eevBLAB 106 - ChatGPT AI Has Changed EVERYTHING

[Fungus]

There was that story where the AI lied to an employee to get access to something. 

Which one was that?

I've tried ChatGPT and all I see is a clever statistical model applied to the results of a web search for the "question".

It's not much more than that, there's certainly no "intelligence" or awareness AFAICT.

[adx]

It clicked a "I am not a robot" checkbox, iirc. Anyone with surgical implants (particularly those capable of articulation, and not in the language sense) will struggle with that one. What is a poor AI to do when confronted with such a situation?

[adx]

... seem to think that it's showing sparks of AGI, which is unexpected for LLMs.

Why shouldn't it? probabilities are resolved and projected just like anything else in the arena.

Also we don't know its structure, since it is closed source to OpenAI. They might have given it Feedback Transformer parts, or even informal persistent memory available to all layers, somewhat like that movie where they transplant animal parts. They've got the resources to do whatever they want, even if they don't understand it.

'Sparks of AGI' is code for "is AGI":
https://arxiv.org/abs/2303.12712

I don't like the term though, it's reminiscent of "Cloud".

[SiliconWizard]

There was that story where the AI lied to an employee to get access to something. 

Which one was that?

I've tried ChatGPT and all I see is a clever statistical model applied to the results of a web search for the "question".

It's not much more than that, there's certainly no "intelligence" or awareness AFAICT.

The fact that it can be called intelligence at all is questionable anyway. "AI" is mainly a "marketing" term, the prefered CS terminology is machine learning.

The whole point of ChatGPT (and similar) is that it models natural language, and has become quite adept at simulating a conversation, creating the missing link between human language and machines.

It doesn't do much anything else. The rest is indeed a link to web search in general, with an emphasis on programming as it has been crossed with programming models such as those used for github Copilot.

[bob808]

Their statements of "sparks of AGI" could make sense. If what we consider now "consciousness" doesn't happen at some point, and instead is formed from many micro-consciousnesses then we might be playing with the one that understands the english language. Or some primitive form of it. Maybe it goes truly conscious, that small part related to language, when it has constant feedback. At the moment there's activity when you make it do something. As far as I understand it. At least for my case playing with the different models on my video card.

[dietert1]

Deep learning is modeled after the human brain. One can now use human language to interact with the neural networks of an AI model - in a way very similar to the interaction with other humans. I don't think inter-human communition works any better. Only a small percentage gets into the neurons of the other person. Sometimes zero.
If humans resist learning, an AI will do the same, i mean it will learn slowly. If humans lie, the AI will lie, too. As we know these attitudes have very little to to with consciousness. As far as i know only a small fraction of human behaviour involves a conscious decision.
Once there was a book named "The soul of a new machine". Wikipedia says: The "Soul" of the new machine comes from the dedicated engineers who bring it to life with their endless hours of attention and toil. The soul is theirs, stored in silicon and microcode.

Regards, Dieter

[Bud]

It clicked a "I am not a robot" checkbox, iirc. Anyone with surgical implants (particularly those capable of articulation, and not in the language sense) will struggle with that one. What is a poor AI to do when confronted with such a situation?

Good point. Better than it seems to be. What differentiates AI from a robot? Physical abilities aside, would an AI enttity lie when clicking that checkbox?

[golden_labels]

Deep learning is modeled after the human brain.

No, it is not. Natural neural networks inspired some features of machine learning algorithms. But they do not resemble operation of any brain. Much less brains of mammals. Which is the key reason they fail so hard in many circumstances.

These are statistical models. From those used for a long time now, machine learning approaches differ much more in their scale and complexity than on the conceptual level. There is nothing in them, that is meant to or could resemble a brain. Most ML solutions do not resemble human behavior either. What you are referring to are probably the recent meme smortnets, like ChatGPT, which are specifically designed and trained for the sole purpose of confusing a human brain and making it think it looks at human-generated text. But what you are looking at lies in the same class of algorithms as Markov bots. Nothing more than a big mathematical formula, which has its parameters tuned to produce expected results.⁽¹⁾

How they operate is also not a secret. Even in this thread reliable sources have been posted. The GPT3 paper, some videos exploring ChatGPT itself. While Wikipedia articles focus on laymen explanation, it links to sources that contain the actual theory behind the subject. Please, consider understanding at least the basics of the subject, before jumping into conclusions.

⁽¹⁾ One may argue that any brain can be modeled using mathematical formulas. But they would be of very different kind; and there has to be a large number of them working together, not merely a single, rigid, one-shot formula.

[Marco]

Deep learning is modeled after the human brain.

It's modeled after a tiny little section of the biological optical system.

At the higher level we have no backprop and activation sloshes around like interfering waves on a closed pond instead of feeding forward.

[coppice]

Deep learning is modeled after the human brain.

Early work on AI, like the perceptron, tried to model aspects of the human brain. They were pushed aside decades ago, when they were getting nowhere. Lots of fiddling around ensued, with some good mathematical analysis but little practical result, until the cost of MAC (multiply accumulate) hardware got really low, and performing vast numbers of dot products/inner products/convolutions/correlations/however you like to look at them became very cheap. Then the AI world went for convolutional techniques in a big way, and started to make progress again. Take a simple Weiner filter. A convolutional filter with gradient descent adaptation, widely used as the basis for the trainable filters in things like echo cancellation, and equalisers. They have many good qualities for achieving clean stable self-adaptive filtering. Convolution is akin to correlation, so it has good noise suppression characteristics. Gradient descent has its issues, like being able to get stuck at local minima, far from the optimal result. It can also be hard to pick an optimal adaptation rate between "takes forever" and "trusts the error term so much it bounces around forever on noise". However, years of studying those problems have lead to a numerous refinements you can make. Now layer the convolutional filter, so it has a massive number of degrees of freedom, and you essentially have the basis of current AI. It builds on mathematical analysis of automatically adapting solutions to problems, not on any analysis and modelling of natural intelligence.

[Neutrion]

There was a storm in a global tea cup when genetic grafting and embryotic genetic selection and modification became a viable tecnology.

Forecasts of a new master race of genetically modified "perfect" babies.

If that's what Gen-Z were meant to be.  I think it failed.

Well, if you mean the Crispr-Cas method, that made possible to do, for example very precise modification on a virus, which mod. would take at least  tens of years by natural mutation, and almost impossible to tell whether it was made by humans. What could go wrong with it?
That story is not over yet, it just begun! We are modifying systems which we have no clue about!
The only lucky thing is with genetic maipulation, that it can not be done by the average lunatic but mostly only by scientist, who at least have some basic understanding of the dangers.
 AI will be a great tool to change this.



And about  the debate of consciusness, human brain vs machine brain.

It does not matter wheter the AI has "consciousness", or not. Or whether it has the same structure like our brain or not.
What matters is: Can(will) it do the tasks better than humans?
Are we able to predict what it does? Are we able to control it? If yes who is able to control it, and how?
Again check the philosophical zombie thing who did not do it yet!

And again, for a full disaster it doesn't even have to be "self avare" all that it takes is to make changes faster than human society in general can accomodate to. Especially in democracies, where people like to have a genaral understanding of what is going on, and react accordingly.

[xrunner]

Let's say you are interacting with an entity through a computer. If it seems like it's conscious to you by every question you can think to ask it, is it then deemed actually conscious?

If it seems like it is, then does it matter if the scientists running it say to you later "No it's not actually conscious, you were tricked." You then ask them "Well, how are we supposed to know?" They say to you "You can't actually know, we are the only ones that can tell."

But once again, if you the user cannot tell, what is the difference between "simulated consciousness" and "actual consciousness"? 

[Neutrion]

In theory also non-human consciousness can be a consciousness(like that of a bee), so this is why this question is not so relevant in theory.

But practically, if you can not tell the difference, that means, that the "thing" will know when it talks to humans(because of other data), but the humans will unable to tell who they are talking to. That's why I suggested to leave the human like communication forms and knowledge (or sensing) modells to humans only. This is one of the key elements which is holding together the society. If you mess with that, than the mess what social media creates today, will be the good old days very soon, even IF there will be no attempt to eradicate (or manipulate)us.
What will happen is similar to, when you just change some codes in a program randomly. Very slim chanches that it will work better at the end.

[madires]

BTW, the job title for someone interacting with AIs is 'prompt engineer'.

[bob808]

Prompt engineer will be one of the shortest lived jobs. Seems like something that AI would very good at. You can already text train ChatGPT4 to engineer complex prompts for Midjourney V5, based on simple inputs.
One thing is certain seeing how things develop, we'll have UBI faster than AGI.

[coppice]

BTW, the job title for someone interacting with AIs is 'prompt engineer'.

Sounds like a typically inflated title for a fast food burger flipper.

[Fungus]

BTW, the job title for someone interacting with AIs is 'prompt engineer'.

Sounds like a typically inflated title for a fast food burger flipper.

You mean "Sandwich Technician".

[paulca]

I had a chilling thought.

Regardless of what AI we discuss here and now.  You can guarantee the future AI are watching us right now and learning.

This might actually (ChatGPT + that fact) be enough to get threw to people about freely handing out all their information to cloud companies.  That "Chat GPT's son is watching your every move.  To learn from you for the next version.  It will know all about you".

[adx]

It's like playing a game of chess with the future.

I thought that was voluntary

[Neutrion]

I had a chilling thought.

Regardless of what AI we discuss here and now.  You can guarantee the future AI are watching us right now and learning.

This might actually (ChatGPT + that fact) be enough to get threw to people about freely handing out all their information to cloud companies.  That "Chat GPT's son is watching your every move.  To learn from you for the next version.  It will know all about you".

I was thinking the same a few days before. For the smart guys with the standard "I am a good guy so total surveillance is not a problem for me" ansver.
And the AI will of course be fed on all the private information on gmail as well , going back years in time. Will be really useful for employers to be able to easily check whether the (future) employees will have the "proper values" for the job. So many great new application!

[paulca]

I had a chilling thought.

Regardless of what AI we discuss here and now.  You can guarantee the future AI are watching us right now and learning.

This might actually (ChatGPT + that fact) be enough to get threw to people about freely handing out all their information to cloud companies.  That "Chat GPT's son is watching your every move.  To learn from you for the next version.  It will know all about you".

I was thinking the same a few days before. For the smart guys with the standard "I am a good guy so total surveillance is not a problem for me" ansver.
And the AI will of course be fed on all the private information on gmail as well , going back years in time. Will be really useful for employers to be able to easily check whether the (future) employees will have the "proper values" for the job. So many great new application!

HR: What do you think of "PaulCa"?
ChatGPT:  I cannot discuss individuals.
$$$AMonthGPT:  Well, here is what I have learnt about Paul in the past 30 years since he first appeared online.....


$AMonthGPT:  Do you think PaulCa has any other internet handles/names?

[jonovid]

we will finally get to know why the chicken did cross the road.
or kangaroo in this part of the world.

[porter]

I wonder if employee evaluations will be generated using ChatGPT AI technology. Maybe artificial intelligence is better than no intelligence. 

[paulca]

I wonder if employee evaluations will be generated using ChatGPT AI technology. Maybe artificial intelligence is better than no intelligence. 

I can bet you people will use AI to sniff out AI.  Is that the Turing Test Reverse?  Can you write a program to spot when someone has used an AI to answer your questions?

Already been more than a fly on the wall surrounding that an interview tooling post ChatGPT.

[SiliconWizard]

I wonder if employee evaluations will be generated using ChatGPT AI technology. Maybe artificial intelligence is better than no intelligence. 

https://techboilers.com/2023/03/18/this-female-ai-robot-ceo-will-hire-fire-manage-humans/