Install Theme

what GPT-4 “knows” about the northern caves

image

A lot of this is pretty close to right!

I wanted to see if it knew, on some level, which parts were definitely based in reality, and which parts were mere guesses.  So I asked it to express numerical confidences.  Screenshots under the cut

image
image

I don’t know what I expected… the two claims under 70% are both false, so maybe that’s something?

  1. cimmerian-chaos reblogged this from nostalgebraist
  2. nostalgebraist said: @leshwi oh interesting, i didn’t think to look up that name. fwiw, the other inaccuracies are “Neil Henderson” (not in TNC, dunno where it got this name); “Sam Keeper” (actually the name of a blogger i occasionally argued with around the time i wrote TNC); “1970s” (the in-story book TNC dates to the 90s); and the claim that the in-story TNC is a series (it is a book, a successor to an earlier series called Chesscourt).
  3. leshwi said: naming the in-fiction author as VM Straka I think shows that it’s also pulling in information about S, a weird high-concept mystery that happens in the margins of a fictional book. I wonder if that could account for the other inaccuracies?
  4. l33tminion reblogged this from nostalgebraist and added:
    Confusing Leanord Salby with V. M. Straka strikes me as a surprisingly sophisticated mistake.
  5. holyscream reblogged this from nostalgebraist
  6. nostalgebraist said: @eightyonekilograms yes, see prev replies
  7. eightyonekilograms said: Did you see the claims that supposedly the base model is highly calibrated and then the RLHF wrecks it.
  8. nostalgebraist said: @how-about-a-nice-game-of-chess they were probably referring to token probabilities, though they don’t actually say. but i bet instruction tuning and rlhf messed up this kind of stuff too.
  9. mashivan reblogged this from raginrayguns
  10. 91625 reblogged this from nostalgebraist
  11. how-about-a-nice-game-of-chess reblogged this from nostalgebraist and added:
    There was a mention in the GPT-4 tech report (page 12) that the pre-RLHF model had pretty good calibration, but RHLF...
  12. transienturl reblogged this from nostalgebraist
  13. nostalgebraist said: @nationalmissiledefense i don’t think so? that doesn’t explain any of the errors. and it correctly identifies that there is a fictional book in the story with that title