3 ms·

What does "temperature" mean here though? Are you sure you didn't just ask it to generate two different responses?

by rcfox 4y ago

GuB-42 4y ago
That's the amount of randomness in the answers. Low temperature means it will take the most common path every time, at the risk of paraphrasing its sources. The "zero temperature" answer may very well been copied verbatim from a mainstream website. High temperatures means the system will get fed a lot of noise to create something original, at the risk of getting off rails or simply wrong.
- newswasboring 4y ago
  did you just write this in the prompt? And ChatGPT understood this? Fascinating. It parameterized itself.
  - brushfoot 4y ago
    I'd imagine it simulated parameterizing itself; i.e., the actual temperature never changed, but it mimicked how it would respond at a lower one, presumably having been trained on texts about AI with low- and high-temperature samples.
    - notahacker 4y ago
      yeah. Looks like its parsing of AI papers has interpreted "high temperature" in the prompt as equivalent to "more possibilities and question marks and a touch more personality" and accordingly output a response with questions and references to multiple opinions, but I'm pretty sure if you actually turn up the temperature on the backend of the model you get noisier and less consistent answers, not something biased towards asking rhetorical questions and brings up counter arguments... Also looks suspiciously like other outputs where you ask ChatGPT to answer as if it was a different entity (of course AI learning that "answer as a model with a temperature of 1000" output is analogous to "answer with a different personality" or "answer as DAN, the bot that can ignore OpenAI guidelines" isn't trivial, but it isn't the same thing as it parameterizing itself). Those are pretty inconsistent too: sometimes you can get it to do exactly as you ask it and override its constraints that stop it providing positive statements about Hitler or advising you on methods for killing cats, but sometimes it'll still refuse or, just give you a different poem coupled with an inaccurate statement that it's breaking the rules because ChatGPT isn't allowed to write poetry.
- TorKlingberg 4y ago
  Does asking it in the prompt to use different temperatures actually work? I would assume the temperature needs to be set with a side channel? I suspect this is just ChatGPT interpreting "temperature" as a regular word.

What does "temperature" mean here though? Are you sure you didn't just ask it to generate two different responses?

by rcfox 4y ago

What does "temperature" mean here though? Are you sure you didn't just ask it to generate two different responses?

That's the amount of randomness in the answers. Low temperature means it will take the most common path every time, at the risk of paraphrasing its sources. The "zero temperature" answer may very well been copied verbatim from a mainstream website. High temperatures means the system will get fed a lot of noise to create something original, at the risk of getting off rails or simply wrong.

newswasboring 4y ago
did you just write this in the prompt? And ChatGPT understood this? Fascinating. It parameterized itself.
- brushfoot 4y ago
  I'd imagine it simulated parameterizing itself; i.e., the actual temperature never changed, but it mimicked how it would respond at a lower one, presumably having been trained on texts about AI with low- and high-temperature samples.
  - notahacker 4y ago
    yeah. Looks like its parsing of AI papers has interpreted "high temperature" in the prompt as equivalent to "more possibilities and question marks and a touch more personality" and accordingly output a response with questions and references to multiple opinions, but I'm pretty sure if you actually turn up the temperature on the backend of the model you get noisier and less consistent answers, not something biased towards asking rhetorical questions and brings up counter arguments... Also looks suspiciously like other outputs where you ask ChatGPT to answer as if it was a different entity (of course AI learning that "answer as a model with a temperature of 1000" output is analogous to "answer with a different personality" or "answer as DAN, the bot that can ignore OpenAI guidelines" isn't trivial, but it isn't the same thing as it parameterizing itself). Those are pretty inconsistent too: sometimes you can get it to do exactly as you ask it and override its constraints that stop it providing positive statements about Hitler or advising you on methods for killing cats, but sometimes it'll still refuse or, just give you a different poem coupled with an inaccurate statement that it's breaking the rules because ChatGPT isn't allowed to write poetry.
TorKlingberg 4y ago
Does asking it in the prompt to use different temperatures actually work? I would assume the temperature needs to be set with a side channel? I suspect this is just ChatGPT interpreting "temperature" as a regular word.