My Thoughts on AI Agents and the Price of Fish

ShelLuser said:
<snip> So my gf didn't like the idea of 'my' Copilot AI "just" accessing her data (perfectly reasonable to me). Well, first thing I've noticed is that I can't casually tell Copilot to "just look things up", I need to be very specific. For example: "Did I get an email?", then it'll tell me: "I dunno...". Only when I'm specific does that work: "Check my emails to see if I got a new one". IMO that's a reassuring idea. <snip>
Click to expand...
A ping-pong ball was better!

To design a certain type of hospital wings you need to make sure air from the patient room(s) is not entering the hallway (or visa-versa). To achieve this, you would measure the delta air pressure from inside to the outside of the room. This device is typically complicated and expensive--and also subject to breakage-. But imagine a clear tube installed in the wall of each room with a orange ping-pong ball in it. Now a person--with just one quick glance can look down the hallway to see if they see any orange. As a bonus, the simple system also results in an easy-to-teach system: Orange==Bad!

So, I'm not sure I'd share your GF's aversion to having something--or just yet another thing--access my data but I would certainly be looking at you with a side-eyed look.
 
AlfredoLlaquet said:
Crumb 2217: Non-Intelligence:
Click to expand...
I like that one. And that crumb reminded me of a quote... Found it. Sorry if my thought/comparison was a bit too off tract but I thought the crumb was good.

"I give myself credit for having seen clearly in a number of important situations. In itself, this is not so difficult to achieve, and yet it is rather unusual. To my mind, it is less a question of an exalted or shrewd intelligence, than of good sense, goodwill, and a certain sort of courage to enable one to rise above both the pressures of one's environment and the natural inclination to close one's eyes to facts, a temptation that arises from our immediate interests and from the fear which problems inspire in us. A French essayist has said: 'What is terrible when you seek the truth, is that you find it.' You find it, and then you are no longer free to follow the biases of your personal circle, or to accept fashionable cliches."
~Victor Serge
 
AlfredoLlaquet said:
If the confirmation is controlled not by the agent, but by the email client, you are right that it should be enough of a guarantee in that respect. But an agent can still "decide," for example, to email all your contacts. There have been cases. That's an example of what can happen. They are very capable, so suddenly they can start doing things you didn't predict because it's very difficult to predict every little possibility and say: don't do this, or this, or this, or this.

This crumb explains a real thing that happened to me, related to all this:

Crumb 2014: The Rogue Agent: In a project whose instructions require to consult the file Crumbs.txt, I told ChatGPT to provide a detailed analysis of the crumbs from several standpoints. Accidentally, the file Crumbs.txt wasn’t present because I had deleted it and forgotten to update the new version. The result was that ChatGPT (in full agentic mode) decided that the mission was the mission and, instead of reporting the missing file, roamed the internet for almost 4 minutes and consulted every crumb it could find on YouTube, DeviantArt, Instagram, etc., even attempting to download a ZIP file with an old version from the FreeBSD Forums. In the end, it generated the analysis. It was fantastic. Tom Cruise would be proud.
Click to expand...
ChatGPT especially but all those models are created to please, even at the expense of lying or desperate pointless scavenger hunts on the Internet .
Reminds me of people....they are trained on.

It is still evolving. The guard-rails are not very firm. However until it plants keylogger to sniff out your root password, it can be tamed. As I said, in the meantime I do not let it to roam free in my sensitive domain i.e. contact lists.
But even if it did, I am pretty sure I never took a d1ck pic to be send around. But why risk, it will auto-generate one.
 
AlfredoLlaquet said:
What worries me is that I don't think I'll ever be able to trust an AI agent. The possibility of one performing actions on your behalf that you would never approve will always exist and, if that happens, you'll have no one to blame other than yourself, but when everyone uses AI agents for everything, not using them will leave you at a great disadvantage, so probably one day I'll have to accept the risk and be one with the flock just to keep up.
Click to expand...
500 years ago, there is limited number of books. People have little or no information, myths and legends are spreading. Now we have abundance of information and statements generated by AI. The final result becomes similar to the lack of information in middle ages and even worse - now you read "facts" with detailed explanations and evidences which can be an outright lie.
 
You must log in or register to reply here.
Back
Top