Every scalable-oversight test on record keeps an answer key in the loop
Weak supervision recovers at most 80 percent of the performance gap to a stronger model, and both it and debate have only been measured on tasks a human could already grade alone.
Answer order alone swung a language model's verdict by 80 points
Judged by ChatGPT, the same two chatbot answers swung from a 2.5% win rate to an 82.5% win rate depending only on which one appeared first.
GPT-4 names Tom Cruise's mother but not Mary Lee Pfeiffer's son
Berglund et al. tested about 1,500 real parent-child pairs and found forward accuracy at 79% and reverse accuracy at 33% — a training update that only ever runs in one direction, not a fact the model forgot.
Epic's sepsis alarm missed two-thirds of the patients it was built to catch
An outside test of 38,455 hospital stays found the widely used score performed far below Epic's own numbers, and traced the gap to how the model defined sepsis in the first place.
Two more papers turned a robot's backflip trick into InstructGPT's training recipe
Christiano et al. 2017 spent tens of thousands of human comparisons on Atari games and simulated robots, not one sentence of text, and getting that recipe to work on a language model took five more years.