BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Can Humans Supervise Increasingly Ultracrepidarian AI? - Jose Hern
 andez-Orallo
DTSTART:20241205T150000Z
DTEND:20241205T160000Z
UID:TALK225295@talks.cam.ac.uk
CONTACT:Matthew Colbrook
DESCRIPTION:Large language models have evolved to solve increasingly compl
 ex problems but still fail at many simple ones --from a human point of vie
 w. This discordance with human difficulty expectations strongly affects th
 e reliability of these models\, as users cannot identify a safe operating 
 condition where the model is expected to be correct. With the extensive us
 e of scaling up and shaping up (such as RLHF) in newer generations of LLMs
 \, we question whether this is the case. In a recent Nature paper\, we exa
 mined several LLM families and showed that instances that are easy for hum
 ans are usually easy for the models. However\, scaled-up\, shaped-up model
 s do not secure areas of low difficulty in which either the model does not
  err or human supervision can spot the errors. We also found that early mo
 dels often avoid user questions\, whereas scaled-up\, shaped-up models ten
 d to give apparently sensible yet wrong answers much more often\, includin
 g errors on difficult questions that human supervisors frequently overlook
 . Finally\, we disentangled whether this behaviour arises from scaling up 
 or shaping up\, and discovered new scaling laws showing that larger models
  become more incorrect and especially more ultracrepidarian\, operating be
 yond their competence. These findings highlight the need for a fundamental
  shift in the design and development of general-purpose artificial intelli
 gence\, particularly in high-stakes areas where a predictable distribution
  of errors is paramount.\n\nThe talk will be based on the recent paper: L 
 Zhou\, W Schellaert\, FM Plumed\, YM Daval\, C Ferri\, JH Orallo (2024) "L
 arger and more instructable language models become less reliable"\, Nature
 \, 61-68\n
LOCATION:Centre for Mathematical Sciences\, MR14
END:VEVENT
END:VCALENDAR