GiladJun 9, 2023

ואמר רבי אלעזר אמר רבי חנינא: כל האומר דבר בשם אומרו מביא גאולה לעולם

Question: Does this apply only to humans? What about quoting an #AI inspired Dvar #Torah?

In other words is the point to give credit to the originator of the idea to honor him (which presumably only applies to people)? Or possibly to teach you humility (this lesson would be taught even if the idea came from a machine)?

#mazeldon #jewniverse

Show threadMoshe ZadkaJun 9, 2023

@gilad73 "AI" is probably a bad category.

If you text and use autocompletion, do you attribute your phone's keyboard's autocomplete algorithms?

LLMs are basically fancy autocompletes, so I doubt they would fall under this category.

If you train a Reinforcement Learning agent on, say, chess, and then ask for a good chess move, I think you should attribute the answer.

Show threadMoshe ZadkaJun 9, 2023

@gilad73 So my answer would be: "it depends on the agency of the AI". Per the fence rule, we should probably put the lower boundary for how agentic an AI should be before you have to attribute things to it so that we are certain we are not violating the law.

Any AI that has coherent, complex, goals should be one that you attribute things to.

Show threadGiladJun 9, 2023
@moshez interesting distinction. Of, as I understand it, LLMs are trained and later fine tuned by prompts. They don't merely regurgitate the data stuffed into the model, it's using abstract rules from the training to generate new ideas. This should be no different from me taking a Ramban on the Torah and applying it to understand the comments of the Ktzot HaChoshen on Choshen Mishpat. Applying knowledge from one area and using it in a different context is called binah. That's why I specified generative AI
Show threadMoshe ZadkaJun 9, 2023

@gilad73 Again, no more binah than if you usually text people about the Rambam, and then use your phone's keyboard's autocomplete when you text about Rashi. It's also "applying knowledge from one area to another".

I think the mechanics of lacking agency make it unworthy of attribution, not the lack of "generation of new ideas".

Show threadGiladJun 9, 2023
@moshez I think there are qualitative differences between Gen AI and auto complete. Gen AI is capable of original content creation and can generate unique outputs based on given prompts. Auto-complete, on the other hand, assists users by suggesting the most probable next words or phrases based on context and user patterns. I think that's a significant difference.
But you raised a different point about agency, that's an interesting idea I'll need to think about.
Anyway, it's almost Shabbat, so shabbat shalom!
Show threadMoshe Zadka

@gilad73 Hahe, timezones exist, so fine if you don't see this until Motzei Shabbat. Shabbot shalom!

It might interest you to know that under the covers, the way Gen AI works is by....predicting the most probable next token.

Jun 9, 2023 at 4:11pmWeb