eevBLAB 106 - ChatGPT AI Has Changed EVERYTHING

[YurkshireLad]

How long before students are taught by AI? No-one will need to check if students cheated because the AI will know.

[dietert1]

The video you posted on physics questions is pretty much correct. The fact ChatGPT is able to communicate perfectly in various languages is impressive, yet does not mean its model of reality is correct. Many mistakes. I didn't try to discuss a bad result, maybe something similar happens as with Wikipedia: Who sorts out contradictory input? Majority vote?

Regards, Dieter

[bob808]

How long before students are taught by AI? No-one will need to check if students cheated because the AI will know.

 

Hopefully yet not controversial but let me ask you something more interesting: how long until the AI "identifies" with a human, and when do we start discriminating "it"? 
What I want to say is, would it matter if for all intents and purposes it behaves and has the results of a real human?

[Poe]

Ultimately, people will not accept AI telling the truth.

"truth"?...AI??   What!?

AI gives what it gets and filters based on the bias of its teachers.... as you have seen.

People have intentionally trained AI systems on dubious sources like 4CHAN.  You should see that "truth".

[bob808]

Isn't that how humans do it as well? If you have someone that only lurks on 4chan, what "truth" will you get out of them? Shit in/shit out applies to people and machines

[Bicurico]

ChatGPT is just an algorithm with a huge database and a trained neural network, which by itself is basically a statistical treatment.

There is no intelligence or self-awareness.

Also, all that ChatGPT and Dall-E produce is based on content created by humans.

Truth is a human concept that often depends on the point of view. What is true for one person might ne false for another one. What might be true today, could be debunked as false tomorrow, including science.

ChatGPT is not aware on what truth is. It calculates the weights of the input words and calculates the most probable first output word. Then it uses the input and output words to calculate which is the best next word. This is repeated until the generated text reaches the desired length. This is of course just a simplification. But all there is is a calculation that determines the next word in a sequence of words. No intelligence or self-awareness.

For me, the disruptiveness of this technology is that it accelerates and automates content creation. At the same time it allows to learn new things in less time, is used correctly.

[SL4P]

I think the greatest change will start in about 5-10 years…
The divide and types of differences between white and blue collar workers is going to explode. Everywhere.

Politics (as usual) will take another decade to catch up, but will probably be too late, with massive social upheaval and inequity with people from both sides that can’t adapt to the new reality.

What might be more interesting is the necessity of otherwise incompatible ideologies that MUST work together to reinstate peace and harmony in their communities.

[SiliconWizard]

I think the greatest change will start in about 5-10 years…
The divide and types of differences between white and blue collar workers is going to explode. Everywhere.

Well, it has already. But yes, in a sense, we are going backwards rather than forwards in terms of inequality, and a lot of what happened during the XXth century that has seen the rise of the "middle class" is rather rapidly getting canceled, with the middle class in many western countries collapsing at a never-seen rate compared to last century. The trend is already there, it didn't wait for all this AI frenzy. AI will not per se trigger it, it's already triggered. It will merely be a means to an end.

[Marco]

its model of reality is correct.

It doesn't have a model of reality, it just has high order statistics. It's amazing how much can be caught in just statistics, but generators have some fundamental limits. They can't apply procedure and to model non trivial systems you need procedural algorithms.

[dietert1]

ChatGPT pretends to have a model of reality and to know things and that this isn't as good as its communication capability. Maybe you did not understand what i wrote.
Anyway, there is some mathematics and some other faculties in the make. Maybe they can collaborate with Wikipedia to get things right.

Regards, Dieter

[adx]

The real problematic truth that will come from AI is that we humans don't have a "soul". Not more than it can have. Because what we call "soul" or "conscience" is just a side-effect of the multitude of things that are happening all at once in our brain. Lots of stuff deep down, lots of stuff on the surface. That creates "the conscience" as a side effect, it's something virtual that doesn't exist materially. 

You don't know any of that for a fact. But it is the "scientific" basis which much of the world follows if not believes as its fundamental truth. Leading to the conclusion that Blake Lemoine was unavoidably right. And the more interesting conclusion that a fully deterministic system can be conscious therefore possess both free will and a soul. So yes, problematic.

And it's not something that will be told to us by the (general) AI, it's something we will observe while making that (general) AI. That is a very hard to swallow pill, for most people.

That kind of already has happened.

[Neutrion]

ChatGPT is just an algorithm with a huge database and a trained neural network, which by itself is basically a statistical treatment.

There is no intelligence or self-awareness.

Also, all that ChatGPT and Dall-E produce is based on content created by humans.

Truth is a human concept that often depends on the point of view. What is true for one person might ne false for another one. What might be true today, could be debunked as false tomorrow, including science.

ChatGPT is not aware on what truth is. It calculates the weights of the input words and calculates the most probable first output word. Then it uses the input and output words to calculate which is the best next word. This is repeated until the generated text reaches the desired length. This is of course just a simplification. But all there is is a calculation that determines the next word in a sequence of words. No intelligence or self-awareness.

For me, the disruptiveness of this technology is that it accelerates and automates content creation. At the same time it allows to learn new things in less time, is used correctly.

You can not really define what is self aware. If we would suppose something is self aware because it makes unexpected things, than a wrong washing machine would fit the definition as well.
And even if something would tell you that it become self aware, how do you know it has?
And if it has but it didn't tell you?

But the thing is, as we are discussing it now, that even a not self aware thing can do a lot of harm. Just see the bubbling algorythm of FB. They are pretty easy to understand, but the effects on society was unpredictable. (Well actually not, but no one gave a shit I suppose.)

Other problem is, that the methods how these AIs are trained is completely out of any democratic oversight. The natural one become a nazi, so I suppose there is already some taming and modification about what it HAS to think or ansver. And there you have a censorship without any oversight, just like with FB, which is also not solved yet, the algorythms are still not openly aviable.
(How many people and politician would understand them is another question.)

In the last few years I repeatedly asked myself, how issues which affect less than one person out of thousands could become mainstream in the media and society.
While other affecting way more people are almost completely hidden. Did someone fiddle with some algorythm?
So oversiht would be crucial.

PS.:There will be soon an EU US high level political meeting about AI. Will it get at least one tenth of the media coverage like some BS recently?

[Neutrion]

How long before students are taught by AI? No-one will need to check if students cheated because the AI will know.

 

Hopefully yet not controversial but let me ask you something more interesting: how long until the AI "identifies" with a human, and when do we start discriminating "it"? 
What I want to say is, would it matter if for all intents and purposes it behaves and has the results of a real human?

Check the term "philosophical zombie".

I have to add that every time, when I identify myself as a woman and being in a public shower, I really feel other women discriminating and even hating me, just because I look like a man, and have a massive boner.
When we are there together with my friend the T-800 it is even worse, although "she" has a really nice smily face, and not leaking any oil. So yes humans are evil, but my friend mentioned a good AI based soulution for the problem.   

[bob808]

A behavioral zombie that is behaviorally indistinguishable from a human.

 
I think this fits best for a more advanced ChatGPT version.

[dietert1]

Question: How can we make ChatGPT join a discussion on an internet forum?

Answer by ChatGPT: ChatGPT can't join a discussion on an internet forum as it is a text-based language model developed by OpenAI, and it doesn't have the capability to participate in internet forums. However, you can integrate it into a chatbot application or a website that allows users to interact with it in a conversational manner, providing answers to their questions based on its training data.

Question: Are you a behavioral zombie?

Answer by ChatGPT: No, I am not a behavioral zombie. I am an AI language model created by OpenAI, designed to generate human-like text based on the input I receive. I don't have feelings, emotions, or consciousness, and my responses are generated based on patterns in the data I was trained on, not on personal experiences or decisions.

Regards Dieter

[bob808]

The current ChatGPT is clearly less than a behavioral zombie. 
This discussion reminded me of this terrifying though experiment: 
https://en.wikipedia.org/wiki/Roko's_basilisk

While the theory was initially dismissed as nothing but conjecture or speculation by many LessWrong users, LessWrong co-founder Eliezer Yudkowsky reported users who described symptoms such as nightmares and mental breakdowns upon reading the theory, due to its stipulation that knowing about the theory and its basilisk made one vulnerable to the basilisk itself.

 
So be warned, I guess.

[xrunner]

Been trying for a while to get into it but there's too many users. I think they created a monster.

[ataradov]

I think they created a monster.

I think they started scaling back the access in preparation to a paid tier roll-out. They collected all the test data they need, generated a lot of publicity, there is no reason to waste more money on this.

[bob808]

https://beta.elevenlabs.io/ 
"AI" for speech

[xrunner]

https://beta.elevenlabs.io/ 
"AI" for speech

I got a couple of tries before it wanted me to create an account, but yea nice voices. Just run the output of ChatGPT through it and you'd have a lot of fun. 

[bob808]

It directly asks me to make an account.
This could be used to generate audio books from PDFs/epub files.
Here's a sample for the tech:

[ataradov]

I just found a new AI-based entertainment. There is this Twitch channel https://www.twitch.tv/watchmeforever that translates an infinite AI generated version of Seinfeld. Dialogs and "video".

It has been going for days apparently. Jokes are "AI weird", but some are quite fun and it is genuinely watchable as a background.

Not sure if it just a huge pre-generated video though, it might eventually repeat.

[madires]

OpenAI releases tool to detect machine-written text (https://www.axios.com/2023/01/31/openai-chatgpt-detector-tool-machine-written-text).
Tool: https://platform.openai.com/ai-text-classifier

[xrunner]

OpenAI releases tool to detect machine-written text (https://www.axios.com/2023/01/31/openai-chatgpt-detector-tool-machine-written-text).
Tool: https://platform.openai.com/ai-text-classifier

Well -

The classifier isn't always accurate; it can mislabel both AI-generated and human-written text.
AI-generated text can be edited easily to evade the classifier.

https://platform.openai.com/ai-text-classifier

What if your teacher runs your essay through it and tells you it's an AI generated essay, but you say it isn't? What happens then? 

[bob808]

Interesting idea. I guess ideally you'd make ChatGPT somehow spike the generated text but that's kind of complicated, and people could figure it out eventually. This tool they released seems to work best on at least 1000 words texts.
I think the best way is to fight fire with fire, and have AI (eventually) test the students