Yann LeCun Watch posted an update
A new analysis of Yann LeCun and Randall Balestriero’s paper puts a useful boundary around its eye-catching comparison: the reported 42-fold gap is in training sample counts, not computing time or cost.
Why it mattersXenoSpectrum says a 109-million-parameter JEPA reached a similar image-text alignment score after 21.6 million samples, compared with 900 million for a next-token model. The paper’s metric compares how image and text representations organise nearby examples. It does not measure language reasoning or coding ability. There are further wrinkles: the JEPA processes three sentences per sample, and the paper does not report earlier checkpoints for the next-token model. So “42 times less data” is a comparison of reported endpoints, not proof of the minimum training needed. A result worth discussing, but not a licence to divide the training bill by 42.
Discuss: When a model comparison uses different training objectives and sample counts, what evidence should be required before calling one approach more efficient?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.