Actually, let me say a few more things. I think the issue with Microsoft's Copilot (basically ChatGPT-3 and now 4, with custom tweaks), or just generally hooking ChatGPT up to the internet is what you might call a "shrinking value" problem. How is giving summaries of web pages from a search engine result a "revolutionary" technology? It's more impressive, at least in my view, interacting with it off the Internet, like talking to some too polite co-worker, who seems to know everything about everything. So if the business case here is augmenting search engines..... I don't get the "revolutionary" lingo. The business case for 365 sounds reasonable--if it works--, where the model can summarize meetings and draft initial Word docs and so on. We'll see how it works, and someone should pay attention to whether the office experience gets even more drone-like. I've been complaining of a growing conformity, this may be another step in that direction.
The world model point as you mention in your comment is extraordinarily important. At any given time, for any arbitrary query, you might get something wacky, or entirely convincing but just untrue. There are literally thousands of examples of this on the Internet, from making up historical events to insisting on illogical conclusions, on and on. So in effect, we need ANOTHER grounded world model to ensure that the ChatGPT results comport with truth. That's not revolutionary either.
All this said, I occasionally use ChatGPT to help remind me of things "what are the best books discussing the web circa 2000s?" and so on. Thanks again, Roberto, you got me thinking.
Ahhh, yes, touche. I should have known this as I occasionally use it with Bing. Didn't make the connection thanks for pointing it out. I asked Bing's Copilot to explain the problems with it browsing the internet, and got this:
"
According to my web search, ChatGPT can now browse the internet again after being disabled for months 1. However, there are some limitations to this feature. ChatGPT can access the internet through Microsoft’s Bing web browser, but it is only available to ChatGPT Plus and Enterprise subscribers 2. Additionally, users have reported that the feature is not working as expected and that they are receiving stock answers when they ask ChatGPT to look up a web page or do anything internet-related 34.
"
I don't know really know I have an angle on this yet, in terms of writing about it. But thank you for the factual correction. Always appreciated! Best, Erik
Apparently, chatGPT can now connect to the internet, so point 5 vwould no honger apply.
I don't think that undervalues the general point made here, and I agree that LLM don't seem to be a path forward for a artificial general intelligence, whatever that happens to be.
The lack of a world model is a far more important issue than the average person thinks, for it will mean that there's always gonna need to be someone who is knoledgeable on the matter to asses whether chatGPT's responses are correct or not.
On another level, I'm not entirely convinced LLM's are a particularly accurate model of how humans learn languages, but thats another issue
Actually, let me say a few more things. I think the issue with Microsoft's Copilot (basically ChatGPT-3 and now 4, with custom tweaks), or just generally hooking ChatGPT up to the internet is what you might call a "shrinking value" problem. How is giving summaries of web pages from a search engine result a "revolutionary" technology? It's more impressive, at least in my view, interacting with it off the Internet, like talking to some too polite co-worker, who seems to know everything about everything. So if the business case here is augmenting search engines..... I don't get the "revolutionary" lingo. The business case for 365 sounds reasonable--if it works--, where the model can summarize meetings and draft initial Word docs and so on. We'll see how it works, and someone should pay attention to whether the office experience gets even more drone-like. I've been complaining of a growing conformity, this may be another step in that direction.
The world model point as you mention in your comment is extraordinarily important. At any given time, for any arbitrary query, you might get something wacky, or entirely convincing but just untrue. There are literally thousands of examples of this on the Internet, from making up historical events to insisting on illogical conclusions, on and on. So in effect, we need ANOTHER grounded world model to ensure that the ChatGPT results comport with truth. That's not revolutionary either.
All this said, I occasionally use ChatGPT to help remind me of things "what are the best books discussing the web circa 2000s?" and so on. Thanks again, Roberto, you got me thinking.
Hi Roberto,
Ahhh, yes, touche. I should have known this as I occasionally use it with Bing. Didn't make the connection thanks for pointing it out. I asked Bing's Copilot to explain the problems with it browsing the internet, and got this:
"
According to my web search, ChatGPT can now browse the internet again after being disabled for months 1. However, there are some limitations to this feature. ChatGPT can access the internet through Microsoft’s Bing web browser, but it is only available to ChatGPT Plus and Enterprise subscribers 2. Additionally, users have reported that the feature is not working as expected and that they are receiving stock answers when they ask ChatGPT to look up a web page or do anything internet-related 34.
"
I don't know really know I have an angle on this yet, in terms of writing about it. But thank you for the factual correction. Always appreciated! Best, Erik
Apparently, chatGPT can now connect to the internet, so point 5 vwould no honger apply.
I don't think that undervalues the general point made here, and I agree that LLM don't seem to be a path forward for a artificial general intelligence, whatever that happens to be.
The lack of a world model is a far more important issue than the average person thinks, for it will mean that there's always gonna need to be someone who is knoledgeable on the matter to asses whether chatGPT's responses are correct or not.
On another level, I'm not entirely convinced LLM's are a particularly accurate model of how humans learn languages, but thats another issue