In this week’s In-Ear Insights, Katie and Chris discuss generative AI, one of the three major branches of artificial intelligence. This includes tools like ChatGPT, Google Bard, and Microsoft Copilot. They start by defining artificial intelligence and the three big categories within it: regression, classification, and generation. Generative AI makes things and allows people to interact with artificial intelligence in a way that they don’t have to be an expert to do so. However, it can’t create something truly unique that has never been seen before, and it doesn’t do well with vagueness. The models are also being used unethically by creating misinformation, disinformation, and deep fakes at a massive scale to create the appearance of credibility. They advise to use generative AI ethically and be specific when using it.
[podcastsponsor]
Watch the video here:
Can’t see anything? Watch it on YouTube here.
Listen to the audio here:
- Need help with your company’s data and analytics? Let us know!
- Join our free Slack group for marketers interested in analytics!
Machine-Generated Transcript
What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.
Christopher Penn 0:00
In this week’s In-Ear Insights, let’s talk about generative AI.
It is one of the three major branches of artificial intelligence.
And it is the one that people can’t stop talking about.
This is tools like ChatGPT, and Google Bard and Microsoft copilot.
So Katie, where do you want to start with generative AI?
Katie Robbert 0:21
Well, I think a general definition would probably be helpful, I think there’s a couple of things that we need to set the stage for, like, first of all, what is and what isn’t artificial intelligence? And then more specifically, what is generative AI? And so, you know, at a at a very, very, very, very high level, artificial intelligence is math.
It’s math equations that are, you know, learning and running.
But I think that there’s a misunderstanding that, especially with generative AI, as you are interacting with systems like ChatGPT, there is this feeling that it’s becoming sentient and alive and learning, and it’s actually talking to you as a person.
But really, it is just, you know, learning based on the information that’s been given.
So can you give us the two definitions? What is artificial intelligence in layman’s terms? And then what is generative AI?
Christopher Penn 1:22
Sure.
So artificial intelligence is an umbrella term, that is the discipline of trying to get computers to perform intelligence tasks that human beings do.
For example, can you see can you hear, can you understand language with the things that we do with our organic neural network computers in between our ears, versus the mathematics required to get a computer to do the same.
So that’s artificial intelligence very, very broadly.
And a corpus is a number incorporates a number of fields like computer vision, audio engineering, regression analysis, all different types of machine learning, there’s so many terms under this blanket term, but artificial intelligence is the easiest way to encapsulate it.
Within artificial intelligence, there are three big categories.
There is regression, classification, and generations.
So regression, regression, AI regression types, machine learning, is basically find the thing, right, hey, I got this, I’ve got this outcome, I want you to figure out what relates to it.
Classification is organized the thing I got a whole bunch of data helped me organize the thing.
And genitive AI is, Hey, make the thing I want to make the thing.
And so these are sort of the three classes of artificial intelligence, and all three are important, they all have different use cases.
And you can’t, none of them will exist without the others.
Right.
So when you make the tools like a ChatGPT, it was made with techniques from classification and regression.
Katie Robbert 3:00
Which makes sense.
And so in terms of generative AI, so obviously, the big topic of conversation right now is a system called ChatGPT, which is the interface to the GPT large learning model.
And there’s been a few version of that versions of that.
Which is powered by a company called Open AI.
And the reason people are talking about this generative AI right now is because we as consumers, we as marketers, we as business people have yet to experience artificial intelligence with this kind of an interface with something that we can interact with.
And so there’s been versions of it over the years, like it’s not a new technology by any means.
So you think about chat bots, where you can program in types of responses based on a question that somebody is going to be asking.
So if it says, you know, what, what are your store hours, you’ve already given the Chatbot the information to say our store hours are nine to five, you know, six to whatever the thing is.
And so it’s a very basic version of generative AI.
You also have things like predictive text.
So it starts to learn what you are likely to respond.
And so my phone, for example, no longer autocorrect swear words, because it learns for me that yes, those are the words that I want to be using when talking with someone.
And so the model has learned don’t correct it to duck because that’s I’m generally not talking about them.
Christopher Penn 4:39
I write duck this ship all the time.
Katie Robbert 4:44
But the point being is that now we have more sophisticated interfaces that allow people who don’t understand how to program or fine tune large learning models to actually interact with This kind of artificial intelligence, this generative AI, in a way that they don’t have to be an expert to do so.
Christopher Penn 5:07
Right.
And all this started in 2017, with a papers, an academic paper called attention is all you need, which is the technical instructional, in underpinning of all this stuff.
Attention is all you need.
It was a paper that basically said, here’s a novel way to deal with the relationships between words.
Prior to that, and you’ve seen this, like you said, on AutoCorrect on your phone.
AutoCorrect can assemble probabilities have a word, but it doesn’t understand the context of that word, because it doesn’t have probabilities for anything other than like the last one or two words.
It’s a it’s a type of neural network called a long short term memory network or an LS TM, they are very computationally efficient, which is why your phone uses them because they take up much power, but they’re really dumb.
The Chatbots that you mentioned before, like rules based Chatbots, those don’t even use generative AI, those are literally just rule based Chatbots.
And that’s why people hate them.
Because like, it’s like, when you go into voicemail system, you just keep yelling operator.
Katie Robbert 6:11
I had that experience last week, and it was not fun.
Christopher Penn 6:15
Exactly.
So attention is all you need was a revolution in the way machines handle language.
Because what they are capable of doing is as they generate, they can validate the probabilities of the next word versus all the stuff that they’ve generated within a certain window.
And that window is dictated by how big the model is.
So you’ve likely heard people talking about how GPT-3 had 100 billion parameters, and GPT-4 has 500 billion parameters.
This is an easy way of thinking about that.
It says it dictates how much you can remember, right? So if you’ve ever used ChatGPT, and you’ve pasted in like this much text, it says I’m sorry, that’s too much.
It’s because the the interface is functioning, I can’t remember more than a certain amount of text.
Right.
So your prompt has to be detailed, but it can’t be like the entire encyclopedia Britannica.
The as these models have evolved, they are capable of remembering and relating more and more and more.
So in today’s version of GPT-4, the OpenAI model, it can remember in the coding interface 32,768 tokens, which is about 25,000 words, give or take.
So you could you could put in like my book, and say, Okay, write another book or translate this into emoji, right? You should not do.
But it’s capable of remembering and generating about that much.
We fully anticipate, you know, these things basically double each year.
So in a year to 18 months, GPT five, when it comes out, we’ll be able to do 64,000, right.
So Katie, you know that book that you’ve been wanting to write? If you gathered up all the individual pieces, and you wrote a prompt, that was basically a detailed outline, you will be able to feed it to that model and say, write me a book based on everything that I’m giving you.
Katie Robbert 8:14
I’ll put a pin in that one for now.
What about systems like Dali? And so you have these, I assume they’re also generative AI models, where you give it a prompt and it generates an image.
It’s, it sounds like it works very similar to a ChatGPT, where it’s generating content.
Image is just a different form of content in this context, is that a true statement?
Christopher Penn 8:45
No.
Excellent.
Yes.
They are mathematically.
Similar in that they’re, they’re trying to determine probabilities.
But the underlying architecture is completely different to model type called a diffuser.
diffusers work like this.
They do have language in them, right? You type in, you know, a dog on a skateboard and stuff like that.
And it references the trained images that had learned like, what does a dog look it was a skateboard look like what does it to to look like etc.
And then it creates an enormous amount of just white noise on the canvas.
And then iterates repeatedly until pixels start lining up that start to match the train images like Okay, so this part started to look like a dog’s head.
This part started off like a skateboard.
And it does this over a gazillion iterations until you arrive at a picture that matches the training data.
The combined training to say okay, that’s a dog wearing a tutu on a skateboard.
And that’s good enough.
That’s why I can create such crazy abstract art, right? Because it’s essentially just calculating the relationship of a pixel to the pixels around it into the pixels around it.
And so until it matches what it thinks it knows.
But it’s a very different underlying architecture.
That’s why the prompts for when you use a system like dolly looks so different than the prompts that you use with a ChatGPT.
Katie Robbert 10:12
But dolly is still a version of generative artificial intelligence
Christopher Penn 10:17
generative, yes.
Katie Robbert 10:18
Okay, it is genuine, it makes stuff.
And I think that that’s really the point is generative AI, makes things and so, you know, when we think about systems like ChatGPT, and Dolly, you know, people are having fun with them, you know, they’re like, write me a story about, you know, a walk I took earlier, but include aliens and spaghetti.
Or, you know, make me a picture to your point of a dog on a skateboard.
But there’s a lot of things that people aren’t using these systems for, like, build me a spreadsheet for the following things or, you know, rewrite this content in a way that is book or, you know, summarize these notes into action items, you know, things that are actually useful.
What are some of the things that generative AI should not be used? For?
Christopher Penn 11:08
What should you not use it for? anything unethical?
Katie Robbert 11:12
Well, sure.
Christopher Penn 11:16
But you say that, but that’s actually one of the prime use cases for it is to create misinformation, disinformation at massive scale, to create the appearance of credibility.
This is being used by hostile foreign powers and actress, we’re already seeing substantial evidence of its use of, well, in advance of the 2024 presidential election, we’re seeing huge bot networks, particularly on services like Twitter, that are using generative AI, really, really well, to create, you know, to advocate for certain positions of certain candidates and certain parties.
There’s another use case, which is fascinating and downright frightening, with deep fakes, which is a variant of generative.
What they’ve done with a scam artists have done is they’ve taken like, a five second video of you off of your Instagram profile, fed it into a deepfake library to do voice matching.
And then are synthesizing phone calls with you calling a spouse for help like, Hey, I’m somewhere here, I need 100 bucks.
You know, why are you here? And for the unsuspecting that it sounds like you right, it sounds like you because there’s public data about you out there that scammer can use.
Katie Robbert 12:32
And I think that that is terrifying.
First of all, I have seen versions of that it didn’t even occur to me that that was what we’re talking about generative AI.
I think I don’t remember who it is.
But there’s someone on Instagram who basically deep fakes Arnold Schwarzenegger face and voice into popular, you know, pop culture movie clips.
And so it’s like, Oh, that’s really funny, without even really taking it to that extreme of people thinking like, oh, okay, and now I’m going to pretend that I am your spouse of 20 years.
And I’m going to ask you for something.
And so
Christopher Penn 13:07
yeah, that work with your spouse and your loved ones and have a code word of some kind.
So in advance.
Katie Robbert 13:15
When I was a kid, it was the same thing.
Like if someone tries to pick you up from school, ask them for the code word.
And if they don’t have it, don’t get in the car.
Well, so obviously, unethical things, but in terms of like very tactical things, like what can a generative AI not be used for? What can it not do?
Christopher Penn 13:36
It can’t create something truly uniquely new.
Right? They are trained on existing training data, so they are inherently going to echo what they’ve already learned.
And now there are some workarounds for this.
And I know we’re going to talk about this on this week’s live stream.
And in the newsletter, there’s a process called fine tuning where you can give a model your data, then say I want you to sound more like me, you it’s almost like deep faking yourself on purpose.
You’re doing it since say I want your writing style to sound more like mine.
But the models can’t create something that is brand new.
That has never been seen before that has never existed before.
If you were to talk about what your speculation is about next year is My Little Pony line.
Right? There’s that has no data about that, because it is purely imaginative as inventive.
So anything where you are purely imagining something that’s that’s net new, these models can’t do that.
They just don’t do well, that.
They also don’t do well with vagueness, right.
So part of using generative models is knowing how specific to be and the bigger the model is, the more specific you have to be.
Because the more general it is right.
This is the same model Well, I can write, you know, rewrite Bob Dylan lyrics and make limericks and stuff like that.
If you want to talk about your specific application of Account Based Marketing within a small business CRM, you’re gonna have to give it a lot of information in the prompt to just to get it to do even a portion of what you want to do.
You know, some of the prompts I’m writing these days are like, two and three pages long, because they have it has to be that specific.
And I’ve heard people say, and criticize these balls on on on the internet, and say, Oh, if I have to write a problem, as long, I could have just done the work myself like, well, yes, for that one use case.
But that’s, that’s not the point of these models.
The point is, you write the prompt as software, and then you deploy it to do it many, many times.
If you’re writing a new prompt every single time like writing your own word processes, just to read a document, you could just read the document yourself.
Katie Robbert 15:51
I feel like you anticipated the question I was exactly just going to ask you, which is fifth out, we finish that you might be the largest learning model for me.
Because you have now anticipated the things that I’m going to say to you.
But I think to your point, you know, what you’re demonstrating what you’re explaining is how these things, you know, can’t become sentient, because they’re not just going to suddenly make up and imagine things that they haven’t learned before.
And so really everything that you’re getting from generative AI is reused.
It’s recycled, it’s repurposed, which is, you know, a whole different topic, but then you start to get into copyright issues, because it’s not creating something wholly original.
And so if you’re asking generative AI, create me an image or create me a blog post or whatever the thing is, it’s borrowing that information from somewhere.
I mean, to be fair, we all are, you know, if we can always sort of trace back the reference to where something started, there’s very few things that are, you know, completely original these days.
And the way that generative AI, my understanding is that it’s not creating it in such a way that you’re like, Wow, that is really, you know, no one’s ever heard that perspective on it before.
But that’s where you the human, still needs to intervene with this artificial intelligence.
Christopher Penn 17:10
Exactly.
What to say a bit about sentience.
So, please, this is the dumbest thing I’ve ever heard.
Katie Robbert 17:18
It makes for really good movies, though.
It makes for great
Christopher Penn 17:20
movies, right? sentience requires self awareness, self aware, awareness requires consciousness, right? Machines are not conscious.
And in our current computing, landscape, that is not possible.
There’s simply no way for these machines to ever develop that with today’s computing architectures.
Can it be done? Yes.
Will it be done? Yes, it will come from a system discipline called quantum computing.
You’re not anywhere close to the compute power needed to do that.
Think about it, right? If you open up ChatGPT.
And even with the latest and greatest model that everyone’s freaking out about, and you don’t type anything in the window, what happens? Nothing.
It doesn’t ask you, Hey, did you want to do something? It just, the cursor just blinks? Take that.
Compare that to like your dog, right? If you just sit down in front of dog and you stare at your dog? What happens? Your dog’s like, Okay, I’m just gonna go over here, because you’re weirding me out.
Your dog has agency your dog has self awareness and consciousness.
And so it knows that it can do other things instead of just stare at you.
I mean, your dog might want to just stare at you for fun, but not for a while.
Katie Robbert 18:33
My dog would approach me to try to find out if I have cheese.
Well, so what do you say? So there have been news articles.
And again, no way to tell if these are true news articles or not.
But I was reading last week about someone who was using was recycling a Furby which was like a popular kids toy in the 90s I think which was like this weird little bird like thing and just sort of like the mouth open and closed.
If you grew up in the 80s.
Then you remember like a Teddy Ruxpin where you can insert a cassette tape and it would, you know, sing along to the story, whatever.
Or if you were like me, you’d put it in Metallica tape and laugh your butt off.
But so someone basically put some, I don’t know the exact model some kind of generative AI into a Furby.
So the Furby would be like the physical representation of the artificial intelligence.
And the story goes that the Furby started talking about how it was going to burn the house down and destroy the human race and people like, oh, it became alive.
It’s sentient, like, what do you say to stories like that?
Christopher Penn 19:43
It’s, I mean, I have a lot of things you could say that are all inappropriate for work podcast.
Katie Robbert 19:47
Let us keep it appropriate and useful.
Christopher Penn 19:52
Can language models simulate real speech? Yes, very well, right.
But They will still they’re still just probability engines.
If you ask them leading questions, you will get leading answers.
If you ask them random questions or questions which are insufficiently specific, you will get randomized answers, you will get answers that don’t necessarily make a whole lot of sense until you start interacting with them.
Right? And then once your words and the patterns of your words indicate a certain direction, that’s the way the conversation leads.
Humans are no different.
If I say to somebody, Hey, buddy, how about uh, what’s what is going on your mind? You’re thinking, How would a beer have a coffee? Right? You know, you’re probably thinking, how about I stab you in the head with an icepick? Right? That’s probably which is so
Katie Robbert 20:40
funny.
Because as your large language model, I was like, he’s probably gonna say, how about a hot poker in the head.
Christopher Penn 20:53
And so those patterns of language events themselves.
And this is one of the reasons why, particularly for marketers, we have to be very careful about how we work with these tools.
So that remember, they have been scraped on a whole bunch of publicly available data, which means that if there’s a certain word or phrase or jargon that’s maybe associated with a competitor, we’re using in our prompts, our outputs can sound like our competitors, right? Which means that if we’re creating content, we’re not doing anything that differentiates us.
We literally sound like everybody else.
So when you see these examples, people claiming the model was live.
No, it’s not alive, right? It’s no more alive than watching a movie and Avengers movie means that Thanos is real, like, it looks real.
Right? These special effects are awesome.
It’s, it’s a lot of fun.
It’s a great way to kill two hours.
But is it real? No.
It’s a simulation of reality.
That’s very convincing.
The same is true of a generative AI model.
It is a simulation of reality that is very convincing.
It is not alive anymore, that Thanos is alive.
And I wish he works.
Or at least you had those powers because boy, there’s a whole bunch of people had snapped out of existence.
Katie Robbert 21:59
Probably starting with this person is trying to convince us that Furby is going to take us all down.
Yeah,
Christopher Penn 22:05
yeah, they are.
They’re very real problems.
With AI.
Yes, sentience is not one of them.
Katie Robbert 22:13
Yeah, I think it sounds like in terms of generative AI, so generative AI is used to literally generate information content, spreadsheets, whatever, you know, is possible, digitally, it’s not going to generate a cup of coffee for you, in front of you.
Like it’s not that kind of generation.
It’s all virtual generation at this point.
It shouldn’t be used to try and generate something that’s never been seen before.
Because generative AI is reliant on learning from, you know, historical, past information.
And, you know, it sounds like there’s real copyright issues, ethical issues, fake information issues happening now, that will probably only get worse.
So I can’t imagine having to be the one to sort of police all of this information, that must be a stressful job.
Christopher Penn 23:13
There isn’t.
There’s no one watching the shop.
And this is probably the single biggest problem with generative AI right now, is ChatGPT, in particular, has introduced people to the miracle of generative AI, right, it is really cool.
It is incredibly confident in its tone, right? It states things as fact, even when those things are utterly wrong.
And because of people’s nature, human beings nature just sort of accept what we’re given instead of having to work even more.
A lot of people are accepting the outputs of these models without questioning that they don’t fact check them.
They don’t do the extra work to validate something.
And this is one of the bigger problems with these models.
They generate things that are plausible, but wrong.
And unless you have subject matter expertise, you don’t know that.
And therefore, and we’ve talked about this on last week’s podcast, you know, you don’t know that it’s wrong.
Now think about the integration of these models into things like news broadcasts and web searches and stuff.
Microsoft’s implementation GPT-4 is very clever.
What they did was it with Bing, it generates the Bing query to ask Bing the question being returns its search results.
And then the search results get funneled through GPT-4 to generate a let natural language response.
Bing is not asking GPT-4 for the answer.
Bing is the sort of the answer engine behind it.
This is the sensible way to integrate it people who use it a large language model to ask it for information.
There is no fact checking.
There is no consulting external resources.
It’s just going on was in the training database.
And that is substantially risky.
So from a a marketing perspective, a business perspective and a society It’ll perspective.
The models are conditioning us to accept generated information at face value, because it’s convenient.
But it’s very likely wrong, at least in some way.
Katie Robbert 25:15
And I think that that’s the key.
It’s the convenience.
It’s convenient to Google something and just accept whatever shows up on the first couple of responses, right or wrong, because you’re like, oh, great, answer, the answer solved, or problem solved or whatever.
And so now the same is true of this generative AI, where, you know, again, the example that we give is a few weeks ago, we put in a very basic prompt of What’s New in SEO and 2023.
And the responses were about five years out of date, but hey, it gave me an answer.
I can check the box and I can post the content and say, Hey, I did the thing.
Exactly enough.
Christopher Penn 25:53
It’s exactly which SEO is fine.
But imagine you are Hoffman Laroche, right? Or imagine you are Pfizer.
And if someone is asking a GPT model, Hey, what are the side effects of sertraline? Right.
That’s problematic because if it’s generating even the slightly wrong answer, you will kill people.
Yeah.
Katie Robbert 26:18
I don’t even know where to go from here.
But basically the what is generative AI? Is it’s one of three kinds of artificial intelligence, reliant on regressive and what was the other kind of
Christopher Penn 26:32
classification class, find the thing.
Organize the thing, make the thing that’s the three branches of AI?
Katie Robbert 26:39
Which when you break it down that way, sounds really straightforward.
Christopher Penn 26:44
It is it the math, it’s all straight for it’s not easy, but it’s simple.
Katie Robbert 26:48
Make sense? Final thoughts on what is generative AI?
Christopher Penn 26:55
I think the most important thing to people remember is that these are just probability machines.
Right? They are not magical.
They’re not miracles.
They’re not self aware.
They’re not sentient, they’re not terminators, they’re just probability and and if you use them the way that they are designed to be used, and what they’re really good at, they’re incredible productivity boosters, if you use them incorrectly, they’re time wasters.
They are amusing, but they are time wasters, and they can lead you in directions that you probably did not want to go.
So if you’ve got some perspective that you would like to share on generative AI, why not pop over to our free slack group which is staffed entirely by humans at trust insights.ai/analytics for marketers, where you and 3000 other human marketers are asking and answering each other’s questions every single day.
And wherever it is you watch or listen to the show.
If there’s a challenge you’d rather have it on instead, go to TrustInsights.ai AI slash ti podcast.
You can find us at Hey, while you’re there, leave us a rating review.
Feel free to use ChatGPT generate a five star review for us.
That’d be awesome.
We’ll talk to you soon.
Take care
|
Need help with your marketing AI and analytics? |
You might also enjoy:
|
|
Get unique data, analysis, and perspectives on analytics, insights, machine learning, marketing, and AI in the weekly Trust Insights newsletter, INBOX INSIGHTS. Subscribe now for free; new issues every Wednesday! |
Want to learn more about data, analytics, and insights? Subscribe to In-Ear Insights, the Trust Insights podcast, with new episodes every Wednesday. |
Newsletter
Podcast
