[Avodah] AI and Jewish Law

[Micha Berger]

On Thu, Sep 07, 2023 at 05:30:43PM -0400, Joel Rich via Avodah wrote:
> Let me just articulate one issue which I haven't really seen discussed. As
> I understand it AI assumes something along the lines of that the problem
> was presented to the decider and he went through an intellectual process
> and came up with the result. While this may be often true I can't help
> but of [RYBS]'s comments in [C]ommunity, [C]ovenant and [C]onversation
> where he
> says he knew the result for the question concerning drafting chaplains,
> and just had to come up with a logic to support it...

Again, today's AI, GPT and other Large Language Models, are doing neither.
There is no reasoning out of concepts.

Words are indeed mapped to vectors, sets of numbers that relate to the
words' meaning. But they are manipulated as tokens of language. That's
why they're large language models, not large world models.

What you get from chatGPT is something written following the patterns,
meta-patterns, meta-meta-patterns... of the texts it was trained on,
brought to the words in the prompt.

It produces results that usually seem like it knows what it's talking
about, because the taraining texts generally make sense. But that's why
at times it fills in the pattern with nonsense. Hence what people call AI
"hallucinations".

There is no parallel to logical analytics vs. gestalt thought in any
of this.

(And I still think no large language model should be all that heavily
used as a tool for a poseiq until you train one on a lot more teshuvos,
shas and posqim than on general meterial. You have to get the right
meta-patterns dominating.)

Tir'u baTov!
-Micha

-- 
Micha Berger                 The same boiling water
http://www.aishdas.org/asp   that softens the potato, hardens the egg.
Author: Widen Your Tent      It's not about the circumstance,
- https://amzn.to/2JRxnDF    but rather what you are made of.