build1 distinct publisher
A completer that scores 0.546 on its eval scores 0.070 on the thing users see
Every dial in pycomplete won a sweep against held-out next-token accuracy. Then its author scored the ghost text, and it came back right one time in fourteen.
Publishers:dev.to
Reality
- Evidence62
- Adoption
- Insufficient
- Hype gap−12
- Incentives34
- Confidence55