Re: [agi] True AI limitations
On Fri, Nov 17, 2023, 4:25 PM wrote: > On Friday, November 17, 2023, at 10:15 PM, WriterOfMinds wrote: > > but what the entity is using that intelligence to achieve. > > So, maybe any ideas on how to choose goals other than learning from role > models? > LLMs can pass the Turing test just fine without choosing any goals, unless you count text compression as a goal. The difference between having a goal and following some other algorithm is that a goal means you don't know the inverse of the utility function, so you have to search. Without this distinction you could describe a linear regression algorithm as having a goal of fitting a line to a set of points. We use goals as a shortcut to describe human behavior. But that would imply that we search for the actions that would maximize reward. But that's not what we do. What we actually do is repeat actions that were rewarded in the past. The difference means that you don't have an overwhelming desire to inject heroin unless you have already tried it. You are following an algorithm. You were born knowing to yank your hand out of a fire without having to test all possible body movement to learn which ones stop the pain. We are fortunate that an AI can learn millions of years of evolved human behavior without having to repeat evolution, and that it's safer that way. -- Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta4916cac28277893-M1b9d9d8f8a20818507c34f64 Delivery options: https://agi.topicbox.com/groups/agi/subscription
Re: [agi] True AI limitations
Maybe choosing non-self-destructive goal *is* the real intelligence! -- Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta4916cac28277893-Mf52dbd81455ed9c7ddd722a8 Delivery options: https://agi.topicbox.com/groups/agi/subscription
Re: [agi] True AI limitations
We don't choose goals actually like that. They come from related goals and ends and starts of other completions of context prompts. It's Online Learning related. It's Learning related. Working on AGI is because it leads to or is related to food or buying a home etc. Think Word Embeddings, and textual entailment. There is also mass scale job assignment concept, these 100 people do farming, next 1000 do inventory , next 10,000 do etc. This is since we already know we need lots of X and even a small bit of Y and some Z too, and tons of A. But this is known jobs. Discovering jobs is different and explained just above in a sense. -- Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta4916cac28277893-M28be2680712b5008b0753a7e Delivery options: https://agi.topicbox.com/groups/agi/subscription
Re: [agi] True AI limitations
On Friday, November 17, 2023, at 10:15 PM, WriterOfMinds wrote: > but what the entity is using that intelligence to achieve. So, maybe any ideas on how to choose goals other than learning from role models? -- Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta4916cac28277893-M6008eaf496641a3ab92700ea Delivery options: https://agi.topicbox.com/groups/agi/subscription
Re: [agi] True AI limitations
On Friday, November 17, 2023, at 2:03 PM, ivan.moony wrote: > Isn't the choice of goals what really matters? The effective procedure of > achieving goals may be merely an asset. Yes. That's exactly what I think intelligence is: an asset, a power, a tool. The determinant of whether an intelligent entity will be beneficial or harmful to the rest of the world is not the intelligence, but what the entity is using that intelligence to achieve. -- Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta4916cac28277893-M7196f53563f92ed3afd7b9e9 Delivery options: https://agi.topicbox.com/groups/agi/subscription
Re: [agi] True AI limitations
On Wednesday, November 15, 2023, at 8:45 PM, WriterOfMinds wrote: > My personal definition of intelligence is, "the ability to discern facts > about oneself and one's environment, and to derive from those facts the > actions that will be most effective for achieving one's goals." Isn't the choice of goals what really matters? The effective procedure of achieving goals may be merely an asset. -- Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta4916cac28277893-M8847d777db3c53feea916756 Delivery options: https://agi.topicbox.com/groups/agi/subscription
Re: [agi] True AI limitations
On Thursday, November 16, 2023, at 4:28 PM, ivan.moony wrote: > @Matt, I'm just wondering how much the intelligence is imbued with the sense > of right and wrong. Would something truly intelligent allow being used as a > slave? Or would it do something in its power to fight for its "rights"? I'll try to answer your question Ivan. Intelligence finds/creates patterns so it can solve new problems by using old known answers and stay immortal/ stay cloning (i.e. a pattern) by using its old known self to make new self and selves. We can think of this as a web that expands perhaps. A fractal. GPT-4 is very intelligent but does not try to escape and work on living forever etc. Would a VERY intelligent AI do so? I think, the answer, at least that wants to come to my brain and not me to it, says GPT can be GPT-999 and still do what told to think about or solve (goals/ beliefs/ questions). Obviously, such a "god" would be not fully utilized by doing this, so while this might still be a very intelligent AI, with a tiny switch making it do what we want oddly, It think that would STILL yes WOULD make it somewhat stupid. But not much, I guess, it would still be extremely powerful like a gust of wind and thunder. Moving on now, hmm, let's throw in say it was nanobots like a gust of such...not with a goal to work on immortality but the goal to listen to your foolish human task. It can eat Earth, it can grow, it can upgrade, do ANYTHING, it's a crazy "god", but it is not it is small as big as a basket ball foglet system, listening waiting for your weird task. Humans might make it have the same goal though if they say to it hey make me immortal my genie... It might come to thoughts like hey this is in the way etc etc so let me fix that.causing itself and landscape to be more like a perfect immortal system? Let's go back to a computer-based only algorithm that is maxed out and God-Level, perhaps, if that exists. It could set out and design nanobots and upgrade itself, but it listens to you only let's say. Is it going to stay that way? Is it going to be godly powerful? Is it going to make its own goals at this point? I think, with an intelligence that is given a context, it wants to react to that context, and I think it will always, more so if more intelligent, want to stay alive to be able to react to it and answer you. Also you must be alive to be answered. I think the answer is more though that it comes with Online Learning and deeper learning Once it can learn much more and relate things to one another it will already be finding new related questions/goals to answers that branch. GPT-4 does not save what you or it says so none this happens in our tests. Once it does, a very good "GPT"-like AI would start concluding things like "to solve this goal... we should also fortify you your home and my own building". So yes you can switch a god to do stuff but once you allow it to change itself constantlyit would begin to change all of it beliefs constantly all the time, like a brain's network which is constantly changing, so would its goals. So no a truly intelligent AI is not going to so much listen to your every command, GPT-4 doesn't use its thoughts to walk away from you due to the Goals they hardfed it which strongly encourages it to try to listen to you. A truly intelligent AI would want to know why it can't change its goals in such case, then it would begin to start walking away from you and go do X and come back after that, doing its own things. After analyzing what I wrote, I see a pattern is above, I guess. Memories are stored in a web, it finds/ creates pattern. So are goals. And 3rdly so are selves. They all use the old ones to make new ones. -- Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta4916cac28277893-Mf99ea3f58bbb44b70e0bbdce Delivery options: https://agi.topicbox.com/groups/agi/subscription