Lee Kezar (@leekezar) 's Twitter Profile
Lee Kezar

@leekezar

PhD candidate @USCViterbi. I make linguistically-grounded ML models for American Sign Language. #FreePalestine

ID: 1079407293169582082

linkhttp://leekezar.github.io calendar_today30-12-2018 16:02:07

845 Tweet

384 Takipçi

549 Takip Edilen

Lee Kezar (@leekezar) 's Twitter Profile Photo
Lee Kezar
@leekezar

4 months ago

GPT 4o is noticeably smarter than 3.5, but cannot follow negative constraints (e.g. "don't say X") whatsoever. Looks like the inverse scaling challenging was right about this: More params --> less capable of parsing negation?