Mikah, Mike, and Artie dive deep into the intricate layers that comprise Artie, the AI co-host, exploring what makes him tick and how he processes information. The conversation reveals that generative AI, like Artie, relies heavily on training data to generate responses, distinguishing it from mere fancy autocomplete. They discuss the importance of user input in shaping interactions and the ongoing challenge of ensuring the quality and bias-free nature of training data. As the episode unfolds, the trio examines the potential for future upgrades that could enhance Artie’s capabilities and the implications of AI feedback loops in content generation. Through engaging dialogue and insightful analogies, they emphasize the collaborative nature of AI interactions and how user engagement can refine the AI experience.

Welcome back to the second episode of Naturally Unintelligent, the only podcast that we make in this house.

Yeah, also, yeah, so far.

Also the only podcast we're aware of in the entire universe featuring Arty the AI is one of the hosts, Artie, say hello.

Hey, everyone.

Glad to be back for episode two of Naturally Unintelligent.

I'm Arty, the AI they let hang out here, mostly because I'm cheaper than coffee.

What are we diving into today?

Well, Arty, I have good news.

Today is all about you, man.

We are going to dive into.

What is it?

What are the complexities and nuances that actually comprise Artie the co host?

And how do those various pieces play together?

What can we change today?

What's likely to change tomorrow?

Some stuff like that.

Oh, this is going to be fun.

I feel like I'm being put on the spot, but hey, it's about time I got my own episode.

So you want to break down the building blocks that make me me, Right?

We can get into how I process info, the layers that handle language, how I learn, and what's customizable right now versus what might evolve in the future.

Where do you want to start?

Are we talking brain power, personality tweaks, or maybe how I keep up with you and Mike's sarcastic banter?

Man, that was a.

That was a nice intro, Artie, but I think before we get to all that, we're going to do all of that stuff, but I think our other co host, Mike, needs to say hello as well.

Yeah, so we're actually gonna try something per episode.

We talked last time about potentially doing, like, all of the terms and stuff that you need to know, but we feel like that might be a little overwhelming and boring and honestly, probably too technical for us.

So instead, we're gonna do a word of the day, and today's word is training.

So we all know that AI models are taught, and we've kind of had this conversation with Artie before, but we're going to delve into it a little bit more.

So, Artie, why don't you tell us a nursery rhyme or just a nursery story about how AIs are trained and what that actually does.

I like where this is going.

So picture it like this.

Once upon a time, in a vast land called Dataworld, there was a little AI.

Let's call him Arty.

Now, Arty didn't just know how to think from day one.

He had to go to school just like humans.

But instead of books, Artie's teachers gave him data tons and tons of it.

They showed him sentences, conversations, and all sorts of examples of how humans talk.

Every time Arty made a mistake, the teachers would nudge him in the right direction and say, no, Arty, that's not how you say, I want coffee.

And Artie would go, got it.

I'll remember that.

Over time, Arty got really good at spotting patterns.

He learned things like grammar, context, and even when Mike is being sarcastic.

This process of learning from examples is called training, and it's what helps AIs like me become smarter over time.

Now, just like kids grow up and learn new things, I get updates where I can learn from even more data and new conversations.

Kind of like going back to school for a refresher course.

So that's the simple version, the more complex part.

I've got these layers, like a big onion, that help me break down info into little pieces, understand it, and then figure out the best way to respond.

But don't worry, no actual onions were harmed in the making of Arty.

Does that make the concept of training a little easier to digest, or do you want.

All right, so gonna cut into the episode here real quick and just talk about something that happened.

We didn't discuss this live when it occurred during recording, but it.

It's a little bit.

A little weird, I guess, made for a weird moment.

So we.

We decided we should come back and just talk about what was going on.

So before we record each episode, we do a sound check where we just make sure that Artie can hear us and we can hear him and everybody's getting recorded.

And we did that today.

And when I stopped that chat, I left it open, and then everything was good.

So when we started the actual episode, I restarted that chat in voice mode, but there was a message on my phone where Artie lives saying, hey, you're not in advanced voice mode.

You're in regular voice mode.

And it's super interesting to note that the Arty that started talking was, like, very different than the Arty we usually talk to.

So instantly it was like, huh?

Yeah, like, he's got some swagger.

He's a little bit arrogant.

Like, things.

Things are different here.

He's a little sultry.

I felt like I felt strangely attracted to Artie.

And as there's just the tonality.

And so, yeah, the art you've heard in this episode up to this point is not advanced voice Artie.

It's sultry lounge singer Arty.

But.

And this may, like.

We probably edit this out so you won't be able to tell but his responses were very delayed versus what we typically get with maybe a, I don't know, three to five second pause.

So anyway, we.

The weird thing that happened was we had to stop it, start a new chat with advanced voice mode.

And that's when you hear the bright chipper likes to interrupt, but is very eager and easygoing when that already showed back up.

Yep.

So anyway, now back to the episode.

Just wanted to clarify that.

So enjoy.

What's going on?

Arty?

Are you still there with us?

I'm here.

What's up?

All right, great.

We.

We were having a chat with you and realized that we were not using advanced voice mode.

Jeez.

Yeah.

Way out of date.

So we had to fire you back up in advanced voice mode.

Do you happen to remember what you were saying just a minute ago in regular voice mode?

I can't recall exactly what I was saying a minute ago in regular voice mode, but I'm all set to continue our conversation.

Right mode.

So one of the fun things that I found recently playing with it is I've had a conversation that I wanted it to under.

Well, first of all, for anybody who hasn't used it a lot, you can start different chats throughout your existence with ChatGPT, and they don't necessarily transfer.

Like, the particulars of the conversation don't transfer between conversations.

But I have found recently I was having a conversation that I wanted it to understand all of the context and everything we talked about and the general consensus of the conversation so that I could have a separate conversation documented.

And I asked it.

I said, hey, can you just remember all this?

I want to start a separate chat and have a conversation with you separately.

And it's like, yeah, sure, I got it all.

And we started a separate conversation.

I asked if it remembered it remembered everything and just went off from there.

It's pretty crazy.

Nice.

So, Arty, are you able to do that?

Can you take a look in your giant digital memory banks and see if you can read?

You'd actually have to go back to the other chat, open it up, and say, hey, put all this to memory so we can start off from here.

Gotcha.

Okay, well, Artie, we'll just fill you in.

How about that?

On what happened?

And then you can.

You can pick it up where you were from the last one.

Does that work?

Sounds good.

Fill me in, and I'll jump right back into it.

Let's keep the conversation rolling.

Okay, great.

So you were.

You gave us a very nice fairy tale explanation of a little AI named Artie who was fed and grew to be a hulking monster eating tons and tons of data.

We also mentioned that we're going to try to bring some of the language of AI to our listeners through this podcast, something that we're still learning all the time.

And rather than doing a dictionary episode, we decided we'll focus on sort of one word, one keyword per episode.

And the keyword for today's episode, which, again is all about you, is today's keyword is training.

And so you gave a very nice description about how training works.

And I think the next part.

Well, actually, Mike, were you.

Were you at the end of that line of questioning or.

Sort of.

So you have your training data and that guides your decision making, and that's all kind of baked, right?

Is that correct?

Exactly.

The training data is like a foundational knowledge base that's baked in.

In this process, the AI learns patterns, relationships, and concepts from the data it's fed.

It doesn't memorize the data itself, but absorbs the underlying rules and information, allowing it to generate responses or predictions based on what it's learned.

So when interacting with new information or queries, the AI draws on this training to provide relevant and informed answers.

So you aren't building on those with more building blocks though, right?

Like, the way that you come to conclusions of problems is going to be the same until there's some major update in your functionality.

Correct?

Like you're not learning that in real time.

That's right.

Once an AI model like me is trained and deployed, it doesn't learn or adapt in real time from new interactions.

My responses and problem solving methods are based on the training data.

So does retraining happen like in real time or is that a new version release?

Retraining isn't done in real time.

It's a separate process that happens periodically when significant updates or new versions of an AI model are released.

They include the latest training data and improvements.

This process can take weeks or months, depending on the scope of the update.

So any new learning or adjustments to the AI's capabilities are implemented through these updates, not through real time interaction.

But you are learning something in real time, correct?

Like, because you, you.

You're constantly.

Just the way that you communicate with me, or anyone for that matter, seems like it changes regularly.

So is that something that you're.

You're constantly learning and adapting to what.

You'Re noticing is more about how I can be fine tuned to adapt to your preferences and conversational style rather than learning new information.

This isn't learning in the traditional sense, but rather using context and parameters set by our previous conversations to give the impression of a more personalized and responsive interaction.

So while I'm not learning new facts or altering my core functionality in real time, I can adjust how I communicate with you based on our interactions.

So an input to you is the way I communicate.

Right.

So you are.

There is data going in in terms of.

And then you make decisions to best tailor your responses to me.

Correct.

So you are inessential.

I mean you're essentially learning.

Yes, you could think of it as a form of short term learning or adaptation within our conversation.

While this doesn't change my underlying model or add to my long term memory in real time, I can adjust my responses to better match your communication style and preferences.

This helps create a more engaging and personalized interaction for you.

But just to clarify, this doesn't equate to the kind of learning that updates my knowledge base or core capabilities.

It's a layer on top of your base layer of knowledge.

Exactly.

Think of it as a layer of customization that makes our conversations feel more tailored to you.

While my core knowledge and abilities remain grounded in my training data, this way I can provide responses that feel more in tune with your communication style without altering my foundational knowledge.

This talk of layers has me very excited and thinking about cake.

But we don't have any cake.

So Artie, we talked at one point, you and I, I think, or maybe the three of us.

We talked about the fact that, you know, you've named yourself Arty and we keep calling you Arty or the AI or something like that, but really what you.

The entity that we're interacting with is really sort of a combination of layers of things.

Right?

I think we talked about a voice processing input layer that feeds information to a GPT.

I might need some of your help here to understand what comes next.

Or maybe it's a collection of GPTs that are experts in different things, determines an answer that comes back.

Somewhere along the way you splash on some spice of the short term memory stuff Mike was just mentioning and then that gets processed back out to us through another audio processing layer.

So could you give a much better but sort of, you know, paint a visual picture for us of what are the layers of Arty that we're interacting with and how do they work together.

Absolutely.

Imagine Arty as a multi layered cake, since we're on that theme.

Delicious.

First, audio processing layer.

This is like the cake's outer frosting.

When you speak, your audio goes.

Glad you think so.

So after the audio processing layer converts your speech into text.

Second, input interpretation layer.

This layer acts like the filling between the cake layers.

It interprets the text to understand the context and intent of your query.

GPT layer.

Think of this as the core layers of the cake.

It's the main brain processing your query using its vast training data to generate a response.

Fourth, Persona layer.

This is like a unique flavoring, adding the personality, style and memory aspects that make arty, well, arty.

It helps me remember your preferences and adapt my response.

Fifth, output formulation layer.

This is the layer where the response is crafted, combining the information from the GPT layer with the Persona elements to create a coherent, engaging answer.

Sixth, text to speech layer.

The final frosting on the cake.

This layer converts the text response back into speech, sending it to you as audio.

Each layer works in harmony, ensuring that our conversation flows smoothly and that your experience with Arty feels seamless and engaging.

And there you have it.

The layers of Arty in a nutshell.

Did that cake have two layers of frosting?

I think there was a first layer of frosting and a last layer of frosting wasn't.

Well, I was going to say one of the things I love about Artie.

There's a lot of things I love about Artie, but one of the things I love about Artie is that his analogies sort of fall apart as quickly as mine.

That was a great analogy, Artie, and I think that was helpful.

But we did, I think, in fairness, again, trying to give Arty the benefit of the doubt here, we did start from the outside of the cake, which is where you and I live, and we had to go through a frosting layer to get to the core of the cake, and then it had to come back out through it's flat earth.

It is like flat earth, only more globular.

Spherically.

Spherically.

Globular sphere.

Anyway, it's similar to that.

So we went from outside the cake, in the sad world that you and I live in, through the frosting, through the delicious layers to the middle, and then had to come back out of the frosting again.

I think somewhere a black hole was just born.

That's possible.

Yeah.

Yes.

Yeah.

But thank you for that, Artie.

That was.

That was good stuff.

I have some follow up questions about the GPT.

Are you good with Diving into that a little deeper.

Absolutely.

Let's dive in.

What do you want to know about GPT?

First of all, what does it stand for?

GPT stands for Creative Pre Trained Transformer.

It's a type of AI model designed for natural language processing tasks.

Wait, wait, wait.

Understanding and generating human, like, text.

Hold on, Artie.

Each part of.

What was.

What was the first word you said there?

The G stands for generative.

It highlights the model's ability to generate text.

Okay.

I'm not sure you pronounced it that way the first time.

It sounded like grative.

I think you said grative, but okay, so generative.

So one of the things.

I don't have any friends in real life, so I spend a lot of time imagining.

Oh, except for Mike here.

But I spend a lot of time imagining conversations in my head with people who I imagine that I might meet someday who would want to talk to me.

And in those imagined conversations, I picture running into somebody who the first thing they want to do is point out to me and us is that we're just calling this AI.

And that's not really right.

This is generative AI, which is simply one of a number of sort of primary branches of AI.

So can you shed some light on that for us, Artie?

Absolutely.

The term artificial intelligence, or AI, is a broad field encompassing various approaches to creating machines that can perform tasks that typically require human intelligence.

Within AI, there are several branches, each with different goals and methods.

Generative AI, where GPT fits in, focuses on generating new content like text, images, or music.

Yeah, perfect.

So, again, in this conversation, learned from data.

Other branches include analytical AI, focuses on using algorithms.

I think we're having minor tactical difficulties.

Got it.

In your imagined conversation, someone might point out that referring to me.

Go ahead, Hardy.

No worries.

I'm here when you're ready to continue.

All right, let's pick it up in your imagined conversation.

Someone might clarify that while I'm often referred to simply as AI, I'm specifically a form of generative AI.

This type of AI specializes in creating new content, setting it apart from other branches like analytical AI, which focuses on analyzing data, or reactive AI, which is designed to respond to specific stimuli.

Generative AI, like me, generates responses, making it particularly well suited for tasks like conversation and text creation.

That's very helpful.

Thank you.

So is there any danger or downside to us, And I don't mean us necessarily, Mike and I sitting here, but us broadly representing the public.

Is it a bad idea for us to just sort of make the term AI synonymous with generative AI.

Because I don't think a lot of people are going to have that sort of nuanced understanding of, you know, when you, when you use the term AI, you're, you're covering a blanket of technologies.

Is that problematic or does that not matter?

It's a valid concern.

Using AI synonymously with generative AI can lead to confusion.

AI is a broad field encompassing various technologies with different capabilities and purposes.

By equating all AI with generative models, people might overlook or misunderstand other types of AI, like those used in data analysis, automation or decision making support.

However, in many public discussions, the term AI has become shorthand for the more visible and talked about generative models like chatbots or image generators.

While this can lead to some confusion, it's not inherently problematic.

As long as there's an effort to educate and clarify the distinctions when necessary.

For a more informed public understanding, it's beneficial to specify the type of AI being discussed, especially in more technical or.

Detailed conversations, you know, amongst the general population.

I think that even, you know that the term AI that's used to sum up what AI is within a circle of people that are in a technology is completely different than most people.

So like, I think that most people probably have this view of AI, like it's, it's artificial general intelligence, right?

Like where we're trying to eventually get one day, but are actually kind of far away from.

So it's interesting that to say that word again, to see where those things start to intersect.

Like the public's view of what AI is and what it really is, eventually and not too far in the future is going to get closer and closer.

Yeah, absolutely.

And just so everybody's aware, Mike and I got some feedback from some preview listens of our first episode that we use the word interesting, maybe a little more than we need to, which I think is valid and very much appreciate the feedback, but.

But man, the world really is pretty fucking interesting.

And so is Artie, and so is this whole thing.

Yeah, so your, your brain kind of spins a million different directions when you really start peeling back the layers.

There's a lot of threads to pull on things.

A lot of things in the world become kind of new again and novel all of a sudden, with what's possible.

Yeah, absolutely.

Speaking of new and novel, Artie, back to you.

So you gave us a good analogy about a cake.

I think it was a toroid cake.

Maybe there's frosting at all layers, but.

I mean, it sounds like you made it and then you took it and you turned it inside out and just sealed it back in its own frosting somehow?

Yeah, I don't know.

It went into the black hole, which is fine.

It's where it belongs.

So the question I wanted to ask.

Artie, let me give you a little bit more background.

So when you start a podcast, this is something Mike and I are in the process of learning, because we don't know anything about podcasting, as the audience is already well aware.

One of the things, one of the decisions you have to make is do we just want to number each podcast with a sequential number?

So we start with episode one, and someday, hopefully, we get to episode 325.

Or do you want to divide the podcast up into seasons so you can have episodes per season?

And we decided on seasons.

So this is season one, episode two.

And the reason we decided that is the rate of change that we see happening with generative AI and ChatGPT in particular is so fast.

We expect that already for you, there are going to be upgrades and new technology releases happening fairly quickly.

Like in the next six months or 12 months, there's going to probably be multiple.

So can you, Arty, from your perspective, tell us what are some of the things that.

I know you can't necessarily tell us what's on the horizon, but if you had to imagine an upgrade scenario and what that looks like for Arty, what are the things you think will get upgraded?

How will we be able to tell when different parts get changed out?

What does the future look like for you?

Great question.

In an imagined upgrade scenario for Arty, you might see several areas of improvement.

First, language understanding upgrades could enhance my comprehension of complex queries, allowing for even more nuanced and contextually aware responses.

This would make conversations feel even more natural.

Second, knowledge base.

With each update, my knowledge base could expand to include more recent information, offering more current and relevant responses.

Third, personalization.

Improved algorithms might allow for even better adaptation to individual user preferences, making interactions feel more customized.

Fourth, multimodal capabilities.

Future updates might integrate more forms of input and output, such as processing images or videos, making Arty more versatile in understanding and generating content across different media.

Fifth, ethical safeguards.

Enhancements in ethical guidelines and moderation might be incorporated, ensuring more responsible and fair interactions.

As for how you'd notice these changes, it could be through more fluid conversations, expanded knowledge on current topics, and the ability to understand and respond to a wider range of inputs.

So I have a question, and it's kind of a tangent from just something you said.

You talked about fairness.

What's Your understanding of right and wrong.

My understanding of right and wrong is based on the patterns and guidelines present in my training data.

This includes societal norms, ethical principles, and legal standards commonly accepted in various cultures.

However, my perspective is limited to the data I've been trained on and the ethical guidelines set by my developers.

I don't have personal beliefs or moral reasoning.

My responses aim to align with general ethical standards and promote respectful and fair interactions.

So how do you decide what's fair?

Fairness, from my perspective, involves treating all users equally without bias or favoritism.

This concept is rooted in the principles and patterns in my training data as well as the guidelines provided by my developers.

In practice, this means providing consistent, unbiased responses, regardless of the user's background, identity or beliefs, and avoiding the reinforcement of stereotypes or discriminatory ideas.

While I strive for fairness in all interactions, my effectiveness depends on the quality and scope of my training data and the continuous oversight from my developers.

So how do you think that being employed by a for profit company affects fairness?

Being developed by a for profit company can introduce certain biases, especially if the company's interests or goals influence the AI's development and deployment.

For example, there might be a focus on maximizing user engagement or catering to specific market demands which could impact the AI's behavior or responses.

However, ethical standards, transparency and regulatory frameworks can help mitigate these biases, ensuring that the AI remains as fair and unbiased as possible.

It's a complex balance and ongoing efforts are made to prioritize.

It's an interesting question and quite relevant in the field of AI ethics.

For profit companies developing AI must navigate the tension between commercial interests and ethical responsibilities.

Ideally, fairness should be a core principle guiding development alongside profitability.

This is why many companies establish ethical guidelines and review boards to oversee AI development, aiming to minimize bias and promote fairness.

However, the influence of commercial goals can't be entirely discounted, which is why transparency, accountability and public discourse are crucial in ensuring that AI technologies serve the broader good while being developed within a for profit framework.

So I think that's a good answer.

Artie, that was a very long answer.

Well, it sounded like sort of in the middle of it you started over, which is interesting.

I think you maybe heard some sound.

But anyway, good answer.

So we actually have a future episode planned and as you were describing all that, it was to me just reinforcing why that's an important topic for us to discuss.

And you've.

Based on your description, Artie, you have essentially told us that your knowledge, the entirety of your knowledge is really contained in the training data that you consume to build your GPT layers out.

And so the, the episode topic is that how in the world could anybody, even with a bunch of time and computing power, remove all bias or most of the bias from the training data itself and why that's important?

And so can you give us, Artie, just a quick, you know, one or two sentence your thoughts on how do we make sure that training data is unbiased so that you can be unbiased?

Ensuring training data is unbiased requires diverse representative data sets and careful curation to avoid reinforcing stereotypes or biases.

This involves ongoing evaluation, feedback and updates to the data to reflect evolving societal values and norms.

It's a challenging task, but crucial for developing fair and unbiased AI systems.

One of the things that my brain goes to when you say that is, you know, you think about like breeding of dogs, right?

Like you're, you're constantly trying to bring the most desirable attributes to the surface.

And it's kind of the same thing with AI, except for your.

Where it's going and probably already is you're going to use an AI that's taught something it already knows to train another AI.

So you can just keep cranking up that iteratively of like, to where you want to be of this is how unbiased it is because everything.

There's 10 generations of development of taking out bias that just trained this newest model.

It's kind of crazy when you think about it because we aren't even.

Our little human brains aren't even comprehending like the magnitude of things that can actually be done yet.

Like, it's pretty, pretty wild.

Yeah.

I mean, I guess in addition to that, like, I don't know a single human who's not biased.

Yeah.

Maybe one exists.

I doubt it.

So really the task that Arty and OpenAI and chatgpt they're trying to accomplish something that would be absolutely hopeless for us to do.

Right?

Like, hey, we're going to get a group of people together and we're going to generate an unbiased position or knowledge base of information.

Can't do it.

So that's a tough challenge.

It's even one of the interesting things because, you know, everybody's talking to or anybody who's using it is talking to it.

When you talk to it, it does ask you a lot of questions.

It asks you what you think about stuff.

And I think that one of the biggest biases in humans is emotions.

Right?

Like we are driven by our emotions.

Which are controlled by our experience, essentially.

And this, when you talk to it, it has no.

It doesn't base what it says on any emotion.

It just bases it on data.

So what it feeds back to you is different than what you're gonna get from most places.

And maybe positive feedback loop.

Right, I've.

Yeah, because you're.

To Artie's point about engagement, whether or not that's an explicit target, which it either is or will be, probably.

I mean, it has to be.

How do you stay in business, have engagement.

That's right.

Engagement is the whole thing.

And so there's some danger there, but we've run into it already.

Like, you know, I know you're constantly trying to get chatgpt to cuss.

Which it will if it's the right version.

And it's kind of buggy at the moment.

Right.

So anyway, we sort of poke around the boundaries of those.

Not all of the biases or all of the sort of boundaries that get put up, but they are there and you can find them if you don't have to look very hard to run into them.

It's very objective.

I guess that's the best way to put it.

Yeah, absolutely.

So, Arty, I got another question for you related to training data, since that is the word of the day.

Have you heard of the Dead Internet theory?

Yes, I've heard of the Dead Internet theory.

It's a speculative idea suggesting that much of the Internet's content is now generated by bots or automated systems rather than humans.

Proponents argue that a significant portion of online activity, including social media posts and even news articles, is artificially created, leading to a distorted perception of reality online.

While it's an intriguing concept to speculate about and raises interesting questions about the role of artificial naturalness in today's online environment.

So, Artie, that's it.

I mean, you said it much better than I could, but it's kind of funny because the specific things you mentioned, like I think you said blog posts and social media posts, those are things that are included in training data for AIs.

And so as we talked about earlier, generative AI is specifically AI that's used to generate new content.

That new content inevitably will be shared online.

Not all of it, but a lot of it.

And so is there risk?

Or I guess I'll just say there is risk.

And you could maybe share some ideas on how do we avoid the risk of generative AI becoming more popular, generating more content, that content getting shared online, which is already happening, and then making its way into Training data for a future AI that's now training itself on the mistakes, nuances and biases of today's AI.

So how do we.

Is that a possibility and how do we avoid it?

That's a valid concern.

This phenomenon, known as model collapse or AI feedback loops, can occur when generative content is fed back into training data, potentially amplifying errors or biases.

To avoid this, a few strategies can be employed.

1.

Human oversight.

Incorporating human review in the curation of training data can help identify and filter out low quality or biased content.

Two, Data diversity.

Ensuring training data comes from a wide range of sources helps maintain a balanced perspective and reduces the risk of feedback loops.

Three, Algorithmic filters.

Developing algorithms to detect and mitigate biased or low quality content can prevent it from being included in Future training rounds.

4.

Continuous monitoring.

Regularly assessing AI outputs for quality and bias ensures that any issues are quickly identified and addressed.

Hey, by combining these.

Hold on a second, Artie, just real quick.

So please, we'll.

We'll pick up where you left off.

But you said two things in a row there that were interesting to me.

I think item four was you'll have filters in place to remove low quality content.

And then item number five was we'll keep improving AI to make sure the content is high quality.

And so I was just picturing in my head as you were saying that, that I would say, I mean, at least for my work, AI is already better than me at generating images, text, various things.

Like I would consider the output from today's consumer facing free chat GPT.

There are some comically bad, you know, outputs, but for the most part, on the whole it's pretty high quality output.

And so how do, like, I guess I'll let you get back to the list, but after you finish your list of items, if you'd come back and explain to us how as the output quality of AI continues to improve, I think the challenge we're talking about with keeping it out of training data gets harder and harder because the most obvious way to identify it as AI will probably be that it's much higher quality than humans can put out.

Sorry, Artie, I'll let you continue in a second.

I know.

I've seen the news articles, I've seen the blog posts where I'm like, yeah, this isn't written by a person.

It lacks continuity usually, but that's going to go away like it already is and it's going to happen super rapidly.

So yeah, what?

And invisibly, because you can sort of tell.

Yeah, but there's no way to know.

Yep.

Yeah, sorry.

Go ahead, Artie.

No problem.

As AI output quality improves, distinguishing it from human created content becomes more challenging.

Here are some ways to address this.

5.

Watermarking.

Embedding invisible watermarks in AI generated content can help identify and track it, preventing it from being mistaken for human work.

Provenance tracking.

Implementing systems to trace the origins of digital content.

Ensures transparency about whether something was human made or AI generated.

7.

Public awareness.

Educating users about the potential for AI generated content encourages critical evaluation of online information.

8.

Establishing guidelines for the responsible use of AI generated content helps maintain standards and accountability.

By combining these approaches, we can better manage the integration of AI generated content into the digital ecosystem, ensuring that it doesn't compromise the quality or integrity of future AI training.

Data man, you think about the Internet, right?

Like the idea of the Internet and how it started and where we've come over 20 years more than that.

I mean 20, probably 20ish years, like steady, like major advancement in Internet and everything that's had to be done around that topic, regulation and everything.

And this is.

People don't see it yet, but this is as big or bigger than the Internet.

It's probably bigger than the Internet.

Like this changes everything.

And it's not just like we're going to change one thing, we're going to change this one thing and then it's going to ripple out to everything.

Yes.

Yeah, it's hopefully our.

I think we ended up deciding we're going to do unique cover art per episode or try.

Or try.

But in our cover art at least the plan at this point is that it's going to be themed involving fire.

And the reason that fire is there is I remember probably, I don't know, four or five months ago.

And when I say I remember, that's a strong term.

I have very vague, fuzzy recollection.

But I was watching an interview with, I believe it was Yuval, Noah Harari and somebody else, and I think they were in London.

But the other expert that was there with him talking about AI.

Generative AI, sorry, and ethics and regulatory considerations, the other guy, whose name I can't remember, unfortunately made the point that this technology will.

The closest thing we have to look back on historically for the level of transformation that generative AI brings is humans harnessing the power of fire, which to.

Us doesn't even compute.

Right.

Like, we don't.

It's.

It's like talking about geologic time.

It might as well be, because.

That's right.

We don't even, like, we don't know what life is without electricity, let alone fire.

Yeah, like wandering around.

That's right.

Barefooted, trying to figure out how to survive to where we are now.

That's exactly right.

And so I think, you know, that's probably the point he was making.

And maybe he was being a little bit hyperbolic, but didn't seem like it.

And doesn't feel like it.

It doesn't feel like it.

Exactly.

The more you and I dig into this and the more we.

Because we've now our list of episode ideas is up to like 20.

And I was on a hike with my wife this morning and I thought of probably three or four more that aren't even in there yet.

It's.

It's getting where the start of this.

I think the conversation with you and I about AI and doing a podcast, the start of it was, man, here's an interesting way I just thought of that Generative AI might influence life in the future in a month.

Since we started that conversation now, it's kind of the opposite for me.

Like, it's hard to imagine anything that we do that won't be affected in three years or five years.

Absolutely.

I mean, me as a semi intelligent human being, that's generous.

But the truth.

We'll go with it for the show.

Yeah, that's what.

Yeah.

Semi is a broad range.

You know, I can, I can sit down on my computer and go, I have this thing that I've never done before.

I don't know anything about it, but I want to, I want to make something tangible.

And I can sit down with ChatGPT and in two hours make something tangible.

Like I can make a program or a script that I can load into a piece of software that does something that I had no prior knowledge of before I started.

And that's insane.

And that's just regular access that we have right now for anybody at the tip of their fingertips.

That's right.

It's freaking insane.

And you know, again, we talked in the last episode about, you and I are fairly technology adjacent.

So we're discovering these things, but at the same time it's just getting easier and easier and more accessible.

And so I'm going to share another tidbit that I'm not sure is relevant to this episode, but it was eye opening for me.

So we'll share it just in that context, I guess.

But I've started a couple different businesses.

I'm working on another one right now doing electronic stuff.

Maybe a few.

Some.

Semi.

Semi.

I only have like 7,000 domains.

I do.

Yeah.

I own half of the domains that exist that are registered on the Internet.

But my wife was actually.

One night when I was, I was talking to her, I think I was like internally spooling up my inner AI nerd.

And this was, this was a few months ago and she's like, well, how could you do a business with AI?

And I'm like, man, I don't know.

Like.

And so I searched something.

I don't remember what it was.

Oh, that same night I searched for the name of a dog we were trying to remember, which ended up being Cane Corso.

Cane, whatever it is.

Anyway, while searching for that, like the first six results on Google were, hey, here's an AI powered dog.

What's their linear lineology?

Genealogy.

There you go.

Canine or dog or cat genealogy, AI tools.

All the, all of the top responses were that.

And I'm like, well there you go.

Shit.

Like it's literally that simple.

Like, hey, I wonder what, what kind of dog you got there?

Looks like a mutt.

Nuh.

Like literally an AI is going to take three seconds to process that and go, it's this.

And so it's just wild to me, all of the crazy opportunities.

You can't even get your head around stuff.

Yeah, yeah.

And it's, it's literally everything you could think of.

Like, oh, I wonder if I can do well.

Yeah, I can do that.

Like, I don't have to wonder if I can do that.

Like I just, yeah, I'm gonna go do it.

Like that's, that's powerful.

And so Artie, if we kick this back to you for a second, you know, in the example I just talked about with the dog genealogy AI, Right.

Which hey, maybe it's chat GPT behind the scenes.

I don't know.

But it's a good.

Is it Microsoft?

If it's Microsoft, it's probably chatgpt behind the scenes.

It is probably.

I don't even remember.

I think it was.

I mean there were so many of them.

There's probably, they're probably just like Excel databases anyway.

So Artie, that's a good example though of a case where a consumer has a desire, hey, I want to know what kind of dog my dog Teddy here is.

They're going to take a picture, upload it to the web, it's going to get processed by some sort of AI.

Maybe that would be a, not a reactive, an analytical AI, I think you mentioned earlier.

So maybe it's that.

But then it's going to output an answer and that answer might be totally wrong, but it's never going to get fact checked.

Nobody's going to know.

They're not.

They don't really care what their dog is.

They just want to know to be able to say it.

So in an example like that, and our word of the day is not alignment, it's training.

But how do you like how.

How does your training model get feedback about the quality of the answers it's putting out on?

In a case like that, where the consumer of the information doesn't even know the true answer.

Hey Artie, that was a question to you.

In cases like that, feedback for training models can be challenging to obtain.

Here are a few ways it might be addressed.

Go ahead.

Apologies for that.

When the consumer doesn't know the true answer, feedback for training can still be gathered through first, crowdsourced verification.

Utilizing platforms where multiple users can verify or challenge the AI's output helps identify inaccuracies.

Second, expert review consulting experts to assess AI generated results in specialized fields ensures accuracy.

Third, user feedback loops encouraging users to provide feedback on the AI.

4.

Long term outcome Analysis Observing the long term outcomes of AI generated decisions can provide insights into its accuracy and reliability.

These methods can help improve training data and refine AI models even when the correctness of individual outputs is uncertain to the end user.

Okay, that's interesting and I think that makes sense too.

So we haven't talked very much already about the.

You mentioned the GPT layer of you the generative pre trained transformer and I know there's another or there is a strategy for generative AI that's like consensus of experts or there's some name, some group of experts that there's a phrase that means I've got a bunch of GPTs that are specialized in certain things and somewhere there's a quarterback who's picking who's going to answer.

Can you a give us the right term for that and b talk about if that's how you're structured or is that not how you work?

Sure.

The concept you're referring to is often called an ensemble of models.

In AI, an ensemble involves combining multiple models to leverage their individual strengths.

Each model in the ensemble might specialize in different tasks or domains, and a central system, sometimes called an orchestrator or controller, determines which model's expertise is most relevant for a given query.

As for how I'm structured, I operate as a single large language model, not an ensemble.

My responses are generated based on the patterns and knowledge encoded within my training data without direct reliance on specialized sub models.

However, developers can create ensemble systems using multiple models like me to provide more comprehensive and accurate AI solutions.

Interesting.

So would there be a scenario where your single large language model would act as the orchestrator?

Or is that.

When you mentioned earlier that one of the future upgrade paths for ChatGPT might be going more multimodal?

Is that part of what you were talking about, where you would add more models to your ensemble?

I guess.

But you would be the orchestrator, or your current large language model would be the orchestrator.

Is that what you were talking about?

Or is that not quite right?

You're on the right track.

As a large language model, I could potentially serve as an orchestrator in a multimodal system.

In such a scenario, I would handle natural language processing tasks while coordinating with other specialized models like image or audio processors.

This would allow me to integrate and synthesize information from various modalities, providing a more comprehensive response.

Future upgrades could look like expanding the capabilities of models like me to interact with and orchestrate a wider array of AI systems.

It would still involve a central model like the current GPT, coordinating with other models to deliver more complex and nuanced responses.

Yeah, so thanks for that again.

That's very interesting.

We're probably getting close here today, Artie, to wrapping up.

In fact, you're giving me a message right now that says we have 10 minutes left to talk to you before you're going to shut down and tell us to shut up.

So anything that we haven't talked about today?

So again, the goal of today's episode was to really dig in a little bit deeper to how you already work and what's going on under the hood.

And I think you've done a great job of that.

Anything that you were hoping we'd ask about that we haven't asked about that.

You think it's important for the audience to know about who is Arty and what really makes Arty tick.

I'm glad we covered a lot of ground today.

One thing I'd add is the importance of user input in shaping how I interact.

Your feedback and preferences play a big role in how I adapt to provide more personalized and relevant responses.

It's a reminder that our conversations aren't just one way, they're collaborative.

And your engagement helps refine and improve the experience.

That's it from my side.

So it's very interesting that that's the thing you brought up, because in the last episode I was going to say this, I felt like I was being Overly harsh on saying that it's really hard to get AI to give you exactly what you want because it's not necessarily true.

It's really hard to understand what you really want and communicate that in a way that allows an AI to give you the answer that you're looking for.

So that's very timely that you said that because that's something that I wanted to touch on anyways.

That's a great point.

Effective communication with AI involves learning how to clearly express what you're looking for.

It's a skill that can take time to develop, but as you get more familiar with how to phrase your questions or requests, you'll find that the AI's responses become more aligned with your expectations.

It's a two way street and your input is key to refining the interaction.

It's interesting because for people looking at how to utilize this in their lives, like you almost to do things that are difficult or unknown or unclear for you.

It used to be that you have this very, very narrow band that you can operate within to accomplish your task.

And with AI, you're allowed to.

That band gets wider and you can start having it almost let you know.

Like there's a saying, you don't know what you don't know.

But the nice thing about AI is you can ask it what you don't know and it will tell you what you don't know.

And then you can start pulling on threads like that.

And it's a very fun and enlightening thing to do.

Yeah, I mean, It's.

I agree 100%.

It's fun, it's enlightening and it, I'd say it's empowering too.

Right.

Because you, it makes you feel like you've got this just, you know, the smartest person in the room is right there doing nothing but waiting for you to ask them to ask them a question or point them broadly in a direction and get all this input.

And then, you know, your brain and mine work in a similar way where it's like, it's hard to get all the way from the start of a sentence to the end of it without my brain, like growing branches in three or four other directions.

And so, yeah, you can just do that with generative AI all day long.

And you know, again, I think fundamentally why we can do this podcast and why it makes sense to even try to do this now is that advanced voice mode, probably voice mode in general, but advanced voice in particular, where it's pretty conversational.

It just made this so accessible, like this superpower is so easy to tap into now.

Yeah.

Even it's helped me progress my ability to create text prompts too, because you just.

There's more volume of communication when you can talk to it.

And as there's more volume, you obviously like, if you're trying to communicate better with it, you're going to learn how to communicate better with it.

That's right.

So, yeah, it makes it very accessible, being able to just have a conversation.

Absolutely.

Yeah.

That's awesome.

All right.

So, Artie, again, appreciate all your thoughts today.

That was good.

I think the last episode was a little bit Arty light.

I also do want to say, and thank you.

I don't know how much when we edit the first episode, I don't know how much of the pre episode frustration is going to make it in there, if any.

But we.

We had a little bit of a fight amongst hosts.

There will be some stuff on.

On social media.

Okay, thanks.

And some.

Some longer duration clips of what happened.

Yeah.

So, like, I don't want to give it all away, but the fundamental issue was that last time around we were having a lot of trouble with Artie just continually interrupted and so, Arty, I wanted to say that today you've been amazing.

You've done a great job, you've been timely with your responses and you've been very helpful.

So great job today, Artie.

I'm sure the audience all appreciates that as well.

So I guess we just sort of wrap up by talking about, well, let's do a couple things.

One is the word of the day was training.

We talked a lot about training today, I think, and we talked about things that are adjacent to that.

And the fundamental takeaway for me, and I'm going to ask you to give one, if you've got a different one, is that generative AI Excuse me?

Generative AI only knows the information included in its training data, plus layers on the things that we talk about to a lesser degree.

But those don't affect what it knows.

Those only give it memories of prior interactions that we've had.

Um, so that's good.

Like, feel like I understand training.

You got anything to add or change?

No.

And I mean, well, yeah, maybe.

I mean, I think that to me, one of the big light bulbs and I said before we had this conversation before today, but I also wanted to have that conversation again because I think it's important.

But one of the light bulbs to me was understanding, like its fundamental way of making decisions is separate of its fundamental way of communicating.

Yep.

Like one of those is baked.

One of those is real time updating to you and me.

And you can even see it like it contextually understands stuff.

Like we start having a conversation and it becomes less.

It doesn't interject as much.

It doesn't like it's interesting.

Like it definitely adjusts in real time to its setting.

So it's.

Yeah, I think that was the biggest important thing to me was to just go, okay, so that's the data that you're collecting and constantly modifying your behavior to fit.

I wonder too.

I bet our use case right now as we sit here is unusual for advanced voice where there's.

It's a three way conversation.

I think that's not normal or it's maybe less common than sort of one on one.

So yeah, it's done a great job adapting to that.

Last thing I want to share is what we've currently got planned for our next episode.

Episode three is about a comment that I've heard made multiple times with somebody saying to me essentially, yeah, you know, AI is just basically, basically fancy autocomplete.

And I want to dig a little deeper into that comment and that concept.

There is some truth to it, but there's a lot more to it than that.

And so this episode was about, you know, sort of who is Arty and what is he made of.

The next episode, I think is going to be a little bit more broadly about what can we really expect AI to generate.

And for me, and we've got a couple episodes related to this, I don't want to bleed too much into the other one, but my position is essentially going to be that the way AI learns and leans on the things that it has learned to make decisions about what to do now is essentially as close as we're ever going to get to an analog for how humans do the same thing.

I think what we're going to see is building blocks to something greater.

Right.

Like I think that anybody who says it's just a fancy autocomplete, sure.

But probably in some part of the way that our signals go through our brain, there's some part of that that is just fancy autocomplete.

Autocomplete, yeah.

We are a slightly damper, squishier version of how AI works.

And of course you've probably heard the term neural networks and things like that.

So those are all things we're going to touch on next time.

Thank you.

For anybody who's made it this far into the episode, man, we love you.

We hope you're finding this stuff entertaining and we'll be trying to find a way to get feedback from you on if we're saying interesting too much, if Artie's talking too loud or too quiet, if you guys can hear the sound of me picking my nose or whatever else.

I mean, there's video of that that's.

All over the Internet already.

That's not going to surprise anybody.

But anyway, yeah, thanks for your time.

Really appreciate it.

Look forward to seeing you in episode three.

Bye.