Bina 2.0: Why AI Won’t Replace Us

The latest iteration of AI – specifically OpenAI’s ChatGPT product – has created an explosion of global proportions, with the fastest take-up of any online product ever (a million users signed in just five days). Users can engage with it conversationally in the way science fiction has previously imagined intelligent robots. Science fiction writer Arthur C Clarke famously said in 1962 that “any sufficiently advanced technology is indistinguishable from magic”, and ChatGPT certainly looks like magic. It successfully passed the Turing Test, a test of a computer’s ability to communicate indistinguishably from a human, which is considered a test of ‘intelligence’.

This revolution has raised many questions: Has AI finally crossed the point that will lead to the replacement of humanity by computers? If it does, what does it mean for employment? If we are no longer needed, what would we do all day? And if we are no longer needed, what makes us uniquely human? Can computers develop sentience or a soul? Will the computers ultimately take over? Should I watch The Terminator yet again but not as science fiction?

These days it’s very quick to jump from magical technology to apocalyptic scenarios and existential angst. However, we can instead look to concepts within the Torah that shed light on such contemporary questions. First, it always helps to take a deep breath and understand a little more about AI, generative AI, and AGI.

What exactly is AI?

When people use the term “AI” today, they are most commonly referring to “conversational AI” or “generative AI” programs such as ChatGPT. Those forms of AI use something known as “large language models” – an underlying software technique that allows computers to converse. It’s important to know the difference between all of these terms so as not to be distracted by the hype that surrounds them.

This is distinct from AGI – “Artificial General Intelligence” – a far more advanced intelligent computer that can learn to accomplish any task that a human can do. Such a technology could in theory be able to reason, make judgements, have knowledge and much more. It can border on being sentient, which is what people have termed “the singularity”: the point at which technology growth is out of control. The good news is that (a) this is entirely hypothetical, and (b) current AI like ChatGPT are not even close, as I will explain.

What are Large Language Models (LLMs)?

The key to getting over the ChatGPT hype is to understand what LLMs are, and a great way to understand them is through the lens of the Sefirot – the divine faculties described in Kabbalah through which God is revealed, and which are reflected in the human condition.

An LLM takes input: a question or brief typed into a chatbot, and produces output: a response. It examines the sequence of words in the input, and refers to its internal model built from being trained on almost everything on the internet. How do these models actually work?

A great way to understand them is to consider something similar but far more simple – predictive text on your phone. Let’s say you start writing a message to someone with the word “good”. Your phone has an internal model that knows what words most commonly follow “good” at the start of a sentence – usually “morning” or “night”. That’s simple. Let’s say you choose “morning”; what happens next? My phone came up with “my”. Why? Because most often if we say something beyond just “good morning”, it’s followed by a preposition like “my” and then an adjective like “friend” or “love”. In that way, I can start by just typing “good”, and just relying on the options presented by predictive text, could choose from the options presented by my phone to come up with “good morning my dear friend”.

Predictive text like this is relatively simple. For the popular words and phrases that people type on their phones, they would store the common words that are likely to immediately follow. Rules like “word A is usually followed by either word B, C, or D” come from the huge volume of text people type into messaging apps around the world. In AI terminology, this is a language model ‘trained’ on some text. Your phone might have a few thousand rules like this, and these rules are known as ‘parameters’.

When we take these concepts and scale them up using huge computing resources, we get “large language models”. Unlike your phone, ChatGPT has about 175 billion parameters. These were created by training the model on about 570Gb of datasets scraped from the internet – web pages, books and more. This is the equivalent of more than three times the amount of text in the Library of Congress. How many books have you read in your life? ChatGPT has done the equivalent of reading 1.3 million books to learn how to converse with humans on a vast variety of topics and produce meaningful responses.

Personification and Metaphors

Note that I have used terms like “know”, “think”, “learn”, and “read” when describing attributes of a computer program. This is something we have a habit of doing in part because technology is so embedded in our lives that we describe it as having human characteristics, and also because they are a convenient shortcut. We say “my phone has died” but it was never alive and cannot die. Rather, it’s a short way of saying my phone has stopped working and hasn’t immediately started working again. Saying “Google thinks I’m a NY Giants fan” is similarly incorrect and lazy. Google has used your browsing and search history to determine a high probability that you want to read more about a particular sports club.

Using this sort of language about a computer and experiencing the sort of conversation and interaction with a computer that feels significantly more human than any previous interaction might lead us to think that the program has now developed the ability to think, develop ideas or construct sentences in the way we humans do. In fact, they have not done anything of the sort.

LLMs in Kabbalah Terms

Kabbalah explains that there are three intellectual faculties: chochma, bina and da’at. which form a hierarchy or a progression.  Very briefly, chochma is the spark of an idea; bina is the development of that idea into something substantial, and the derivation of one idea from another; da’at is when those developed ideas are integrated and assimilated. Each faculty has certain capabilities in an of itself, and also to the extent that it links to the others (and the faculty of da’at is the bridge to the emotional faculties chessed, gevurah, etc).

LLMs essentially take a bunch of words and try to guess what word should come next. They are based on an existing body of text and draw on that to generate new text. Because they have done this at a scale that has never been achieved, it allows them to actually converse. But what they are really doing is mimicking human language and understanding. That inherently will not allow them to come up with anything new – rather something based on something already written and used to train it.

In Kabbalistic terms, LLMs have programmed the intellectual faculty of bina. The definitive characteristic of bina is “l’havin davar mitoch davar” – to understand or derive one thing from another. Based on all the text that the LLM has been trained with, it can derive responses to questions. The faculty of bina cannot come up with anything new – that is the realm of chochma. If a LLM comes up with something that appears new, it actually isn’t. Rather, there’s a good chance that you’ve asked it to do something it’s already seen before in the billions of examples used to train it.

As part of researching this article, the obvious thing to do was to ask ChatGPT to write it. I used two prompts: one which asked it to compare and contrast LLMs with chochma and bina, and another which asked it to compare LLMs to bina. In both cases, it was able to construct a coherent thousand or so words talking around the concepts, but the connections it made were superficial and not particularly insightful. It took what it already knew and presented it in a different form. It provided some useful phrases about LLMs which I used, and at the same time showed me its limitations.

Chochma, Bina and LLMs

With this Kabbalistic perspective on LLMs and the intellectual faculties, we can gain a greater insight into the nature of both chochma and bina.

As a way to think about the difference between chochma and bina, and how a  human might approach a task compared with technology, imagine a LLM trained on the entirety of Jewish responsa documented by Rabbinic authorities over hundreds of years. You could ask it any question, and it would be able to draw every documented precedent to identify which are relevant and why, and consider how to rule. With access to more data than any human could remember and consider, could such an AI answer complex questions of halacha better than us? Is this Rabbi 2.0? No. The process of halachic ruling is itself far more than bina – which precedents best match the case. It has elements of chochma – looking at a precedent with a different perspective or finding a connection from one domain to another, and of course shimush – using context and emotions to determine what level of stringency or leniency are appropriate to apply. That said, a Rabbi could use such an LLM to either support their ideas about how to rule, or inform on considerations of which they were not aware.

Further, LLMs use technology to mimic just one aspect of bina. However, because they can be trained on so much more text than humans can possibly absorb and recall, they illustrate how technology can supercharge the faculty to achieve far more than we could (using just that faculty alone). What is astounding about LLMs is that they can do this so well that it can mimic chochma – effectively “indistinguishable from magic”.

True chochma is something qualitatively different – a flash of inspiration. With a slightly different reading of Job 28:12 “and chochma emerges from nothingness”, Kabbalists explain that the word “nothingness” alludes to the Godly creative process that makes something from nothing. God drew upon the divine faculty of chochma to create the world. That same creative spark is within each one of us. Ideas appear to come from “nothing”. We call it a “spark” because it can feel like lightning, coming from nowhere and suddenly shedding an entirely new light on something previously obscure.

ChatGPT-3 cost millions of dollars firstly to build a powerful enough computer, then to train, and then to operate daily. Every single query costs (in computing power) about 0.2 cents, which is over twenty times the ‘cost’ of a google search. All that to deliver something that mimics a quality we all possess through one part of our innate faculty of bina.

Even to the extent that LLMs and their underpinnings (like neural networks) seek to mimic the way a human brain works, they do not get close to the level of efficiency of the human body. Consider the cost in food and drink to ‘power’ a human brain compared to its capabilities.

Where next for humans?

The human faculties – intellectual and emotional – are a mirror of the divine. We can examine one to gain insights into the other. The framework of chochma and bina helps us see both the power and limitations of LLMs and conversational AI. While they can do certain some things far better and faster than humans, we still maintain a clear advantage when it comes to the creative spark of chochma.

This technology will replace some of what we do – entire jobs will disappear. Technology has done this in the past, such as during the Industrial Revolution, and we have adapted. New jobs have and will be created. Could ChatGPT produce another sequel to the Fast & Furious franchise? Undoubtedly, but that says more about the franchise than humans or AI. Creativity and emotion can be mimicked or faked using technology, but in their true form, they are uniquely human. Rather than fear AI, we should celebrate humanity.

—

A version of this article was first published in the Melbourne Young Yeshivah magazine #37, RH 5784