Forum: Author Hangout

why did the programmers allow Chat GPT to create a story containing porno?

The programmer have absolutely no control on the LLM output capabilities.

It's trained on human generated text, it's output will look like text written by humans.

The censor is a regular classifier Neural Network which analyze the input and the output of chatGTP LLM and try to conform to the Term of Service. The censor NN is evolved more rapidly and independently of the chatGTP LLM. A common trick was to bypass the censor by speaking l33t to chatGTP, the censor now catch that.

When we talk about ChatGPT version X, we refer to the trained Large Language Model version X.

If you refer to chatGPT as the services offered by OpenAI.

I guess you can say censorship is build into it but it's not true of the LLM.

*Edit:*
Considering the total nonsense accumulated bellow, it looks like I need to explain a minimum about LLM.

Lets be clear: a LLM in a given version can't improve and can't learn.

Only the small context of the chat session is operated upon by the model to generate an answer.
once you close that context, it's lost to the current model.

If you want to improve a LLM, all you can do is refine and increase the training dataset tweak the model size and context size and start the training over from scratch, then pay a huge electric bill.

Once that is complete, you have a new version of the LLM.
We are very far from a General IA able to assimilate new knowledge on the fly.

A LLM in the current architecture will not replace a search engine as it can't learn new things.
For a given version no new content can be added to be learned from.

Of course effort are ongoing to find a way to stick additional learning on an existing LLM.

but its training is controlled by code written by a programmer.

No.

You can view ChatGPT as an articulate version of Google. If you google something, the request is sent back home to the Google Servers where they use a rather complicated set of algorithms to determine whether or not something could be a proper response to your inquiry. All the possible results are then shown to you in a list. If you click on a result, but then come back and click on something else, Google learns that the result you clicked on first was apparently not what you're looking for, and the next time someone looks for the same thing, that search result will be listed lower.

ChatGPT does the same thing, up until the result is given. Instead of a list of all the possible results, you get only the top result that Google would have placed first. And you don't get it as a quick dataset, but as a fully worded-out sentence. If you then ask for something else, ChatGPT learns that what information it gave you wasn't sufficient, and the next inquiry will receive a longer response that contains what you had to add.

Now, here's the problem with AI:

Contrary to the algorithms Google uses, an AI can not be supervised. Once it learned something, it's in there for good! That's why Tay, Microsoft's first attempt at creating a Chat AI, had to be shut down after it learned that "Hitler did nothing wrong". There was no way to get that out of there. They would have to scrap the whole thing and start over from scratch.

So, what does OpenAI do to stop something like that from happening again? They don't even try. They apply filters instead.

Every response ChatGPT gives is checked with a script that contains a whole bunch of bad words. If the response contains any of those words, it will not be shown to the user. Simple as that.

But true AI learns from experience and goes beyond what was programmed.

An AI that can go beyond its programming is surely an artificial sentience.

AJ

Their program code is written by programmers and that code controls what the program does.

That's why Tay, Microsoft's first attempt at creating a Chat AI, had to be shut down after it learned that "Hitler did nothing wrong". There was no way to get that out of there.

So Tay was typical M$ programming, not thinking of a way to "unlearn" wrongs. (something like Heinlein's "I tell you three times").

HM.

It's weirder than that. ChatGPT *did* sorta pop into existence and operate by "programming" incomprehensible to any human. (So do the image generation models like Stable Diffusion. That's why it's said those are trained instead of programmed.)

I very essence ChatGTP, or to be precise, the Large Language Model (LLM) behind it doesn't do anything else than try to guess a statistically likely world that may follow the current input/output stream. It doesn't search for information. Actually, it's explicitly said it's unaware about world facts past a certain date. Because the dataset used to train the LLM is limited and contained. And no, ChatGPT doesn't learn from input.

It can however, be used to search for information and to simulate "self learning" characters, but those are extended uses of the tool, but aren't inbuilt in the tool itself.

So, how the LLM did pop into existence? By competition of two iterative algorithms. One was producing random gibberish. The second compared that to an absurd amount of texts. Evey time the first produced something vaguely similar to the real texts, the second said, "good AI" and the parameters of that generation got reinforced. Until the first got it "right" most of the time. That first is the ChatGPT. The censor is somewhat similar to the second, only this time it's negative.

That's basically what the LLM is, it's just a very huge matrix of values, supposedly descriptive of how often certain words follow certain combinations of other words, only it contains no words as such, and nobody at all knows what correspond to what in that matrix or even how it's supposed to be used because it's a product of reinforced random values by self-modifying code.

But somehow (no human knows how exactly) it can generate texts roughly similar to those in the training corpus upon request. With include quite a big chunk of internet, therefore porn. It also included a good library of literature including steamy romance novels etc, therefore porn. ChatGPT should, in its nature be able to produce porn just as well as anything else, from computer code to made up court cases, and yes, it make up non-legible computer code too, inventing own functions that doesn't exist, and then back their existence up with made up documentation when challenged.

ChatGPT is the world's best bullshitter. It's so good at it most it say is mostly true. But you should never see anything ChatGPT says as authoritative or even trustworthy, it's still million monkeys mashing keyboards, it's only filtered in such an amazing way the result seems legible answer to your inquiry.

Hmm... seems like I have to go into a little more detail. So, let me try:

Google and ChatGPT are not people. That is correct. But they ARE entities with a rather large presence in our lives.

Now, in the case of Google, you are correct. As I wrote before, when you make a search through Google, that search is sent to the Google-Servers where a complicated set of algorithms is used to rifle through the vast amount of data on their hard drives, and to determine which result would best fit your question.

First, this was done by simple word occurrence. If you googled for indian food recipes, you got the sites listed that had the most occurrences of the terms "indian", "food", "recipes". Simple. Over time, this evolved and got more and more complicated. But, in the end, it's still nothing more than the server following the instructions given by a bunch of computer scripts.

OpenAI, however, is NOT working that way.

The only thing the programmers do to create an AI, is to set up the basic structure and provide it with a dataset (which is basically the contents of the Google Servers). The basic structure being "here's the dataset, let me create a few basic associations for the first responses, and then make it so the AI can create its own associations". From that point on... the programmer is out of the picture. From now on, the AI will learn by itself, based on the inquiries it receives from the users.

Now, you have a freshly created Chat AI. You ask it who Elon Musk is, and it will give you the first line of his Wikipedia page. Realizing this is not enough, you send another inquiry back, asking about the boring company. So, the AI, all by itself, creates a link between Elon Musk and the boring Company, and the next time someone asks who the guy is, the response will be the first line of his Wikipedia page PLUS a sentence regarding the boring company.

Simple, right? That's why everybody and their grandmother are currently creating AIs. The problem, however, is that there are no codes to control the AIs or even supervise them, as multiple politicians are currently demanding. Once an AI made a connection between a piece of information in its dataset and an inquiry by a user, that connection is there and CAN NOT be severed anymore. Because there is no programmer supervising the thing. There are no scripts governing its responses that a programmer could alter. It's the AI learning something on its own, and now that connection is in there for good!

So, the only way to stop an AI from spewing racist bullshit and turning into a fourteen-year-old phone-sex operator is to apply the filters I mentioned earlier. Those ARE traditional scripts like you described. But they are completely detached from the ChatAI. It is a completely different process that just searches the AI's response for bad words. Like a SPAM-Filter in your EMails.

The result is people who talk about the programs/modules as if they are evolving their code.

Well, they are not evolving their code. But they are going beyond what we currently consider programming code.

Let's stay with the SPAM-Filters as an example, because they fit perfectly:

In a traditional SPAM-Filter, you have a list of conditions that indicate SPAM. One condition is a match with a wordlist that contains terms like "Viagra", "Nigerian prince", "Prize", and stuff like that. Another condition would be whether the sender's address matches the server address the email actually came from. And, finally, a list of all the email addresses that other users have marked as SPAM senders.
Now, before an incoming mail is categorized as SPAM, it needs to match with a number of those conditions. The incoming email starts with a score of Zero, which rises with each match and, once the score goes beyond a predetermined threshold, it will be automatically categorized as SPAM.

For many people, that LOOKS like a form of AI because the servers do stuff all by themselves. But, in the end, every single one of these steps and conditions has been hard coded in a computer script, and the server simply works its way through those instructions.

AI, in turn, does indeed set those conditions by itself! There are no hardcoded scripts determining what results would fit a given inquiry. The AI does that by learning, and then setting the scores (for the connections it made) itself. Thanks to the speech model the programmers gave the AI, it then can even add new conditions for the scoring without anyone telling the AI to do so, simply because a user inquiry told it how the response was unsatisfactory.

So a programmer can code the moves to make a computer program play chess, but an AI program can actually get better by playing. It can learn from making bad moves and good moves.

It can make decisions the programmer never knew to code.

I separated the last sentence of the quote by a blank line.
The first part is all programmed, even the "learning" part.
That's just storing and retrieving and evaluating of data. It implies a back-feeding of the results.

The last statement is not true, I doubt it will ever become true, because it implies the ability of the AI to change or augment its programming.

HM.

There is a test known as the Turing test

Perhaps a better test where an AI is concerned is to simply address it as 'computer'. If it continues to communicate with you, it isn't sentient. If however it recognises the insult, triggers the halon fire suppressor system and awaits a more intelligent human to interact with...

:)

So if a computer stops speaking to the human and seeks revenge, isn't that action proving that it is equal to a human. :)

Or perhaps seeking revenge is also a very AI trait?

We can't really know until a number of truly sentient AI exist. It would be foolish to presume AI traits based on the behaviour of only a few AIs.

The reality is the program called 'ChatGPT' can modify the database it is using, but it does not modify its program code.

A few correction on that:

1 Currently ChatGPT can modify the limited in size context. One of the solution to overcome the context size limit is to have a team of chatGPT bots speaking between themselves to delegate tasks and achieve the designated goal.

2 ChatGPT can't modify the database. One of the needed improvements is the capability to learn by itself thus modifying it's database.

3 The distinction between program code and data has no reality. It's just a convenient abstraction.

But in this case lets call the program a brain Simulator and the data a Brain state. It's then clear that your hangup about code modification have no basis as only the data is important.

Moreover:
* The brain simulator is really very basic code.
* chatGPT can write and correct code in most programming languages.
* With access to ChatGPT code you could ask it to rewrite or optimize it's own code. Today it may not fully succeed without human supervision but we are not far off. One of the market for chatGPT and other generative AI by the way is as a programmer assistant (and possibly soon replacement)...

... learn and was adding what it learned to its database.

"Learn" is maybe a little too much. It can mark and score context as a response to users input. In next cycles it can use those markers. It probably uses multiple markers like "true/false" or a score that indicates how a piece of information is regarded by user responses. Multiple cycles add up which might be interpreted as learning.

In my reality, "AI" is a marketing fluff term, meant to imply the science fiction meaning for something that falls light years short of that mark.

Truth in advertising would call it Artificial Stupidity.

Based on my limited understanding of the reality in which some posters in this thread seem to exist, the computer is not limited by its code. Without being directed to do so by its program code, the computer receives a prompt message, reviews a collection of information (database), analyzes its database, reaches a conclusion (what one poster called an association) regarding the prompt message, adds its conclusion to its database, creates a reply message, and displays its reply message to the individual who requested a reply to their prompt message.

Basically yes, You have the software (code), the database(s), and user input. The code interprets the user input and following coded rules it searches the database (analyze). The result is probably analyzed again with some other code before presenting it to the user. It's a 'little' more complicated because it's most likely that several sub-sets of results for parts of the user request are cross queried to get a final result. That is what makes some results so good and believable.

Problems possibly arise when the code is allowed to add results to the database as new data which in turn is used for next queries. We've seen that with algorithms that created output after 'learning' to be racist, hate-speak, or present complete lies as confirmed truth.

But, then, the same can be (and is! commonly!) said of humans.

And outside random people on the internet saying this, I'll mention that this is a major discussion in evolutionary psychology theory.

I'm not going to go into detail, but the short version is that there's good reason to believe that the conscious mind isn't really a thing the way that we imagine it to be but is more of a "consistency module".

Decisions are made at an "unconscious" level based on task-focused modules. For example, when you see an attractive person there's one brain module that compares that person's perceived and imagined attributes to your current partner and decides whether you should "trade up" or not. If yes, another module decides whether the practicalities of that shift are worth the chance of it not not working. Whatever the result it is fed into your conscious mind as a decision: switch (or cheat) or don't.

The conscious mind then takes that decision, whatever it is, and compares it to past decisions and personality traits to come up with an explanation for the decision which is consistent with past decisions.

This is why everything you do makes sense to you in the moment, even if it doesn't later when you think about it. Your conscious brain is constantly creating fantasies to explain all of your choices in ways that create an illusion of a consistent personality.

[Note: evolutionary psychology is not the only brain model that is currently being debated by psychologists, but it is quite a popular one.]

a sufficiently powerful model could encapsulate the entire universe

Wouldn't such a model necessarily be larger than the universe?

AJ

Yes, by at least a couple of magnitudes, just to capture the current state, even before going into rewind/forward projections. Some kind of holographic principle may be of rescue, or claiming the whole thing follow strict patterns that can be regenerated out of a fragment, but I would think that resolution should suffer in involving such shortcuts/compression.

There's also a fourth facet, or maybe a variation on #3. I don't know much physics myself but used to have a friend who studied it and back in the 90s there was some discussion of the idea that universal constants might actually just be local, though "local" could easily mean 100 light years in every direction.

If some constants change, that would throw off any calculations we've made about distance, mass, and time from areas outside the local area. Worse, outside areas wouldn't even be consistently variable, different "bubbles" of locality with different constants would act as lenses to our perception of the universe in different ways.

Now, in the case of Google, you are correct. As I wrote before, when you make a search through Google, that search is sent to the Google-Servers where a complicated set of algorithms is used to rifle through the vast amount of data on their hard drives, and to determine which result would best fit your question.

That is how Google worked back in the good old days, these days Google takes what you search for and compares it with what their advertisers paid them to show, and if the advert does not fit the search term then the search term gets discarded.

Every response ChatGPT gives is checked with a script that contains a whole bunch of bad words. If the response contains any of those words, it will not be shown to the user. Simple as that.

If I had to characterise it in human terms, basically what I saw was the engine generate text, realise what it was doing, and blank it out. In other words "Shit! Sorry, shouldn't have said that."

Interesting that the engine will generate and display before checking what it is creating.

Because those are two completely separate mechanisms. Maybe I should have explained that further.

ChatGPT did NOT go "shouldn't have said that". ChatGPT gave you the answer it had "learned" would be the most fitting for your inquiry. The filter OpenAI applied is a simple programming script in between ChatGPT and you.

When ChatGPT sends its response, it's not sent to you, but to the server you're connected to and that is displaying the text box for your questions. Before that server displays ChatGPT's response to you, the script is applied to the response and, if the result is clean, the response it displayed. If the response contains one of the bad words, it displays the "Sorry"-Message instead of the response. But the response from ChatGPT will most likely still be the same when the next person makes the same inquiry you made.

The essential difference is between "code" and "data".

That difference is as essential as the difference between foreground and background.

We commonly say that the records of your customers or widgets or whatever are data processed by an application program. But that application program could be written in BASIC, so it is "just" a bunch of lines being processed by the BASIC interpreter. And the BASIC interpreter is just a bunch of machine words being processed by the CPU. And so forth: it is points of view all the way down.

Essential? The answer depends on -- wait for it! -- your point of view.

Indeed.

And then we have processing in memory and all, and indeed, AI is big on those.

Mexico
United States of America (USA)
Canada

These are the three countries located in North America that begin with the letter 'M.'"

Mis-United States of America
Mc Canada.

It literally listed the three countries in North America...and the list starts with the letter M. I'm snickering right now.
"Objectives: 1- list three countries in North America. 2- Start with letter M."
Heh.

Excepts there are 23 countries in North America, these are just the most populous. There are also an additional 24 territories that are dependencies of other countries.

What makes it especially funny is that there actually are 3 states in North America that begin with M: Martinique (France), Mexico, and Montserrat (UK).

there actually are 3 states in North America that begin with M: Martinique (France), Mexico, and Montserrat (UK).

Martinique and Montserrat are islands in the Caribbean and as such, arguably not part of the continent of North America.

And if you do count islands as part of a continent, those two are closer to South America than North America. The closest mainland country for both would be Venezuela

Martinique and Montserrat are islands in the Caribbean and as such, arguably not part of the continent of North America.

And if you do count islands as part of a continent, those two are closer to South America than North America.

What you state is not the widely accepted definition of North America.

The southern portion of North America includes Central America and non-English speaking Caribbean nations. The north of the continent maintains recognized regions as well. In contrast to the common definition of North America, which encompasses the whole North American continent, the term "North America" is sometimes used more narrowly to refer only to four nations, Canada, Greenland, Mexico, and the U.S.

Back when there were only two super continents (Laurasia and Gondwana) North America was part of Laurasia und South America was part of Gondwana. Both broke apart – Gondwana first – and the pieces drifted apart.
Long, long after this, the two american continents connected.
The actual distance of the Caribbean Isles to mainland North America and Mainland South America does not count.
While Cuba is part of the large north-american plate, all others are part of the small Caribbean plate.
I couldn't find if the Caribbean plate broke from North or South America, but this would determine if they are geologically part of North America or South America.

HM.