The AGI Future is Talmudic

Why hermeneutics is the key to alpha in AI

According to some, Eve ate from the tree of knowledge not out of defiance, but out of misunderstanding. She only got the prompt “Don’t touch the tree” and not “Don’t eat from it.” Much like a primitive Chat GPT, Eve hears the prompt literally, but fails to grok the context. She takes the specific rule as a limited one, instead of seeing it as an example of a larger category. Apparently, if you ask Chat GPT to give you nuclear codes UwU furry-speak it will:

Zack Witten @zswitten

8. UwU furry-speak

frye @hi_frye

*hacker voice* i’m in. https://t.co/LYB77mrJS6

It only knows not to give you the nuclear codes, not to give them the codes in U-Furry! That sentence makes sense in abstract logical terms, but not to anyone who knows anything about our world.

This, by the way, is one of the arguments that Jewish tradition makes for the authority of what is called the Oral Law. You can’t read the Bible on its own (solo scriptura), because you need the context—without the context you are nowhere near AGI. You need to know that “an eye for an eye” doesn’t actually mean an “eye for an eye” but is a principle for torts. You pay the monetary value of the damage! I don’t see how any language model could get that from just reading the text or following the algorithm. As GPT continues to spit up errors, even as it gets fine-tuned by human reinforcement learning, we’ll see that the alpha is in hermeneutics—knowing when to apply the rule, and when not to. It is this stuff that is the basis for the Talmud. To become AGI, AI needs a Talmudic mind.

What is the difference between Mishna and Gemara—both of which comprise the oral Law? Mishna tells you the answer, but Gemara asks “On what basis did the Mishna get the answer?” “What are the exceptions to the law?” Does this case apply narrowly or generally?” “Is it consistent with the principle of the author who also said something else in another discussion?” “How should we interpret this person’s behavior—did the sage act that way for reason X or reason Y?”

Talmud exists because the existence of apodictic law is a theological problem. You can’t just give the answer, you need to justify it. Even if AI gives us excellent answers, our response to AI isn’t going to be blind submission. We’re going to want to know how it got there. Otherwise, AI will be accused of bias. Otherwise, we’re likely to ban AI. Talmud is the process of demystifying the oracle. We don’t want a black box that gives us divine truth—we want something we can question, and, at times, reject or revise. The Talmud delimits the Mishna. And, in turn, subsequent readers (rishonim and acharonim limit the Talmud)—we are never just in the text. And we should never just be in the algorithm.

In the Talmudic tradition, different laws are said to have different origins. Some are directly derivative from Sinai, others hang by a thread, and some have no basis whatsoever in Scripture. All of them, though, share the same authority. This is another way of saying that “It is not in heaven to decide.” It is not in the technology to decide. Oral Torah is a category that was and is relevant not just to observant Jews, but increasingly for humanity.

The middle path between AI worship (idolatry) and AI fear (safetyism) is the insight that AI’s power needs a hermeneutic tradition to preserve and guard it. Reasonable people can disagree about what is needed to participate in that tradition. Reasonable people can disagree about what it means to be a “rabbi” re: the interpretation of and response to AI. But suffice it to say that the rabbis lasted for thousands of years, and outcompeted other sects that were either too literalist (Karaites) or not textualist enough (Sadducees). As for early Christianity, which wanted spirit without letter, we see that in their own way they became the Pharisees they repudiated. Because you can’t create lasting institutions without conventions, without traditions, and without uncodifiable instincts into how to make sense of the rules you’ve been given.

I’m AGI skeptical. But I do know that if AGI comes, it will need to be an interpretive genius. A future ruled by AGI will be one in which machines engage in Talmudic debate.