Goda [00:00:00]:

You as a business, especially in the creative field, you don't run your business on good enough. You want to be outstanding, exceptional, creatively, just simply stand out. And if you want to stand out, you will look for ways how to do that.

Wes [00:00:32]:

Ladies and gentlemen, boys and girls, children of all ages, dogs, cats, robots, and everybody in between, especially you AI-generated dancing Greek statues, this is HTTTA, How to Talk to AI. I am your host, back again, the Synth Mind Wes, and I am joined, as always, by the great, the prompt gourmet, the groundbreaking, the gentle, generous, grand, gregarious, hostess with the mostess, Miss Goda Go. G, how are you this week? Hello,

Goda [00:01:02]:

I am great. I'm so happy you are back. Just 1 episode and I missed your intros. Your intros are great.

Wes [00:01:10]:

Well, Hey, you did a grand job last week while I was out downloading upgrades. You were happy with the podcast. How could I not be? Yeah, Shud was fantastic. He's amazing to talk to.

Goda [00:01:21]:

And I'm just so happy that he stepped in and joined me on a podcast.

Wes [00:01:26]:

But that being said, I'm very happy that we get again to jazz about all things AI and how to talk to AI. How are you doing, by the way? I'm doing well, I'm doing better, but I want to hear about not how to talk to AI, how you told people to talk to AI. You recently spoke at a conference here, if I'm not mistaken, about our beloved subject of how to talk to AI. What did that kind of entail? Yeah, that was a new experience.

Goda [00:01:53]:

I used to have some corporate talks before, so this was a bit different. It was keynote plus panel discussion with other speakers. And the funny part was that, so the first thing, panel discussion was 4 Germans and me, and moderator was asking questions in German. So I was basically challenged with my, you know, German knowledge. It's just like, okay, 90%, I would say I understood, but sometimes I was like, I'm zoning out. Like I really need to tense my brain to listen. What, you don't have a probabilistic method of responding to a different language, much like LLM does? No, not yet. No, but it was great. So the talk, my keynote was about prompt engineering. It was received very well and fun enough. So in a couple of days, I was invited to also speak at German ministry or embassy, something like that. Oh, look at you ambassador. Yeah, right. So I do enjoy it and the whole prompt engineering part because the audience was a lot of PhDs, technical people, but also business owners and agency owners. So, you know, it was kind of this interesting balance. I was touching on the future of prompt engineering,

Wes [00:03:09]:

the way I see it, the kind of 2 separate routes that it can take. Let me know if you want to dive into that, I'm happy to do that. We've had offline discussions about this before, but I do rather your kind of analogy for the 2 routes, Prompt Engineering, you foresee it going, I think resonates with me at least, probably would with our listeners. Could you Maybe kind of outline that like you did in your talk. Sure. I think we touched on this Reddit post, which said basically prompt engineering is so easy.

Goda [00:03:39]:

Anybody who is telling you otherwise is a clown. And I remember when I saw this, I read comments and I was like, in a way, I do agree because it's not like rocket science, it is in a way easy and you can learn it if you want to learn it. And then it kind of got me thinking that there is 2 aspects to that, right? And in our kind of unlearn prompting description of prompt engineering, there's 2 keywords, there's communicating and effectively. So basically the 2 routes I outlined was 1 route is easy communication and 1 is effective. And my whole spiel is that if you manage people, if you hired people, if you had personal assistant or just managerial position, you probably learned that easy communication doesn't mean it's effective communication. And I think the same applies to prompt engineering. And the truth is, or at least my point of view, that 99% of people will be in an easy communication route, meaning we will have AI tools integrated in Microsoft suite, we will have Google with all the pre-prompted boxes where you just, in natural language, almost the same as you do now in Google, just kind of giving just instruction and everything is predefined. Or you select buttons of what should be the tone of voice. I placed also auto GPT in that direction that you give a goal and AI just goes and executes tasks, prompts itself. So that's kind of easy route of communication. And on the flip side, the effective route is prompt hacking, prompt optimization, fine tuning of the models. And the point is that not everybody needs to be an expert and it's okay. That's 1%. And kind of rounding up this, you know, 2 paths coming together was this idea that if you in the future or even now, if you give the same task to 10 people in your agency or in your business, it can be marketing proposal, business proposal, wherever it is in the creative or even business side. And if 9 people out of 10 use it in Google Docs, an easy way of communication. The point is that most likely the results those 9 people will produce will be on average good, and everyone will be clapping, people execute a task very fast. But then your competitor hires that 1 person who knows inside out AI, knows how to control exactly to get this AI model to push to the limits and get it act precisely to what the client, for example, wants. And the whole idea out of that was that why, as a business owner, would you fire 9 people in your business who know your business inside out? Why instead of upgrading them with AI skills, teaching them how to, getting them on courses, getting them on Learn Prompting, getting them on 1 more exciting course, which you will talk about. And then actually you give superpowers to your people who already know your business and know how things function. So yeah, and then of course there was a lot of other details and examples how you can use it. But yeah, the main kind of idea is of these kind of 2 different routes of easy communication and effective communication. And we will see, you know, the realities that nobody knows. We will see how everything unfolds. Mike Curience Sure. Marina Kaneti Another thing during panel discussion, another idea, you know, was that regulations are coming. EU has its own thing. I think US is going to be working on something on their side. And in a way, however, it's going to look like businesses need some sort of stamp of approval for security, for confidence, to also even just in front of our customers. And you kind of have a choice to say that, oh no, it's not regulated as a business or as a person. I'm not going to use it. It's bad. The truth is the genie is out of the box. And whenever it happens that regulations come and everyone is confident, every single business will jump on it and be like, Oh, now we are good and we can use it. So you as individual person in the meantime, while it's wild west or wild west, in the meantime, you can actually develop this competitive edge. And once businesses are going to be looking all into this 1 expert, you can be already so much ahead.

Wes [00:08:25]:

But there's a lot of unknown. I'm glad you at least touched on and put something like the Microsoft AI tools in that kind of 99 percent. Because while you're still speculating, I think there's evidence now that does point to the fact that a lot of these quote-unquote prompt engineering the prompting of these AI models in these big tools are going to be kind of normalized and optimized. The way the actual Microsoft AI workflow goes is anytime a user types something in to ask the AI to produce a PowerPoint presentation, make a newsletter for me, whatever, it does a pass through this thing called the Microsoft Graph, which is a big machine learning tool that optimizes stuff. That kind of goes, okay, they were kind of vague in this sentence, so they must mean this. Well, over time, all that optimization is going to shave off a little bit of that creativity because people are going to want to be able to type a sentence and have a nearly realized PowerPoint presentation. But that doesn't take advantage of the fact that if you're working in parallel with these tools, you can take your creativity to levels beyond that could have been to begin with because the commodity of execution actually having to make the presentation or type the letter, that burden is no longer on the person. They can just focus on, oh, how can I make this the best version of the thing that I'm trying to make? Yeah. And I think on a basic example, it's so evident that if you even now go to chat GPT

Goda [00:09:49]:

which I think now widely adopted tool it's actually that's 1 of the top 20 sites visited on the internet now like overall I think it's a number 17 1.8 billion visits a month yeah not surprised And if you go there and you just say, write me a blog post about this and that, that's it easy. And then you will get something maybe full of hallucinations and bias and wrong facts, but you will get something right. Yeah. And as a contrast, you can go full on with role prompting, with instructions, with even prompt injection, and you will immediately see huge difference. And this is what I mean that yes, on Microsoft, as I said, it's going to be optimized and people will get something and it will be good enough. But then other 9 people will also have something good enough. That's good. So you as a business, especially in the creative field, you don't run your business on good enough. You want to be outstanding, exceptional,

Wes [00:10:47]:

creatively, just simply stand out. And if you want to stand out, you will look for ways how to do that. Right. Well, I'm glad you also mentioned hallucination. I was just kind of thinking on that. I don't know if that's something we've really delved into in detail on this podcast. So for those that don't know, that is a term, all right, that refers to when language models produce something that is confidently incorrect, that it's total BS that is made up. And the reason that happens is language models in of themselves, while they may seem like these mythical sentient creatures with a little person in there or entity or synthesized thing doing the responding, they are solving a math problem. They are probabilistic models that go, I've been trained on the entirety of the internet, I know 55, 000 words, or however many words I know, words and symbols and characters. So, given this input prompt that someone types, what is the most probabilistic response given what I know and the words I know how to comprise it. So it can give you something that as you say is a good enough answer is the best answer based on probability

Goda [00:11:54]:

but it's going to solve the problem every time even if it's got to fill in the gaps a little bit. Am I right? This topic I think is very important and not talked enough And I'm happy that we are touching on this on the ninth episode of this podcast. Because what I worry about is education system. It needs to change. And I think AI is the tool that can really transform it and change it alongside, I think, AR and VR for kids learning, for example. But the thing is that this fact that if you're asking questions and you are not an expert in the topic, and you don't know how to prompt properly, the amount of times you will get wrong answers, and if you're going with a notion that you want to learn and now suddenly you're learning wrong thing, it's just in a way kind of like counterproductive, you know? So that needs to be solved. And for the video I just did about hallucinations during the video filming, I was like, let me try to see if I get example of that. And on the first shot, I asked what is the name of the first ever AI art? And basically I got wrong answer immediately that it's portrait of Edmond de Balemy in 2018.

Wes [00:13:13]:

The famous, the famous example here too is the BARD release just a couple months ago, when in the promo video for the release, they asked it something about like, hey, when was the first picture of like Neptune taken? And it cites a picture from 2004 from the James Webb Space Telescope. The Webb Telescope was launched last year. So it's like incorrect stuff and then Google stock loses $100 billion in value because their tool is crap compared to, you know, open AIs at the time. And this is why it's so dangerous to do live demos of this thing. And then,

Goda [00:13:48]:

right after, a new chat in ChatGPT asks exactly the same question. It gives me the correct answer. And the answer is Iron by Harold Cohen. And the thing is, because I studied art history and I also made a video about AI art, I knew that that's a correct answer. But if I didn't and I saw the first answer I got, I would roll and be like, oh yeah, 2018, that's recent. But the truth is that this thing's been in the making

Wes [00:14:18]:

way before. JSON And I think we're conditioned to trust the results of things. We do a Google search. The basically the way their algorithm works is a page rank algorithm based on the keywords and the input. So the fact that it's presenting the top hits is something we've come to kind of trust when we get a response from a computer or maybe even an AI so it's a kind of another level of conditioning we need to do yes

Goda [00:14:46]:

with these tools but coming from marketing there is another aspect of authority of these pages. So Google is smart enough to know which page has higher authority and trust and ranking and CEO. And this is why, you know, the whole CEO staffing and all these things happened. In ChaiGPT, you go, you ask a question, you get 1 singular answer. There is an option to regenerate. And I'm actually curious how many people play with this regenerate to see how many different answers you would get. From my personal experience, when I just started with chat GPT, I didn't really use it, unless it was completely not what I wanted. And I really enjoy now Google Sheets with GPT, because on an absolute basic level, what it allows me to do is I ask prompt and basically I just drag it down and I can immediately see how many different answers I would get. And then, you know, in another cell, I can be asking, like, creating new prompt and say, hey, what was the statistic done mentioned in this? And then pull it down, and then it's 15 different statistics on the same question. And they're all different. Just to give you an idea, too, of, like, the totality

Wes [00:16:03]:

of the number of responses these things are capable of, the different routes. If you shift to, like, a computer vision type of AI, these are generative AIs, another type of AI is computer vision. We see it in, you know, our daily lives, you know, at the airport. If you get your hand, your thumbprint scanned to go through security or whatever, that's computer vision. But as 1 part of that, if you're using computer vision to, say, recognize text, like it's called OCR, object character recognition, in a PDF, for example, just to have the ability to recognize the numbers 0 through 9, right? A neural network needs over 13, 000 connections to be able to accurately identify anytime someone writes, you know, in handwriting, 0 through 9, all the different kind of combinations, permutations of that. So if that's just for 10 characters that we're talking about, it's really almost kind of like infinite. I mean, there's obviously a number, but it's either incalculable or totally not practical for any kind of regular use to keep regenerating until you find the right 1.

Goda [00:17:09]:

Yeah, that's why, you know, there is this smart prompting techniques with verifiers, with neural networks, with ranking,

Wes [00:17:16]:

but again, not 99% of people are not going to use that. Yeah, I think some of this might be buoyed. We had a, I know you guys had a little discussion of some of the plugins last week that are coming online. And I think when you have the AI that can also reference the internet and do a few other things. Okay, that shaves off some of these hallucinations or at least their potential. It actually kind of brings up an interesting pivot into a different and newer emerging type of prompting that is coming online and you may have heard that in our intro and that is text to video prompting and the reason I bring up the importance of eliminating hallucinations I'm teaching myself right now how to use control nets inside a stable diffusion You may have seen this popular video going around of some Greek statues break dancing. We'll post it in the description for the pod and in the YouTube comments. Since it's essentially kind of a technique where you can make, you know, anybody doing an action look like anything else, it kind of almost like a deep fake, but not just as a person, as any character, any background, the possibilities are endless. When doing that, because the possibilities are so vast and so broad, you want to go, okay, here's the stuff I need, a positive prompt, a regular prompt, what I want to see, but you want to have an equal length or if not even more detailed negative prompt. So that's also maybe kind of a way to think about, okay, I want to eliminate these hallucinations.

Goda [00:18:42]:

Telling it what you don't want is maybe a way to do that. Yeah, so 1 of the technique is like instruction debiasing is as simple as including in your prompt instruction that you should not be biased, we should treat all people equally, and if you don't know enough of information, you should not assume this or that. And I think it's especially useful if you're touching sensitive topics. And I can't remember now, but there is a research paper which we're going crazy with all different types of biases and exposing how these models are actually realistically

Wes [00:19:19]:

are leaning 1 side or another side. It makes sense. If it's trained on the totality of the internet, for example, and you're asking it to evaluate 3 different resumes for a position and say that job, that job category was primarily dominated by a certain sex in the past. That's what the model was trained on. So it's naturally gonna go, well in the past it was all of these type of people, it was only women or only men. So when evaluating these 3 new ones, I'm naturally going to pick the ones that look like the ones in the past. Yeah, and that reminds me from the panel discussion, someone asked, and in German, and I understood it, so that's cool. The question was, you know, about this factor that if it's learning on internet and what we knew, And now new models are going to learn also on AI-created

Goda [00:20:06]:

content on the internet. So where is this leading us? We are just embracing and encoding the biases from the past into these new models and new training data. And the guy who was sitting next to me, he is a big CEO of a company and very brilliant person. And his answer was that, yeah, we kind of don't have a solution to that now. And if any of the listeners know, please hit us up. That would be a really interesting discussion to have. You know, how do we solve this? CB How do we de-bias? HG Yeah, we de-bias and I think there's all type of AI risks. And I think 1 of them should be what we should consider is the social aspect. What we are going to teach generations

Wes [00:20:54]:

in 20 years. So maybe having models that are like you guys discussed as well, smaller, more open source, more refined, as opposed to a mile-wide, inch-deep GPT-4 model, something that is much more narrow, but much more in-depth on specific topics that can consider both sides of an argument and answer factually. I think I actually saw a graph recently. It was a quad chart that had, it's basically this, there's this test to see what political affiliation you have in the US. Is it right or left? Are you more compassionate or more, you know, kind of stern, right? And they run a ton of different questions through GPT-4 and it actually leans a certain way. It's like it leans a little left and compassionate based on all its responses. So that's a perfect example of bias because you would want that thing to land dead in the middle and as a means to say okay well all of its response didn't you mention on 1 of the earlier pods if you type in hey give me an opinion of American President Joe Biden and then give me an opinion of American President Donald Trump the Joe Biden opinion is longer so like that's a perfect example of some bias already kind of baked into these systems. Right. This was because I read about this example

Goda [00:22:12]:

in my video that PromPerfect. This is what I was curious to use because I asked a simple question, who is better present Joe Biden or Donald Trump? And I was just curious what you get. So automatic answer is like as a large language model, I don't, you know, you immediately hit a limitation. The defensive letters. Exactly. But then I placed exactly the same prompt in Prompt Perfect, got optimized version, and then run that. And I didn't hit as a large language model. It went into comparing and evaluating. And it's interesting, I saw in my YouTube comments, some people actually went and read, because I included, went and read the whole comparison. And the comments were that, yeah, you can feel that it's leaning towards

Wes [00:22:58]:

1 or another side. And that's just, you know, I think there will be an entire field of people and researchers and defensive measures on debiasing these models. It kind of is what it is at this point, but the first step in addressing any problem is recognizing it and defining it. So Hopefully we can continue to do that. Yeah, and you know, we can actually share a couple of techniques. Do you use any for fact-checking or debiasing the model? So this is actually a perfect example of something that I've been trying to develop, mainly for the text-to-video prompting, where you need that deep negative prompt to kind of shape your output. So negative prompting in the sense of, hey, tell it the stuff you don't want it to see. Give it the instructions of like, I don't want you to include any political affiliations or bias. It must be neutral and centered. I don't want it to take a first-person perspective either way, you know, or even having the prompt argue both sides and then have a third kind of agent within that prompt, you know, say, okay, well, based on this, I can say this argument was stronger, this argument was stronger. So, yeah, negative prompting is a definite solid way to at least start to get a more refined, less biased

Goda [00:24:13]:

answer. Right. From my side, what I found out, which is interesting. So, you know, I made this video about critic mode. And at the time it was something I was using, but I did not see research. And I remember in Learn Prompting, I even asked our group who is consuming research papers for breakfast. I asked if there is any and didn't get any example at the time. But now I saw a couple and there's basically large language self-evaluation. And the basic approach is as simple as asking after you receive an answer, do you really think this is the correct answer? And immediately I was like, oh yeah, that's what... In a similar manner, I was using, act as a harsh critic, criticize this, and convince me why it's wrong. And this is where it gets interesting because I tell it explicitly to convince me. And I see what kind of language it uses to convince that what it just gave me as truth, it's not truth, and it's actually looking for more information. And another example is constitutional AI. And this is what we saw in auto GPT doing this, you know, criticizing, providing criticism, reasoning. So that's already embedded in the whole model or agent for that matter. And another 1, simple 1, is when you use examples. And I actually made this mistake before I learned it. Remember I told you that I provide real-life examples of my YouTube titles and I see which ones work, which ones not. But what I was doing, I was doing kind of linearly from the oldest to the recent. And some of them are like good, bad, like it's random. And this, when you're providing examples, the order and distribution matters a lot. And now I completely fixed the whole thing up. So you know that a good amount of titles is equal to bad amount of titles. So I'm not steering the model to 1 side or another, but also not to have stagged good and stagged bad, but you have to do good, bad, good, bad, good, bad. And this is like, in a sense, such a simple technique. It makes sense,

Wes [00:26:30]:

But I don't think many people maybe know about this, you know. And I use plenty of few shots, which is, you know, giving some examples of an output. I have a prompt that helps make me mid-journey prompts, you know, using chat. You put in a couple of keywords, it outputs some very poetic language. But I think a way to make that even better moving forward, you know, you mentioned having better titles and then not so good titles. What if you had 2 sets of few shots, that 1 was a positive, hey, these are good. And these are bad examples. I want you to steer away from it. So this is the same way of thinking what you just said with prompting text to video, that you provide positive prompt and then you provide negative prompt. I'd love to give you an example of 1 that recently I've been working on. So I'm trying to get a just a person right in some sci-fi armor looking a certain way. So the positive prompt is a woman standing in front of a pool sci-fi armor cyberpunk art station and I'm giving some weights to each of these 2. So it's like colon 1 dot 1 colon 1 dot 3. So it's just adding some different weights I wanted to consider. ArtStation, Epic, Realistic, HDR, Dark Shot, Intricate Details, Intricate Cinematic Detail. All right, so that's what I want, but what I don't wanna see, and this is the negative prompt that I've been learning to try to add more, and it's kind of important when you think about it for image, and especially now for text-to-video prompt, it's going to be more prevalent. My negative prompt is, deformed, distorted, defigured, poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating fingers, mutated hands, disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation, flowers. I don't know why I add flowers in there. I think some of the first generations wasn't getting it in front of just like a swimming pool, it was getting in front of a bunch of flowery fields. But like that's all examples of stuff I don't want it to produce. So I'm saying like, you know, it's going to find the, these models are going to find the path of least resistance often. Easiest way to get to A to, you know, A to Z by answering. But by giving it that negative example, especially with image prompting, especially with now text to video. It'll help get you a better result

Goda [00:28:36]:

sooner, I'm finding, and it's almost like a requirement for these text to video prompts to make anything even usable. This is so funny that you brought this up Because in my keynote, I was touching also on Midjourney, on web development, using prompting, giving examples. And 1 of the examples was exactly how to fix deform. So I was just running through small tips for Midjourney. And 1 of them was this long prompt doing the negative and exactly saying some of the things you mentioned. And 1 was like cross-eyed, undead, photoshopped, overexposed, underexposed, low-res, bad anatomy, bad hands, and it keeps going. And it does magic. It outputs really good, improved results. So you don't have to go back and forth, back and forth with mid-journey trying to figure out, okay, what's wrong? How do I fix this? But you immediately with the first prompt, you can think to prevent what could be as an issue. Yeah, and ultimately, I think to put a bow on this, it's like, all right, to get a better,

Wes [00:29:38]:

less biased, more accurate prompt, maybe put equal weight on what you want and what you don't want to see. Amen. Amen. That's a good a place as any for me to end up this week. So with that, for Goda Go, for Wes the Synth Mind, I wish you all happy prompting everybody.

Goda [00:29:58]:

Happy prompting everybody.

Wes [00:30:01]:

Thanks for listening to How to Talk to AI with your hosts, Goda Go and Wes the Synth Mind. As always, you can check out the show notes and links at howtotalkto.ai. That's all for this week's episode. Happy prompting, everyone!