4 ms·

I don't think there is any reason to believe this except that everyone seems to want it to be true. An easy way to make it not be true would be to emphasize so

I don't think there is any reason to believe this except that everyone seems to want it to be true. An easy way to make it not be true would be to emphasize some sources in pretraining by putting them in the corpus multiple times.

dleeftink 2y ago
Maybe not 'median' but rather 'sufficiently representative', as with all distributional semantics, given a large enough corpus we can approach the 'true' distribution of word/phrases in a given language.
- krainboltgreene 2y ago
  Except the corpus itself is fractional of all media. This is like saying Twitter is sufficiently representative of all human history.
  - dleeftink 2y ago
    We are truly dealing with immense sizes here, see [0]. I am not saying it is representative of all things, but that large corpora are sufficiently representative of words and phrases in a language (and by extension, the words and phrases assoiciated with topics and ideas). In written form you may miss some slang, pidgins and dialects, but those would be underrepresented in other media as well. [0]: https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-...
miki123211 2y ago
A much better way is to RLHF the LLM until you get the behavior you want. As far as I know, modern LLMs try to strike a balance between being somewhat neutral, while not being too neutral on topics outside of the overton window. They'll give you a "both sides have their good points" argument on abortion, religion, guns or immigration, but won't do that for obvious racism or nazi viewpoints. Early LLMs had a problem with getting this balance right, I feel like many of them were a lot more left-leaning. I don't know how much of the change is caused by us understanding the technology better and how much is just the political winds shifting, though. I felt like we had a moment there when some models were a bit too "well it depends", even on very uncontroversial subjects.
- pjc50 2y ago
  > I feel like many of them were a lot more left-leaning "Reality has a liberal bias"

dleeftink 2y ago

Maybe not 'median' but rather 'sufficiently representative', as with all distributional semantics, given a large enough corpus we can approach the 'true' distribution of word/phrases in a given language.

krainboltgreene 2y ago
Except the corpus itself is fractional of all media. This is like saying Twitter is sufficiently representative of all human history.
- dleeftink 2y ago
  We are truly dealing with immense sizes here, see [0]. I am not saying it is representative of all things, but that large corpora are sufficiently representative of words and phrases in a language (and by extension, the words and phrases assoiciated with topics and ideas). In written form you may miss some slang, pidgins and dialects, but those would be underrepresented in other media as well. [0]: https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-...

miki123211 2y ago

A much better way is to RLHF the LLM until you get the behavior you want. As far as I know, modern LLMs try to strike a balance between being somewhat neutral, while not being too neutral on topics outside of the overton window. They'll give you a "both sides have their good points" argument on abortion, religion, guns or immigration, but won't do that for obvious racism or nazi viewpoints. Early LLMs had a problem with getting this balance right, I feel like many of them were a lot more left-leaning. I don't know how much of the change is caused by us understanding the technology better and how much is just the political winds shifting, though. I felt like we had a moment there when some models were a bit too "well it depends", even on very uncontroversial subjects.

pjc50 2y ago
> I feel like many of them were a lot more left-leaning "Reality has a liberal bias"