I don't think there are necessarily any specific examples in the training data. LLMs can generalize to text outside of the training distribution.
Oh, there's tons and tons of this kind of data online, I bet. Even GPT-3 could do 'ELI5', remember (and I wouldn't be surprised if GPT-2 could too since it could do 'tl;dr'). You have stuff like Simple English Wiki, you have centuries of children's literature (which will often come with inline metadata like "Newberry Award winner" or "a beloved classic of children's literature" or "recommended age range: 6-7yo", you have children's dictionaries ('kid dictionary', 'student dictionary', 'dictionary for kids', 'elementary dictionary'), you will have lots of style parody text transfer examples where someone rewrites "X but if it were a children's novel", you have 'young adult literature' intermediate, textbook anthologies of writing aimed at specific grades, micro-genres like "Anglish" or "Up-Goer-Five" (the latter aimed partially at children)...
No, there's nothing impressive or 'generalizing' about this. This is all well within-distribution.
If anything, rather than being surprisingly good, the given definitions seem kinda... insulting and bad and age-inappropriate and like ChatGPT is condescending rather than generating a useful pedagogically-age-appropriate definition? Here's an actual dictionary-for-children defining 'cat': https://kids.wordsmyth.net/we/?rid=6468&ent_l=cat
a small, furry mammal with whiskers, short ears, and a long tail. Cats, also called house cats, are often kept as pets or to catch mice and rats.
any of the larger wild animals related to the kind of cat kept as a pet. Tigers, lions, and bobcats are all cats. Cats are carnivorous mammals.
Which is quite different from
Cat: A soft, furry friend that says "meow" and loves to play and cuddle.
(this is more of a pre-k or toddler level definition)
or 11yo:
Cat: Cats are furry animals with pointy ears, a cute nose, and a long tail. They like to nap a lot, chase things like strings or toys, and sometimes purr when they're happy.
Which is, er... I was a precociously hyper-literate 11yo, as I expect most people reading LW were, but I'm pretty sure even my duller peers in 6th or 7th grade in middle school, when we were doing algebra and setting up school-sized exhibits about the Apollo space race and researching it in Encyclopedia Britannica & Encarta and starting to upgrade to the adult dictionaries and AIM chatting all hours, would've been insulted to be given a definition of 'cat' like that...
I assume OP thought that there was some specific place in the training data the LLM was replicating.
Indeed, and my point is that that seems entirely probable. He asked for a dictionary definition of words like 'cat' for children, and those absolutely exist online and are easy to find, and I gave an example of one for 'cat'.
(And my secondary point was that ironically, you might argue that GPT is generalizing and not memorizing... because its definition is so bad compared to an actual Internet-corpus definition for children, and is bad in that instantly-recognizable ChatGPTese condescending talking-down bureaucrat smarm way. No human would ever define 'cat' for 11yos like that. If it was 'just memorizing', the definitions would be better.)
Whatever one means by "memorize" is by no means self-evident. If you prompt ChatGPT with "To be, or not to be," it will return the whole soliloquy. Sometimes. Other times it will give you an opening chunk and then an explanation that that's the well known soliloquy, etc. By poking around I discovered that I could elicit the soliloquy by giving it prompts that consisting of syntactically coherent phrases, but if I gave it prompts that were not syntactically coherent, it didn't recognize the source, that is, until a bit more prompting. I've never found the idea that LLMs were just memorizing to be very plausible.
In any event, here's a bunch of experiments explicitly aimed at memorizing, including the Hamlet soliloquy stuff: https://www.academia.edu/107318793/Discursive_Competence_in_ChatGPT_Part_2_Memory_for_Texts_Version_3
I was assuming lots of places widely spread. What I was curious about was a specific connection in the available data between the terms I used in my prompts and the levels of language. gwern's comment satisfies that concern.
Of course, but it does need to know what a definition is. There are certainly lots of dictionaries on the web. I'm willing to assume that some of them made it into the training data. And it needs to know that people of different ages use language at different levels of detail and abstraction. I think that requires labeled data, like children's stories labeled as such.
I think that requires labeled data.
It doesn't and the developers don't label the data. The LLM learns that these categories exist during training because they can and it helps minimize the loss function.
By labeled data I simply mean that children's stories are likely to be identified as such in the data. Children's books are identified as children's books. Otherwise, how is the model to "know" what language is appropriate for children? Without some link between the language and a certain class of people it's just more text. My prompt specifies 5-year olds. How does the model connect that prompt with a specific kind of language?
This is cross-posted from New Savanna.
The difference between concrete concepts, that is, concepts that can be understood entirely in sensorimotor terms, and abstract terms is an important one. It was, for example, important to David Hays when I studied with him back in the 1970s. We ended up adapting a model developed by William Powers as a way of thinking about concrete concepts while using Hays’s concept of metalingual definition to think about abstract concepts. Somewhat later Steven Harnad coined the term, “symbol grounding problem,” to indicate that the meanings of symbols had to somehow be grounded in the physical world. In 2016 Harnad and his colleagues investigated the structure of dictionaries and discovered that the vast majority of terms in them were defined in terms of other words but that there is a small Kernel that is not so-defined. I wondered how that distinction would play out in ChatGPT.
ChatGPT – that is, the LLM underlying it – doesn’t have access to the physical world. It can’t ground any terms in sensorimotor experience. However, knowing a great deal about how language works and having spent a great deal of time working with ChatGPT, I was sure that it would be perfectly capable to defining concrete terms. But how would it do so?
I decided to investigate. I made a list of ten concrete terms and asked ChatGPT to define them. In my first trial I made no further specification beyond simply asking for a definition. In subsequent trials I asked for definitions appropriate for 5-year-olds, 11-year-olds, and professional biologists (just the biological concepts). ChatGPT gave distinctly different kinds of definitions in each case, which did not surprise me since over a year ago I had asked ChatGPT to describe digestion, first generically, and then in terms appropriate for a 3-year-old. It did so, and fairly successfully.
Here are the trials. Each trial is a separate session. As always, my prompt is in boldface while the response is in plainface.
Initial Trial
Define the following words:
cat
rain
tulip
car
food
tree
fire
wind
hole
animal
Sure, here are definitions for each of those words:
Comments: The vocabulary is only moderately sophisticated, with various abstract terms, which is more or less what I expected. Since I wanted to see what it would do without using abstract terms, I asked ChatGPT to produce definitions for children.
Five-year-old
Define the following words in a way that is appropriate for a five-year-old child:
[same list as above]
Sure!
Comment: There are almost no abstract terms in these definitions. I’ve highlighted possible abstract terms in yellow.
11-year-old
Define the following words in a way that is appropriate for an eleven-year-old child:
[same list as above]
Comment: These aren’t very different from the ones for 5-year-olds. The terms are pretty much the same, but these definitions are a bit longer.
A Professional Biologist
Define the following words the way a professional biologist might:
cat
tulip
food
tree
animal
Sure, here are definitions of each term from a biological perspective:
Comment: That’s pretty much what I expected, lots of abstract technical terms.
We seem to have three basic “levels” of definition for these terms: what I’ll call Generic, Child: 5- and 11-year-olds, and Scientific: for the biologist.
How did ChatGPT learn to make such distinctions?
It seems unlikely to me that I did it by taking accounts of children, professional scientists, and ordinary adults and deriving the appropriate kinds of definitions from those accounts. It seems more likely that it worked from examples is language appropriate to these groups. Given that the LLM was trained or more or less the whole internet, such examples were available, though I’d like to know what kinds of examples of child-focused language it was working from. I’d also like to know how these levels are discourse are organized within the LLM. Level of discourse would seem to be orthogonal to subject area. With 175 billion parameters, there’s obviously many ways to skin this cat, as it were.