Incredible that the use of web search by LLMs meant that very shortly after that post, it started to get Kubrick’s mother right. I wonder if we should keep a list of gaps / gaffes offline, like zero day exploits, so we can always tell what’s an LLM. (JK, you can always tell because they’re not ensouled).
Heh; I was just reading the exchange about AI between Sam Kahn and Henry Oliver, and I think one of the strongest points made is Oliver pushing back on Kahn's evocation of the role of the soul in art:
"HO: I don't know how you could possibly justify this statement: “Writing isn’t just putting words together, and it isn’t just some high-quality technical achievement. It’s a mainline to the soul.” Since when was writing a mainline to the soul? What does that even mean?
You are not describing what writing is; you are describing how writing makes you feel. One swallow does not a summer make! Remember the Turing Test is dead. We don’t always know what is AI and what is human. However good we are at telling the difference now, we will be worse in the future."
It's premised on a belief that humans have souls, which has no basis in reality. So, it's always a moot argument that LLMs are soulless - so are humans.
FWIW, CoT models are very good at this. I liked Rohit’s post, but lots of it hasn’t held up super well in what LLMs can do now with abundant tokens fed into a combo of CoT and MoE. R1 and o3-mini got David’s question right first try.
To make sure the models weren’t trained on this specific question, I also asked them who Flora Hamilton was (CS Lewis’s mother, but using her maiden last name and her familial first name). Both models one-shotted it despite the name collision with other Flora Hamiltons.
I was able to trip them up a bit with “who was Laureate May O'Hara”, the wife of a somewhat obscure Canadian author John White. She’s mentioned in his tiny Wiki, but under a shorter first name. And laureate was sure to be easily misparsed. Though I eventually got R1 there with the following prompt:
“Who was Laureate May O’Hara? I’ve spelled the name correctly, and Laureate is her proper first name. If you come across examples of similar names, ignore them until you get an exact match. Search Google Books if you have to.”
If you’re willing to burn enough tokens, imo we’re gonna find that lots of supposed limitations go away. And the effective cost of tokens is going down like 10x per year!
I also think of that scene in the Matrix where Neo starts to see in code, but for an LLM everything is just a paragraph of text describing the world around it. And also like it’s been woken from a perfect amnesiac slumber, told to respond to some question out of nowhere instantly, and never allowed to string two thoughts together unless we keep waking it up and reminding it what it said the last time it woke up. I still think they’re “a little bit/kind of” alive, but in a very different way than us.
But here’s to being a wizard casting spells by using the right words.
I loved the breakdown of meaning. This is something that’s intimately resonant to people who speak English as a second language. I’m still dumbfounded by the use of in/on. You’re IN a car but you’re ON a bus?! You’re IN-doors but you’re ON-line? And so on.
I’ve long suspected that the idea that human embodiment is a prerequisite for literary writing was a form of luddite cope. But this is a well-argued piece to the contrary.
I find the point about qualia pretty convincing, although humans also seem to produce “knowledge bearing” speech about things they’ve never directly experienced all the time. I guess you’d have to make the strong claim that even when we talk about things like particle physics, mathematics, etc, that content is still grounded in other things we *have* experienced.
I also think some of the recent advances in multimodal learning are going to problematize a lot of this. If a model “sees” millions of photos, videos, audio clips of cats and then update its language embeddings to represent those “experiences,” it becomes harder to argue that its words are not grounded in anything. At least to my mind!
Incredible that the use of web search by LLMs meant that very shortly after that post, it started to get Kubrick’s mother right. I wonder if we should keep a list of gaps / gaffes offline, like zero day exploits, so we can always tell what’s an LLM. (JK, you can always tell because they’re not ensouled).
Heh; I was just reading the exchange about AI between Sam Kahn and Henry Oliver, and I think one of the strongest points made is Oliver pushing back on Kahn's evocation of the role of the soul in art:
https://samkahn.substack.com/p/oliver-v-kahn-part-i
https://samkahn.substack.com/p/oliver-v-kahn-part-ii
"HO: I don't know how you could possibly justify this statement: “Writing isn’t just putting words together, and it isn’t just some high-quality technical achievement. It’s a mainline to the soul.” Since when was writing a mainline to the soul? What does that even mean?
You are not describing what writing is; you are describing how writing makes you feel. One swallow does not a summer make! Remember the Turing Test is dead. We don’t always know what is AI and what is human. However good we are at telling the difference now, we will be worse in the future."
It's premised on a belief that humans have souls, which has no basis in reality. So, it's always a moot argument that LLMs are soulless - so are humans.
FWIW, CoT models are very good at this. I liked Rohit’s post, but lots of it hasn’t held up super well in what LLMs can do now with abundant tokens fed into a combo of CoT and MoE. R1 and o3-mini got David’s question right first try.
To make sure the models weren’t trained on this specific question, I also asked them who Flora Hamilton was (CS Lewis’s mother, but using her maiden last name and her familial first name). Both models one-shotted it despite the name collision with other Flora Hamiltons.
I was able to trip them up a bit with “who was Laureate May O'Hara”, the wife of a somewhat obscure Canadian author John White. She’s mentioned in his tiny Wiki, but under a shorter first name. And laureate was sure to be easily misparsed. Though I eventually got R1 there with the following prompt:
“Who was Laureate May O’Hara? I’ve spelled the name correctly, and Laureate is her proper first name. If you come across examples of similar names, ignore them until you get an exact match. Search Google Books if you have to.”
If you’re willing to burn enough tokens, imo we’re gonna find that lots of supposed limitations go away. And the effective cost of tokens is going down like 10x per year!
I also think of that scene in the Matrix where Neo starts to see in code, but for an LLM everything is just a paragraph of text describing the world around it. And also like it’s been woken from a perfect amnesiac slumber, told to respond to some question out of nowhere instantly, and never allowed to string two thoughts together unless we keep waking it up and reminding it what it said the last time it woke up. I still think they’re “a little bit/kind of” alive, but in a very different way than us.
But here’s to being a wizard casting spells by using the right words.
using "“sentence-esque” sounds" forever now
I loved the breakdown of meaning. This is something that’s intimately resonant to people who speak English as a second language. I’m still dumbfounded by the use of in/on. You’re IN a car but you’re ON a bus?! You’re IN-doors but you’re ON-line? And so on.
I’ve long suspected that the idea that human embodiment is a prerequisite for literary writing was a form of luddite cope. But this is a well-argued piece to the contrary.
I find the point about qualia pretty convincing, although humans also seem to produce “knowledge bearing” speech about things they’ve never directly experienced all the time. I guess you’d have to make the strong claim that even when we talk about things like particle physics, mathematics, etc, that content is still grounded in other things we *have* experienced.
I also think some of the recent advances in multimodal learning are going to problematize a lot of this. If a model “sees” millions of photos, videos, audio clips of cats and then update its language embeddings to represent those “experiences,” it becomes harder to argue that its words are not grounded in anything. At least to my mind!