If I could afford it I'd quite happily employ a Ph.D. level researcher with inability to count the R's in strawberry. The real problem is in accurately characte

If I could afford it I'd quite happily employ a Ph.D. level researcher with inability to count the R's in strawberry. The real problem is in accurately characterising where these models fail in relatively subtle task-relevant ways. Hopefully people work that out before they are universally deployed. The overpromising is definitely a problem.