In all the AI journalism I've been reading, this is the first piece in a long time that felt like it delivered on its headline in a smart and satisfying way. Great article
I still believe that LLMs are (basically) bullshit artists (in the Frankfurtian sense). LLMs simply can't tell truth from fiction, and don't use logic to proceed from premises to conclusions.
LLMs also have nothing that actually grounds their output in concrete reality; this also promotes fictional output.
So, yes, absolutely - ChatGPT and its ilk are best modeled by Lanley, a smooth-talking con artist who will tell you what you want to hear.
"Noteworthy: a new study comparing #ChatGPT vs doctors for responding to patient queries demonstrated clearcut superiority of #AI for improved quality and empathy of responses "
Presumably a future advance will be to train our models to spontaneously infer consequences from their 'beliefs' and to adjust their views when that process produces contradiction (or just internal tension). Current generation AI can't do that both because it doesn't have anything like beliefs or a process for spontaneously generating new consequences from it's beliefs but, while it may take some clever new approaches, no reason to believe that this can't be done. Once done AI may have something very much like our ability to generate strongly held countermajoritarian beliefs.
I haven't read this whole article, but ChatGTP produces better fiction (stories) than fact. Ex.
JS
what can you tell me about the San Antonio Dire Wolf?
ChatGPT
The San Antonio Dire Wolf, also known as Canis dirus guildayi, is an extinct subspecies of the Dire Wolf that lived during the Late Pleistocene epoch in North America, around 300,000 to 12,000 years ago. It was named after the location where the first fossils of this subspecies were found in San Antonio, Texas. The San Antonio Dire Wolf was larger than ... Snip ...
JS:
What is your source for fossils found in San Antonio?
ChatGPT
I apologize for the confusion in my previous response. After reviewing my information, it appears that there is no evidence to suggest that the San Antonio Dire Wolf was named after fossils found in San Antonio, Texas.
NOTE: I have had it actually invent papers and authors when citing these types of fictional entities.
I recently wrote about this and included many of the same points (the Vanderbilt email, for example). The best description I’ve heard so far, and which lines up with what you’re saying, is that “ChatGPT is an ideology machine.” Here’s my article if anyone’s curious:
Excellent and thought provoking. I teach in a business school and what I need after this are some exercises to help students learn what are the strengths and weaknesses of language models and how to use them better. Perhaps a set of prompts or guides for conversation with LLMs (maybe multiple for a comparison)
Maybe even have the prompts or contexts vary so students see when it is good to use LLMs.
"most of them involve trusting elites more and masses less. Maybe we can draw some lessons for real life, too."
Hmmmm... I've found a lot of your writing pretty good, until I read this last line. Far more of our current real life problems stem from trusting elites.
Back to AI/LLMs, if there were a way to prompt for the "elites'/experts'" answer separate from the masses' answer (and vice versa) that would be a tremendously valuable feature.
In all the AI journalism I've been reading, this is the first piece in a long time that felt like it delivered on its headline in a smart and satisfying way. Great article
I still believe that LLMs are (basically) bullshit artists (in the Frankfurtian sense). LLMs simply can't tell truth from fiction, and don't use logic to proceed from premises to conclusions.
LLMs also have nothing that actually grounds their output in concrete reality; this also promotes fictional output.
So, yes, absolutely - ChatGPT and its ilk are best modeled by Lanley, a smooth-talking con artist who will tell you what you want to hear.
Bullshit artists is correct. See my other reply below.
Eric Topol also tweeted about a new study:
https://twitter.com/EricTopol/status/1651965137006522369
"Noteworthy: a new study comparing #ChatGPT vs doctors for responding to patient queries demonstrated clearcut superiority of #AI for improved quality and empathy of responses "
The study is here at JAMA Internal Medicine:
https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2804309
Presumably a future advance will be to train our models to spontaneously infer consequences from their 'beliefs' and to adjust their views when that process produces contradiction (or just internal tension). Current generation AI can't do that both because it doesn't have anything like beliefs or a process for spontaneously generating new consequences from it's beliefs but, while it may take some clever new approaches, no reason to believe that this can't be done. Once done AI may have something very much like our ability to generate strongly held countermajoritarian beliefs.
I haven't read this whole article, but ChatGTP produces better fiction (stories) than fact. Ex.
JS
what can you tell me about the San Antonio Dire Wolf?
ChatGPT
The San Antonio Dire Wolf, also known as Canis dirus guildayi, is an extinct subspecies of the Dire Wolf that lived during the Late Pleistocene epoch in North America, around 300,000 to 12,000 years ago. It was named after the location where the first fossils of this subspecies were found in San Antonio, Texas. The San Antonio Dire Wolf was larger than ... Snip ...
JS:
What is your source for fossils found in San Antonio?
ChatGPT
I apologize for the confusion in my previous response. After reviewing my information, it appears that there is no evidence to suggest that the San Antonio Dire Wolf was named after fossils found in San Antonio, Texas.
NOTE: I have had it actually invent papers and authors when citing these types of fictional entities.
I recently wrote about this and included many of the same points (the Vanderbilt email, for example). The best description I’ve heard so far, and which lines up with what you’re saying, is that “ChatGPT is an ideology machine.” Here’s my article if anyone’s curious:
https://3quarksdaily.com/3quarksdaily/2023/05/oh-no-not-another-essay-on-chatgpt.html
Excellent and thought provoking. I teach in a business school and what I need after this are some exercises to help students learn what are the strengths and weaknesses of language models and how to use them better. Perhaps a set of prompts or guides for conversation with LLMs (maybe multiple for a comparison)
Maybe even have the prompts or contexts vary so students see when it is good to use LLMs.
"most of them involve trusting elites more and masses less. Maybe we can draw some lessons for real life, too."
Hmmmm... I've found a lot of your writing pretty good, until I read this last line. Far more of our current real life problems stem from trusting elites.
Back to AI/LLMs, if there were a way to prompt for the "elites'/experts'" answer separate from the masses' answer (and vice versa) that would be a tremendously valuable feature.
I’m hoping it winds up like the calculator. I don’t think calculators actually put people out of business.