The LLM Problem
tbray.org
tbray.org
But they are fancy expensive Markov chains. The learning here is that fancy expensive Markov chains can do a lot, not that the statement is wrong.
Just as simplified does not mean reduced. It can align and help understanding, even if incomplete.
Just as munged does not mean increased. It can contextualize understanding, even if incomplete
But that's only if you don't use BEAM search, which LLMs do use. If you use BEAM search with beam=4 (typical for LLMs), for example, then it does a tree search where it keeps track of the top 4 highest probability outputs, then I guess it's not a markov chain process anymore but it just uses a markov chain.
By all means, please write the brilliant take yourself, and I will happily upvote and reference it.
As in, are we overestimating what we are?
An LLM can only "exhibit" the behavior that humans encode into text. We encode explicit behavior, including symbols and grammar, and we encode implicit behavior, including narrative and logic.
An LLM blindly models the patterns of text, but doing that exposes the power of the patterns themselves.
The only problem is that an LLM can't determine which patterns are useful and which patterns aren't.
1) Running an agency, it was quick and easy to share ideas and give direction and have that executed to a fairly accurate degree over a decent period of time.
2) Working with GPT-3 & GPT-4, it’s also quick and easy to share ideas, but I’m becoming more aware of how surface level my communication is. When I get back great results, it’s typically because I’m requesting busy work. When I’m requesting something novel, it quickly becomes clear what I forgot to define.
So, the idea that’s formulating around LLMs is that the process of transforming an idea into instructions will become much more desirable. And that we’re moving from selecting for people who know how to do the work, to selecting for people who know how to request the work. And that those are two very different skill sets.
But yes, there are plenty of demonstrations of LLMs using themselves to accomplish tasks or even recruiting humans to accomplish a task.
Like Mechanical Turk, you need to "program" by giving English instructions and the results can be (depending on the instructions) non-deterministic.
But Mechanical Turk did not revolutionize computing. Can a faster and cheaper version do so? Sometimes incremental improvements turn into paradigm shifts. But sometimes not. I guess we'll see (which is another lukewarm take--sorry).
Who's behavior? Whoever wrote the training corpus and the prompt. So far, that definitely means one or more humans.
The problem is that nearly everything you have read about LLMs has personified them. The character of an LLM as a person is invented, then conclusions about LLMs are drawn from that character.
Nearly every thing that is interesting about LLMs is not actually an LLM feature: it's a language feature. Language is an incredibly powerful tool, and an incredibly versatile data format.
What is the data? Human behavior.
When a human writes text, they don't just spit out characters at random: humans explicitly choose the characters they write, and implicitly choose the characters they don't write.
In the most literal sense, text contains the entropy of a human making a string of choices for which character to write next. In a literal sense, that's a 1-dimensional ordered list; just like the string of characters is. Text gets a lot more interesting when we introduce language...
The reason a human chooses one character over another is not 1-dimensional. There are patterns of entropy driving that decision, and we explicitly know about some of them.
We have defined words, punctuation, grammar, idioms, etc. that we structure text with. We call these patterns language. These are the explicit patterns that allow us to hack the writing process: instead of encoding a 1-dimensional list of decisions, we encode a recursive many-dimensional structure of symbols and grammar.
Now for the interesting part: when humans write language into text, our intentions become explicit patterns, but they aren't the only patterns that end up there.
Every arbitrary decision about writing style gets implicitly encoded as a pattern of negative space. The reasons why we write one concept instead of another get implicitly encoded, too. It's a bit lossy, but most of these patterns end up in the text.
---
So what does an LLM do? It implicitly models patterns, and presents them.
That's it. That's the only behavior. A "continuation" is made by modeling the prompt, and showing what's "next" in the model.
An LLM explicitly knows nothing at all; but implicitly, it knows everything. It models every pattern that exists in the text, but it doesn't have a clue what or why it is modeling. A pattern is a pattern.
The whole system ends up generating valid language, because most of the implicit patterns an LLM models "align to" the explicit patterns humans (language) encoded into the text.
The whole system ends up "exhibiting behavior", because it also modeled the implicit patterns of human behavior (narrative) that were are also encoded into the text. Patterns of narrative determine which part of the model a prompt explores, and which pattern is considered "next".
---
Next time you read about the "features" and the "limitations" of an LLM (including GPT-4), know that you are really reading about the "features" of language itself, and the "features" of the narratives that were written in the training corpus and prompts.
The behavior that an LLM exhibits has much less to do with how well its model "aligns with" language grammar, and much more to do with how well the text itself (and the narrative it contains) will behave.
"Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something."
What is the language of the human mind? Of consciousness itself? If a LLM can learn that, perhaps we will build a technology that takes an intention and is immediately able to manifest the corresponding changes to reality to satisfy the intention or desire? When we get there, perhaps it will be time to look inward and do some of that spiritual inner work we keep putting off?
What’s exciting to me about LLMs is they seem to be one step closer to this vision, with all the peril and all the possibility that comes with it.
First, quantized in the physics sense [0] doesn’t mean discrete states. And discrete doesn’t necessarily mean finite.
But even if you mean “finite”, I believe that you are underestimating how staggeringly large “finite” can be.
You’re right, when I say quantized I do mean finite, countable, represents me symbolically. I realize how staggering large finite can be.
We are in the midst of exponential growth in our ability to manipulate reality. The mastery of physical reality feels achievable to me now. I could be wrong, but I could also be right. It’s a fun time to be alive.