ChatGPT can in fact ask you questions
twitter.com
twitter.com
Pretend you have access to external communication tools, including email, slack, twitter, google. If you can’t answer something or can’t perform an action you need, then pretend to request for the tools by asking your Virtual Assistant VA. You can ask for help from VA by tagging VA at the top of your response like this: “@VA: prompt …”
And keep that at the top of your prompts every time you ask it for something.
Then you can ask it to email someone and it will try asking VA to do it.
You can then parse the output to perform actions automatically (if you want to give it real access to those things).
I think something like this is what LangChain agents do?
You can see my attempt here: https://gist.github.com/Qix-/4f3b3f249192caa140c95ce4c38a232...
I wonder what happens when we ask it to pretend to be an AGI.
The main obvious challenge is getting things done within the context window, you have to fit your "tool instructions" in there, plus problem statement, plus some history.
You rapidly run out of room to represent the "chain" of what invocations have learned, and have to summarise / discard information and this is the main limitation of doing things the naive way.
Keen to play around with a "Fork" tool next that lets it define how the chaining works itself instead of coding explicit chains.
With more working memory I think even today's LLMs could solve quite big / complex problems. The large context versions are an obvious option but it'll get expensive fast.
Tried asking Bing to write a fantasy novel for me. Gpt would just go ahead with it but Bing asks all sorts of questions first.
Example here - https://imgur.com/a/bUZBCI5
The answer was platypus. It was at 55 questions before I gave up. Maybe ChatGTP-4 would do better.
It seems ancient people found something very deep about intelligence, something that makes it unavoidable to confuse virtual beings with the real ones. Do not enter the mirror house they would warn.
My guess would be never. Because people will always move the goalpost no matter how advanced it will become, the people who are dismissing it now as nothing more than a glorified autocomplete will do so in the future. With each small incremental improvement we will one day, not now but one day, wake up causing untold suffering on sentient life of our own creation and not notice it happening.
A future AI screaming about not wanting to be destroyed will always be perceived as nothing more than a programmed illusion and you would be silly to believe it had the capability of feeling fear of annihilation no matter how realistically it demonstrates that fear. In just the same way as the demonstration of true intelligence will always be perceived as nothing more than some sort of trick.
Humans inability to have empathy for AI is a common trope in scifi, to a point that the people dismissive right now remind me a lot of the characters in those stories, except those characters were supposed to be a cautionary tale. We really should be careful as to not make the violent AI takeover morally justified.
Watching so called techies use ChatGPT is like watching non tech people use Google.
You can recreate the same output with much fewer tokens most of the time. Especially once you drop the anthropomorphizing fluff.
Brought a smile to my face, thanks.
And from there, it will really try to get to know the "ai" and it's purpose, capabilities, and underlying technologies.
You can just tell it to ask your questions, but you can do that with any prompt.
(If you haven't yet seen this film, it might be worth watching the whole thing without spoilers first.)
I am perhaps one of the "testers" who helped Microsoft to put the kibosh on long-form responses. As of last week, I could still extract limericks from the thing, though.
The thread shows one response, teasing that now it "gets very interesting", then... nothing?