4 ms·
If I'm following correctly, they trained a single model with multiple training paradigms and then the single model could perform token predictions for multiple
by f38zf5vdt 4y ago
If I'm following correctly, they trained a single model with multiple training paradigms and then the single model could perform token predictions for multiple dissimilar token sequences for specific tasks. Seems like it is a straightforward result.
- doubtfuluser 4y agoWell… straightforward in a way, yes. But the scale of learning is huge especially with this diverse set of tasks. Not totally unexpected, but certainly not clear that it would work with current networks and sizes.
- f38zf5vdt 4y agoRight, exactly. Something that seemed like it should work but no one had ever tried it.