8.7TB.
Combined views
33.3K
1 Source, first seen 29d ago