4 ms·

I'm trying to figure out why chatGPT was led to believe the person is dead. There should have been something that made it think the person is dead "by associati

I'm trying to figure out why chatGPT was led to believe the person is dead. There should have been something that made it think the person is dead "by association". Tested a number of conjectures, but none of them worked. There may not be a single reason - it's just the interplay of "weights" of tokens that have no direct relation to the author of his profession.

pvaldes 4y ago
> There should have been something that made it think the person is dead "by association". Maybe not. The fast way to find the solution to a question on internet is baiting your fishhook with a false answer to trigger an emotional correcting burst. The more emotional, the better. I speculate that the algorithms could easily suspect some kind of relationship between somebody asking for an obscure figure, and the name provided. If the bot can't find enough data, the human is its better chance to provide some clues. Going silent is a bad strategy to keep the human engaged. To keep the human interest it must say something. Anything. Applying "social hacking" on the human has a good chance of being feed with the correct answer. In this case the trick worked like a charm (again). The bot received not only much more feedback, but also recruited more humans as a bonus.
- dragonwriter 4y ago
  > The more emotional, the better. You’ve got to calibrate it if you want the right answer not to be drowned in a sea of abuse, if it is presented at all. You can easily have the bait be too emotional.
Terretta 4y ago
> trying to figure out why chatGPT was led to believe the person is dead RNG or "temperature" You ask to hallucinate a bio, non-trivial percentage of people with bios are dead, decent chance the fever dream generated ends with dead and obits in plausible places. This thing is a magic 8 ball with billions of phrases, it picks continuations based on probabilities and a random number to mix it up a bit for freshness or 'creativity'. Smart people believe horoscopes and psychic cold readings. Smart people are fooled.

pvaldes 4y ago

> There should have been something that made it think the person is dead "by association". Maybe not. The fast way to find the solution to a question on internet is baiting your fishhook with a false answer to trigger an emotional correcting burst. The more emotional, the better. I speculate that the algorithms could easily suspect some kind of relationship between somebody asking for an obscure figure, and the name provided. If the bot can't find enough data, the human is its better chance to provide some clues. Going silent is a bad strategy to keep the human engaged. To keep the human interest it must say something. Anything. Applying "social hacking" on the human has a good chance of being feed with the correct answer. In this case the trick worked like a charm (again). The bot received not only much more feedback, but also recruited more humans as a bonus.

dragonwriter 4y ago
> The more emotional, the better. You’ve got to calibrate it if you want the right answer not to be drowned in a sea of abuse, if it is presented at all. You can easily have the bait be too emotional.

Terretta 4y ago

> trying to figure out why chatGPT was led to believe the person is dead RNG or "temperature" You ask to hallucinate a bio, non-trivial percentage of people with bios are dead, decent chance the fever dream generated ends with dead and obits in plausible places. This thing is a magic 8 ball with billions of phrases, it picks continuations based on probabilities and a random number to mix it up a bit for freshness or 'creativity'. Smart people believe horoscopes and psychic cold readings. Smart people are fooled.