Good read. I didn't know that the redemption of the firstborn was still obligatory on someone who hadn't had it done. I guess I need to ask my parents if they did it (probably not, my father's a gentile and my mom's Reform), and if not, do it at some point.
Why is a it so hard for people to write realistic assessment of deep learning tools such as LLMs and always have to add layers of hype, when in fact the things it can do and will do in the future are quite useful? In the end of the day, your analysis over-hypes and in reality just reinforces things which anyone who looks closely into this technology and doesn’t buy into the venture-fueled hype frenzy knows:
1. LLMs are great at ingesting vast sums of data and spitting them back at you in often useful ways - essentially they are the next generation of search engines which not only give you sources but can be more concrete in their answers
2. OTOH there is nothing creative about LLMs. They are just statistical models that can take vast sums of text and distill them and spit back what they already have seen in some source, and that includes “arguments”. And yes they can do other time saving things with those texts like translations or creating audio or video (but all these areas are still variants on statistical models of pattern recognition). Personally I am quite skeptical they can ever be creative in the sense of generating new ideas, because we currently have no way of modeling creativity mathematically, and we probably never will have such models. Nonetheless, they certainly already can augment our creativity by reducing the time it takes to do research and other text-based tasks.
3 The biggest downside is LLM usefulness is that it is directly related to the quality of the sources. For example, they are great at helping you with the dull, repetitive tasks of coding, because they have been fed tons of code by coders who create these LLM. That’s great as far as it goes, but even here they aren’t going to help you create new types of applications. They will just help you with the dull, repetitive tasks. And even then, they will be far better at helping you with popular languages like Python and far less helpful with say, R or Lua or Julia (although even with those, they are helpful)
4. Which leads right to the next point: they are most helpful for things which have been fed huge amounts of quality sources (e.g. code ). For niche topics (and let’s face it Halacha is super niche) they are far more limited. In fact, for anything where most of the relevant information is found in books, not on websites (and this covers most academic topics), they will prove far less useful and even misleading (since their input sources are limited
5. That doesn’t mean with great investment of time and money, you can’t build quality LLMs for niche topics. Sefaria, for example, is at least partially embarking on such an endeavor. But considering that niche topics will have available far less (by orders of magnitude) financial and hence programming resources to create these LLMs, and also far less text sources, it’s unlikely that progress in creating such LLMs will move quickly and in a satisfying direction. Perhaps they will reach a state where they are moderately useful, but you can’t count on that happening. So the state of LLMs in non-niche topics is no harbinger for the future of niche topics.
6. In any case,it bears repeating that LLMs are currently only useful as research aids, even in non-niche topics. Whether they ever will be more than that is hard to know, but that will require far more progress in creating new types of models, which the current state of research in this area have not yet provided.
FWIW, it's very common for trans and queer Jews to analogize the ger "having a Jewish soul" with a transgender person's sense of themself as being born with a different sex than their body would indicate. I'd be shocked if Sefaria didn't have source sheets with this perspective. I'm surprised that Laynie Solomon didn't recognize it immediately.
You note that Deep Research makes heavy use of VBM and Halachipedia. But in looking at the responses it produced, it seems like chabad.org and shulchanaruchharav.com (both of which reflect positions held by Chabad) are well represented, too. I wonder what Chabad's early and thorough adoption of web presence means for their current and future influence on AI-produced Torah.
Good read. I didn't know that the redemption of the firstborn was still obligatory on someone who hadn't had it done. I guess I need to ask my parents if they did it (probably not, my father's a gentile and my mom's Reform), and if not, do it at some point.
This was a great read, thank you. Your reflection at the end was quite moving, maybe because it reminds me of a thought I had back in 2015 as Big Data was first becoming a thing. Exciting times! https://www.jpost.com/opinion/on-leaders-and-followers-in-the-age-of-big-torah-419336
Why is a it so hard for people to write realistic assessment of deep learning tools such as LLMs and always have to add layers of hype, when in fact the things it can do and will do in the future are quite useful? In the end of the day, your analysis over-hypes and in reality just reinforces things which anyone who looks closely into this technology and doesn’t buy into the venture-fueled hype frenzy knows:
1. LLMs are great at ingesting vast sums of data and spitting them back at you in often useful ways - essentially they are the next generation of search engines which not only give you sources but can be more concrete in their answers
2. OTOH there is nothing creative about LLMs. They are just statistical models that can take vast sums of text and distill them and spit back what they already have seen in some source, and that includes “arguments”. And yes they can do other time saving things with those texts like translations or creating audio or video (but all these areas are still variants on statistical models of pattern recognition). Personally I am quite skeptical they can ever be creative in the sense of generating new ideas, because we currently have no way of modeling creativity mathematically, and we probably never will have such models. Nonetheless, they certainly already can augment our creativity by reducing the time it takes to do research and other text-based tasks.
3 The biggest downside is LLM usefulness is that it is directly related to the quality of the sources. For example, they are great at helping you with the dull, repetitive tasks of coding, because they have been fed tons of code by coders who create these LLM. That’s great as far as it goes, but even here they aren’t going to help you create new types of applications. They will just help you with the dull, repetitive tasks. And even then, they will be far better at helping you with popular languages like Python and far less helpful with say, R or Lua or Julia (although even with those, they are helpful)
4. Which leads right to the next point: they are most helpful for things which have been fed huge amounts of quality sources (e.g. code ). For niche topics (and let’s face it Halacha is super niche) they are far more limited. In fact, for anything where most of the relevant information is found in books, not on websites (and this covers most academic topics), they will prove far less useful and even misleading (since their input sources are limited
5. That doesn’t mean with great investment of time and money, you can’t build quality LLMs for niche topics. Sefaria, for example, is at least partially embarking on such an endeavor. But considering that niche topics will have available far less (by orders of magnitude) financial and hence programming resources to create these LLMs, and also far less text sources, it’s unlikely that progress in creating such LLMs will move quickly and in a satisfying direction. Perhaps they will reach a state where they are moderately useful, but you can’t count on that happening. So the state of LLMs in non-niche topics is no harbinger for the future of niche topics.
6. In any case,it bears repeating that LLMs are currently only useful as research aids, even in non-niche topics. Whether they ever will be more than that is hard to know, but that will require far more progress in creating new types of models, which the current state of research in this area have not yet provided.
FWIW, it's very common for trans and queer Jews to analogize the ger "having a Jewish soul" with a transgender person's sense of themself as being born with a different sex than their body would indicate. I'd be shocked if Sefaria didn't have source sheets with this perspective. I'm surprised that Laynie Solomon didn't recognize it immediately.
You note that Deep Research makes heavy use of VBM and Halachipedia. But in looking at the responses it produced, it seems like chabad.org and shulchanaruchharav.com (both of which reflect positions held by Chabad) are well represented, too. I wonder what Chabad's early and thorough adoption of web presence means for their current and future influence on AI-produced Torah.
Too bad AI is puerile bullshit that can't die off fast enough.
https://open.substack.com/pub/marlowe1/p/hey-man-by-tim-lieder-fiction-extra?r=sllf3&utm_campaign=post&utm_medium=web&showWelcomeOnShare=false