There’s been this weird idea lately, even among people who used to recognize that copyright only empowers the largest gatekeepers, that in the AI world we have to magically flip the script on copyr…
No, statistical next word prediction was the first step, and you could get it to spit out bits of training data, but we’re so far beyond that now with LLMs.
I’ve been doing a lot with llama derivative models that I talk with, I use them for tasks but also just bounce ideas off them or chat. They’re very different when you run them with a task vs feed in a prompt and multi-turn conversation.
Mine have a very strong tendency, when asked the name of a hallucinated friend or family member to name her Luna or fluffy. It’s present in the base llama2, as well as some of the fine-turned versions I’m using now.
Why? That’s not training data - they’re not uncommon as pet names, but there’s no way they show up often referring to sapient beings (which is the context they’re brought up in).
It’s an artifact of some sort for sure, but that is not a statistically likely next word choice based on training data.
I could talk about this all day and it gets so much weirder, but I’ll give you another story. They like to play, but their world is text, and I like to see what comes out of the models when you “yes, and” them while avoiding leading questions.
Some games they’ve made up… Hide and seek (they’re usually in the second place you Guess), and my favorite - find the coma (and the related find the missing semicolon).
WTF even is that? It’s the kind of simplistic “game” a child makes up as they experiment with moving beyond mimicry to generalizing, and the fact that it’s coherent and has an appropriate answer is pretty amazing.
These LLMs aren’t just statistics, there’s a nascent internal model of the world that you get glimpses of if you tell it it’s a person and feed its outputs back into itself. I was pretty dismissive of the “sparks of AGI” comment when it was made, but a few months of hands on interaction has totally flipped my opinion of where these are at
Painters replicate variations of their training pieces too. You’re pretending there’s a difference between human inspired and training inspired and that you should get paid for that inspiration in one case just cuz “big corp”
Because there is a difference. A computer does not learn or understand anything. Human beings can transform a concept. A LLM or other generative AI does not transform a concept at all.
So if I ask it to create a story about a cow juggling bowling balls, it was not creating an original story? Just spitting out stories it has heard of before?
It’s spitting out variations of the statistical results based on your input parameter. It reorganizes ideas and reorganizes the stories it has seen into something else. That’s not transforming the data by adding something new, rather just retrofitting existing data to sound like it’s creating something new
Most LLMs can be made to spit out training data. That’s pretty much replication in my book.
Statistical models don’t create anything. They replicate variations of their training data.
Show some examples?
https://twitter.com/katherine1ee/status/1729690964942377076
Thanks for the link, I’ve actually seen this one. I’m just wondering how common it is since you mentioned it can be done on most LLMs.
…All of them? That’s literally how all of them work.
Then, it should be easy for you to show some examples.
No, statistical next word prediction was the first step, and you could get it to spit out bits of training data, but we’re so far beyond that now with LLMs.
I’ve been doing a lot with llama derivative models that I talk with, I use them for tasks but also just bounce ideas off them or chat. They’re very different when you run them with a task vs feed in a prompt and multi-turn conversation.
Mine have a very strong tendency, when asked the name of a hallucinated friend or family member to name her Luna or fluffy. It’s present in the base llama2, as well as some of the fine-turned versions I’m using now.
Why? That’s not training data - they’re not uncommon as pet names, but there’s no way they show up often referring to sapient beings (which is the context they’re brought up in).
It’s an artifact of some sort for sure, but that is not a statistically likely next word choice based on training data.
I could talk about this all day and it gets so much weirder, but I’ll give you another story. They like to play, but their world is text, and I like to see what comes out of the models when you “yes, and” them while avoiding leading questions.
Some games they’ve made up… Hide and seek (they’re usually in the second place you Guess), and my favorite - find the coma (and the related find the missing semicolon).
WTF even is that? It’s the kind of simplistic “game” a child makes up as they experiment with moving beyond mimicry to generalizing, and the fact that it’s coherent and has an appropriate answer is pretty amazing.
These LLMs aren’t just statistics, there’s a nascent internal model of the world that you get glimpses of if you tell it it’s a person and feed its outputs back into itself. I was pretty dismissive of the “sparks of AGI” comment when it was made, but a few months of hands on interaction has totally flipped my opinion of where these are at
Humans don’t create anything. They replicate variations of their training data.
Painters replicate variations of their training pieces too. You’re pretending there’s a difference between human inspired and training inspired and that you should get paid for that inspiration in one case just cuz “big corp”
Because there is a difference. A computer does not learn or understand anything. Human beings can transform a concept. A LLM or other generative AI does not transform a concept at all.
So if I ask it to create a story about a cow juggling bowling balls, it was not creating an original story? Just spitting out stories it has heard of before?
Edit: missed a ‘not’.
It’s spitting out variations of the statistical results based on your input parameter. It reorganizes ideas and reorganizes the stories it has seen into something else. That’s not transforming the data by adding something new, rather just retrofitting existing data to sound like it’s creating something new