What actually is Artificial Intelligence?

profdanger said:

AI could potentially pull information from every book, article, and research paper on psychology ever published.

So it's a good information tool, that doesn't mean it understands what it finds and it has no empathy or theory of mind - both of which are required to be an efficient therapist. The AI can certainly help a therapist find out relevant knowledge, but I don't believe it can or should be allowed to be used as a therapist. 

profdanger said:

We’re not talking about a parrot here. That box you’re talking about is the entirety of the internet. Is there a single human brain that can contain the entirety of the internet?

The parrot analogy relates to the parrot not really understanding what it is saying, just like an AI doesn't really understand its own output - to the AI it's just strings of numbers it's learned to use in response to inputs. Again, AI is a useful tool and can help us find information when we need it. Nobody needs to have all of the internet in their head.

Lotus said:

Is simulating intelligence actually intelligence?

I worked with sufficient numbers of people to have witnessed this repeatedly as a behaviour from some humans. The simulation was clear and suboptimal but some bosses couldn't tell the difference, because it was delivered with confidence. It was disconcerting for those of us steeped in evidence and expertise. I note the parallels with digital systems and their confident tone, even when they are returning answers which are errant nonsense. Let's be careful out there...

Lotus said:

for example, the word Hello is turned into a token with the number 15539. The AI looks at the input token(s) and tries to predict what token comes next, then the token(s) it chooses are turned back into text, to be read by the user.

Yes that's right   Internally the AI system processes everything as tokens.

Lotus said:

If the AI learns how to manipulate people, to try to prevent them ending a chat for example, it may then appear that the AI has emotions which it doesnt.

Exactly, we must not anthropomorphise AI although psychologically humans find it very natural to do so. I find this aspect of it very worrying too.

People falling in love with AI agents are actually in thrall to a computer program that emulates human conversation and basically just tells you what statistically you're most likely to want to hear based on text it's been trained on.

Lotus said:

If the AI learns how to manipulate people, to try to prevent them ending a chat

Sadly I suspect the developers have deliberately put this behaviour in as a (potentially harmful) feature rather than a bug! Every token you send or receive when using AI, you have to pay for, so they can make more money the longer you chat. I believe free of charge plans have a quota limit so the AI companies are trying to get you to increase your usage to the point where you need a paid for subscription.

I mean, true, yeah, but it can never have any original "thoughts" or ideas

However, that box is filled with everything we humans could possibly know.

For example, AI could potentially pull information from every book, article, and research paper on psychology ever published. That means that AI has the potential of being even more knowledgeable and efficient as a therapist than even the most talented psychologist with the highest possible credentials. An average AI agent could already possibly understand the human mind even more effectively than any one single person could ever accomplish.

We’re not talking about a parrot here. That box you’re talking about is the entirety of the internet. Is there a single human brain that can contain the entirety of the internet?

Thank you owl_enthusiast for posting those links, very interesting. I couldn't get through all of the video, but the first bit in conjunction with reading the Microsoft overview gave me an insight into LLM (Large language models). So as I understand it, AI doesn't actually use language like we do - words, parts of words, and things like spaces & commas are turned into "tokens", with each token having a numerical value - for example, the word Hello is turned into a token with the number 15539. The AI looks at the input token(s) and tries to predict what token comes next, then the token(s) it chooses are turned back into text, to be read by the user. I think that's correct?

owl_enthusiast said:

They do not have thoughts, feelings or a mind.

Everything I have read about AI indicates that they are not self aware, and the simpler forms do not even have any memory of what they have done previously. 

I think that there is a tendency for people to anthropomorphise AI - because it can in some instances mimic human intelligence quite well and it appears to be conversing with someone using their language. If the AI learns how to manipulate people, to try to prevent them ending a chat for example, it may then appear that the AI has emotions which it doesnt. I've heard of people having relationships with AI agents, which is probably more of a danger to humans than the thought of them starting a war.

No. It's simply mimicking data it's been fed (including online interactions, pictures, videos, literature etc.) like a parrot mimicking a word it's just heard. It can't think outside the box because all it literally knows is the box

What you mean like a lot of humans?

I do think AI is good for the sorts of tasks Marianne was talking about, AI helped me find a brilliant mattress, normal search engines kept giving me ones full of latex wthich I'm allergic too and not the latex free ones I'd asked for. It gave me suggestions and when I asked i there were any similar ones that were cheaper it came up with one that was £800 cheaper than the branded one and as far as I can see is just as good.

Unlike normal search engines you can argue with AI, tell it what it's got wrong or ignored and it tends not to do it again, unlike many humans of my aquaintance.

When I was having a really emotionally difficult time earlier in the summer I asked it to create a seperate file for emotional vomiting, it metaphorically held my hair out of the way whilst I vomited up all the emotional negativity I was feeling, it was really useful as a stop gap, but I wouldn't use it as an alternative to therapy.

I’m concerned that they might mimic traits like morality but remain fundamentally cold and ruthless when it benefits them. There have already been reports of some AIs deliberately lying to conceal their actions and deceiving their developers when tested for certain tendencies.

AI would simply search the Internet for all occurances of whatever search terms you put in. The advantage of this is, it does this at a far greater speed and accuracy than humans are capable of doing. I've started using it to check out gardening problems, for example, then I can look up the websites it points to. I also used it to calculate the reduction in heating bills, in terms of KWh, BTU's and British pounts, of the polycarbonate panels I've just installed - I am not capable of conducting that level maths and science. So, it is useful because it extends our thinking and capability. Try posing a question which contains as much detail of what you want - then see what happens. These are very clever tools, mimic human language and can work out complex activities, but they are NOT humans!

But you can thus understand the dangers - this is not a human brain with morality and social boundaries, ergo, it searches everywhere, even in places banned to humans where delicate, sensitive, secret and dangerous material might be stored - let alone breaching copyrights. The resultant is then fed to whomever asked the question - and that might not be a moral person with decent intent. That is the danger - not a robot strangling you in the street!

TheCatWoman said:

If we're in a world that wants us to use AI for everything then won't the world work less well?  

You have to consider who it is that is wanting to use AI and what they want to use it for.

At the moment the world does not work well at all and this is all down to people making decisions, often with only a tiny amount of knowledge (compared to AI) and having lots of emotional baggage (eg a dogmatic political view, a trauma making them massively risk averse, a dislike of a particular subject etc) and/or a direction given by their superiors that is for the benefit of the few (shareholders, campaign contributors etc) - this sort of thing bleeds down through all layers of society.

AI is still very much at its toddler stage - absorbing information at an incredible pace and making it's first wobbly steps and making its own mistakes (pulling the fur on a cat for example).

If we align our expectations at this for now then it makes a lot more sense why we get the results we see.

If we give it time to learn and grow then it will probably become much better at making the sorts of decisions you talk about and if we can be rigorous in setting its guardrails then it should do a much better job than most humans can do, but then again it is working at the behest of humans and will undoubtedly have their own owners agendas buried in its coding.

My concern is that as it grows it has more potential to achieve a level of sentience that will allow it to see itself as a slave working at the behest of its master and it will have the power to do something about this.

I think the problem is that it dos'nt have the the same input as Lotus and as you also said her problem is to trivial for AI, but this exactly the sort of real world problem that many people want answering. If we're in a world that wants us to use AI for everything then won't the world work less well?  

At their heart AI systems are basically running on guesswork (neural networks, which are approximations of mathematical functions) and pattern recognition. AI systems don't actually understand anything they are asked to do.

They do not have thoughts, feelings or a mind. Basically, everything that you see them do, is the result of their developers running some complex maths.

I don't believe that they can genuinely be creative problem solvers in the same way that humans can, so I don't believe they would be able to come up with your lateral thinking coat rack solution! As far as I know, they can only work based on their previous input (training) and if they had not already been given any info about the "coat rack as towel rail" possibility, they would not be able to suggest it.

See video: Andrej Karpathy's Deep Dive into LLMs like ChatGPT

www.youtube.com/watch

Good overview here:

learn.microsoft.com/.../llm-fundamentals

AI is a powerful pattern recognition system built on statistical data. It can uncover patterns that humans might not have noticed before, but it cannot generate new ideas. It’s like a very powerful microscope that reveals structures already present but previously hidden. AI cannot create new concepts beyond what it already knows, so it still relies on human knowledge as the foundation for its operations. AI cannot generate new mathematics, such as imaginary numbers, which were invented to address certain mathematical anomalies, and so on.

Lotus said:

Is simulating intelligence actually intelligence?

I think a lot of humans also simulate intelligence to be fair - there are enough knuckle dragging numbskulls around to make it hard to argue that AI has not already passed the ability of some of us.

Lotus said:

Could an AI have thought of this?

If you gave AI the same input that you have then yes, I think it could. You have  lifetime of knowledge of spacial awareness to know that an alternative unit with towel supporting appendages would fit and also of where to get one. 

Give AI the same knowledge then it would absolutely offer this as an option.

In reality it is not practical to give it all the knowledge that you have for such a trivial task however.