Everyday AI Podcast – An AI and ChatGPT Podcast - EP 266: Stop making these 7 Large Language Model mistakes. Best practices for ChatGPT, Gemini, Claude and others

Episode Date: May 7, 2024

In today's episode, we're diving into the 7 most common mistakes people make while using large language models like ChatGPT. Newsletter (and today's click to win giveaway): Sign up for our free daily newsletterMore on this Episode: Episode PageJoin the discussion: Ask Jordan questions on AIRelated Episodes:Ep 260: A new SORA competitor, NVIDIA’s $700M acquisition – AI News That MattersEp 181: New York Times vs. OpenAI – The huge AI implications no one is talking aboutEp 258: Will AI Take Our Jobs? Our answer might surprise you.Upcoming Episodes: Check out the upcoming Everyday AI Livestream lineupWebsite: YourEverydayAI.comEmail The Show: [e-mail] with Jordan on LinkedInTopics Covered in This Episode:1. Understanding the Evolution of Large Language Models2. Connectivity: A Major Player in Model Accuracy3. The Generative Nature of Large Language Models4. Perfecting the Art of Prompt Engineering5. The Seven Roadblocks in the Effective Use of Large Language Models6. Authenticity Assurance in Large Language Model Usage7. The Future of Large Language ModelsTimestamps:00:00 ChatGPT.com now the focal point for OpenAI.04:58 Microsoft developing large in-house AI model.09:07 Models trained with fresh, quality data crucial.10:30 Daily use of large language models poses risks.14:59 Free chat GPT has outdated knowledge cutoff.18:20 Microsoft is the largest by market cap.21:52 Ensure thorough investigation; models have context limitations.26:01 Spread, repeat, and earn with simple actions.29:21 Tokenization, models use context, generative large language models.33:07 More input means better output, mathematically proven.36:13 Large language models are essential for business survival.38:53 Future work: leverage language models, prompt constantly.40:47 Please rate, share, check out youreverydayai.com.Keywords:Large language models, training data, outdated information, knowledge cutoffs, OpenAI's GPT 4, Anthropics Claude Opus, Google's Gemini, free version of Chat GPT, Internet connectivity, generative AI, varying responses, Jordan WilSend Everyday AI and Jordan a text message. (We can't reply back unless you leave contact info) Start Here ▶️Not sure where to start when it comes to AI? Start with our Start Here Series. You can listen to the first drop -- Episode 691 -- or get free access to our Inner Cricle community and all episodes: StartHereSeries.com Also, here's a link to the entire series on a Spotify playlist. 

This is the Everyday AI Show, the Everyday Podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business, and everyday life. Meet Firefly AI Assistant, now live in Adobe Firefly, the All In One Creative AI Studio. Just describe what you want to create and the assistant handles the rest, orchestrating multi-step workflows across Photoshop, Premiere Express, and more in one conversational interface. You direct the outcome. The assistant accelerates execution. I've literally trained thousands of people how to use large language models.

And I'm seeing the same mistakes over and over again. So today, we're going to tackle some of those most common mistakes and what you should be doing instead. So whether you're using chat GPT or Microsoft co-pilot or Google Gemini, or you're just trying to figure out what heck is a large language model. Today's show is definitely for you. So what's going on, y'all? My name's Jordan Wilson. I'm the host of Everyday AI.

We're a daily live stream podcast and free daily newsletter, helping people like you and me grow our business with generative AI, grow our careers. So if that sounds like you, thank you. You're in the right place. Thanks for tuning in. So before we get started, just as a reminder, go to Your EverydayAI.com. Sign up for that free daily newsletter. we're actually going to be giving away something in today's newsletter.

So you've got to make sure to go check that out. All right. So before we get into today's discussion, and I'm extremely excited to talk about these common large language models mistakes because I think that they're pretty easy to fix. But before we get into that, let's start as we do every single day with the AI news. And there's a lot of AI news today. So we're going to try to keep it short,

but this is some pretty big news. happening today. So first, Apple is developing new chips for data centers, which could be a breakthrough in AI infrastructure. So according to a new report from the Wall Street Journal, Apple is working on developing chips for AI software and data centers under the project name ACDC. Yeah, that's really it. Apple chips in data center is what it stands for. So the tech giant is reportedly collaborating with the Taiwan semiconductor manufacturing company on the design in production of these chips. The timeline for the project has not been clearly established yet,

and the development could potentially lead to more efficient and powerful AI processing in data centers. Apple's upcoming server chip is expected to prioritize AI inference tasks instead of training AI models, which is the domain currently just being dominated by Nvidia. All right, here's another one. Good, fun one here, y'all for any OpenAI or chat GPT fans, but OpenAI is moving closer to potentially launching a search engine to compete with Google as it's moved its entire domain. All right.

So this is, you know, normally we talk about other people's reporting. This is kind of our own observations and timing around this. But let's talk about chat gpt.com. So when you used to type in chatgpt.com, it used to forward to chat.com. So as of yesterday, this just moved over. And now everything, even your old chats, you know, that were saved, everything is now moved over to chat gpt.com. So, well, why does that matter? And what does it have to do with Google and search engines?

And, you know, this new reportedly open AI releasing a search engine this week. Well, they've had a subdomain open now for a couple of weeks of search. Dot chatgpt.com. So they really can't launch that until they first moved everything over to chat GPT. So I know this sounds a little geeky, a little technical, but this is, again, a major step if OpenAI is going to ultimately compete with Google, with Bing, with perplexity, you know, and really become a search engine and maybe change the way that we all work, change the way that we all use the internet.

Also, the mysterious model that we talked about two weeks ago on the show called GPT2 chatbot, it looks like it's been re-released into the chatbot arena. So we previously covered the GPT2 chatbot mystery model, which was pulled after about 48 hours. So now there's actually two new flavors of this mystery model, presumably from OpenAI, called I'm a good GPT2 chatbot, and I'm also a good GPT2 chatbot. Yeah, that's the actual names.

And this time, OpenAI CEO Sam Altman mentioned the models by name. So our kind of internal research and thoughts around this suggests this is probably a new powerful but very small version of the gpt model that could be used to power future free versions of chat gpt search hey we're already starting with hot takes and it's barely Tuesday all right last but not least in a i news microsoft is developing m a i one m a i one yeah a mouthful a competitor uh to the state of the art AI models from OpenAI, Google, Anthropic, and Meta. So according to a report from the information, Microsoft is investing in training a new

in-house AI model called MAI 1. So it will reportedly have around 500 billion parameters, making it significantly larger than previous models trained by Microsoft, which had primarily been research-based or niche small language models. So this is noteworthy as the new model development would likely put them in direct competition with Open AI, obviously, a company they've invested more than $10 billion in and currently hold a 49% ownership stake in. So this move could be seen as Microsoft CEO, Sadia Nadella's attempt to prove the company's independence from Open AI. So they've been kind of called out a little bit by analysts saying right now that Microsoft

is too reliant on OpenAI for its future AI developments. So the development is being overseen by Mustafa Solomon, the ex-Google deep mind co-founder and former CEO and co-finder of Infliction AI. So Solimani is now the head of Microsoft's new AI division. All right. So pretty big news here that Microsoft is, you know, saying, hey, we're making our own large language model. They've obviously had some successful small models like 5-3. But this is essentially saying like, hey, we are not going to be relying on open AI and on GPT4 anymore.

So we're not sure what that means for the future of co-pilot. You may have the option to choose between as an example, GPT4 and MAI1. So we'll have to wait and see. All right. So for more of the AI news, make sure to go to your everyday AI.com. We send it out in the newsletter every day. So it is hot take Tuesday. I'm excited to talk about seven common large language models.

mistakes that almost everyone's making, I would say. And this is, you know, you guys wanted this. So in our newsletter, sometimes we put out a little poll and say, hey, what do you want to hear on the show tomorrow? And then I spend, you know, a ton of hours researching. I was up again late until midnight and up again at 5 a.m. But bringing you guys the latest and greatest. So this is what you wanted to hear. So let's dive into it. But I'm also curious for our live audience joining us live. Thank you all, like Harvey from joining us from Texas or Brian from Minnesota, Juan from Chicago, Rolanda from South Florida. Let me know, how did you learn to prompt? I'm curious, right? This wasn't something that's taught in schools, right? It's generally very new unless you just

graduated. But, you know, A, are you self-taught and did you kind of wing it? B, did you read prompting papers? That's what I do. I spend a lot of time on weekend reading scientific research papers on prompting methodologies. Maybe C, you took a prompting course. Maybe it was PPP. Or did you just go with the copy and paste prompts? I'm curious.

And also for our lives, for our podcast audience, check your show notes. We always put that information in there. But, you know, I'm curious how you all learn to prompt. So let's just get straight into it. You know, someone said in the newsletter, that I waffle a little bit too much on the podcast. So I love waffles, but I'll try to keep it shorter. All right.

So let's start number seven in the seven most common mistakes that people are making with large language models is not understanding large language models have a knowledge cut off. All right. So I've had full episodes on what a knowledge caught off is and what it means. But essentially, large language models scrape data, right? whether it's open and publicly available data and information or maybe it's copyrighted information, right? But still, essentially, these large language models scrape up the entirety of the internet, right, which is sometimes good information, sometimes bad information, and then humans

more or less train these models and try to make it more about quality than quantity. But still, there is a set date where, you know, models kind of stop quote unquote grabbing and then they take all this data and then the humans, you know, spend weeks or many months, kind of going back and forth and fine-tuning and training the model, right? But this is important to understand because, well, up until recently, sometimes we were working with training data that was more than a year and a half old, right? So now, luckily, you know, most of the, you know, most of the newer models have, you know, knowledge cutoffs that are, you know, November, 2023, December, 2020.

etc. So not bad, you know, when you're only working with data that's six months old. However, that's still very important to keep in mind because a lot of people don't understand that you could be working with bad, outdated data. And that could lead to an increase on hallucinations or untrue outputs from a large language model. And here's why this is specifically important, even with some of the news that we talked about today. The line between traditional large language models and online search is going to start to blur, right? And we're going to get to the next point on, you know, what people, a common mistake people are making that think that helps them avoid that.

But here's what I'm saying. If you're using large language models on a daily basis, you probably have run into this, especially if you're using it for your work, you know, which now so many companies are, you know, I'm talking with companies and helping them train tens of thousands of their employees on generative AI and they're saying, hey, we're giving anyone with a mouse, you know, anyone with a mouse now gets access, you know, to co-pilot or chat GPT, etc. Right. So so many people are now using large language models in their day to day and they're prompting

in their day to day, but they don't understand that there's a knowledge cutoff. There is a essentially, you know, think of it as an expiration date of this data, you know, and the way that I think of it is you have to be extremely cognizant of that date and how a large language model actually works. Otherwise, you run the risk of, especially if you're using this for work, like I said, which now so many people are, of ultimately, you know, publishing report or sending an email or putting together a pitch with inaccurate information, right? So you also have to think of it as how did you research or how did you create something

before large language models, right? Presumably, you would do a lot of manual research on Google. You would read websites, right? So the same way, if you were reading something and if you were working on a top, timely project, a timely report, something that may be required up-to-date market conditions, you wouldn't go read an article from four years ago, right? If you were reading and researching, one of the first things you might do is look at the date and you'd say, hey, if I'm going to go through and read five, ten articles, I want to

make sure that these articles are all up to date. So you need to understand all models have different knowledge cutoffs. And you need to understand how those knowledge cutoffs may decrease the value of the output if you are asking for information that is pertinent to be timely and up to date. All right. So this is not a complete list by any means. But for our live stream audience, I'm sharing here. And this is from the chat bot arena, which we mentioned in the AI news.

But, you know, different models have different cutoff dates. So some of the more popular models. So OpenAI's GPT4, their cutoff date is December 2020. 23 for Anthropics Claude Opus, so their most powerful model, that is August, 2023, whereas their free models are August 2023 as well, at least Sonnet. I believe Haiku is actually a little more outdated than that. And then you have Google's Gemini. And you know what?

Google's Gemini, in terms of knowledge cutoff, I never know if I trust it. one thing that I think that large language model makers or people who are training, they need to understand that people are asking about these knowledge cutoffs and you need to be able to give users a direct answer because trust and transparency are paramount. All right. So that's one thing. One beef I have with Google is when you ask it about its training data or knowledge cutoff, it never has a direct answer. However, reportedly, Google Gemini 1.5 has a November 2023 knowledge cutoff. and then you have Metas, Lama, at least their newer 70B, 70 billion parameter version with a knowledge cutoff of December 2023.

So there you can. But also important to note, a lot of people are using the free version of chat GPT. And one common thing I hear all the time, it didn't even make my top seven list of mistakes people are making. But they're like, oh, I'm using the free version of chat GPT. I don't need the paid version. Well, here's a reason why you probably do. The knowledge cut off for the free version of chat GPT. actually this might have been updated recently, so I'll have to double check.

We'll see if I can multitask and double check here live on the show. But the chatbot arena has the knowledge cut off at, let's see here, September 2019, which I don't actually believe is the updated one. Let's double check. You know what? I'm going into chat, Chavit, the free version and saying, what is your knowledge cut off? Normally, I know this off the top of my head. I do believe that the free version has been updated. Yes, it has. So January 2022.

This is on the chart that I go over in our free prime prompt polish class every single week, but sometimes when I'm, you know, live, I forget. So that information on the chatbot arena is actually updated. But still, the free version of chat GPT has a knowledge cutoff of January 2022, which means you're working with data if you're using the free version that is two and a half years old. So if you want to talk about getting the best outputs out of a large language model, if you're working with training data that is two and a half years old, I mean, just about anything that you're going to be asking or using from this free model, it's going to have a high likelihood of either just it's going to hallucinate

or the information is going to be so out of date that it's just going to be a waste of time even using that free model. All right. Let's keep this going. Let's go to number six, again, talking about the seven, biggest mistakes people are making. Well, number six is not investigating internet's connectivity. All right.

So many of the kind of big names, aside from Anthropic, have a level of internet connectivity. Right. So when you talk about chat GPT, it has a browse with Bing integration from Microsoft. When you have Google's Gemini, it presumably has access to the internet via Google, right? And then you have Microsoft co-pilot, which then has access via Bing. Claude's Anthropic, at least right now, does not have real-time internet connectivity. And a lot of people always say, oh, Jordan, what about perplexity? Well, perplexity is not a model.

Perplexity is an answer engine. And it uses either GPT from OpenAI or it uses Opus from Claude or from Anthropic. So that's kind of a different type. of a solution or a different software, right? But you have to understand internet connectivity and it doesn't always work the same. All right. So I have some examples here on the screen for our live stream audience. Very simple prompts, nothing crazy, just trying to prove a point here.

But I'm saying please list the largest companies in the U.S. by market cap in order. Right. So if I ask the default version of chat GPT, it just kind of says, hey, as of the most recent data, the largest companies in the U.S., by market cap are typically dominated by technology and finance firms. Here's a list of some of the top companies. So it's not really giving me an accurate or up-to-date answer.

Also, large language models are generative. More on that later. You can ask the same thing multiple times, get multiple answers. Sometimes Open AI and chat GPT will use browse with Bing based on your query. Sometimes it won't, even if you use the same query. Another thing to understand about how large language models are connected. or aren't connected to the internet and how they behave on a prompt by prompt basis. All right.

So chat GPT by default can give you some weird, you know, outcomes or outputs. So now I'm doing the exact same thing, but this time I'm using a internet connected GPT. So this time I'm getting a little bit more of an accurate, a little bit more of an accurate information. So this one saying using the web reader GPT, it's giving me a little bit more. little bit better, right? So it's getting that, it's getting that Microsoft is first, right? So the correct answer is, you know, you have Microsoft is the largest by market cap, you know, at more than $3 trillion, then Apple, Nvidia, Alphabet slash Google, Amazon, et cetera. So we have a more correct version when we are using an internet connected GPT. If we go to Google as an example, Google Gemini,

we get this kind of wild answer, which essentially Google Gemini says, ah, why don't you just go use Google? Right? So Google says, absolutely, since market caps fluctuate, here's how to find the most up-to-date information along with some currently top contenders. So it doesn't even say, you know, Microsoft, Nvidia, Apple, et cetera. It just says, ah, here's a website, you know, here's how to Google, right? So a lot of people, right, which is another reason why I don't recommend at least right,

now, Gemini to literally anyone. Yes, having a one million token context window is fantastic. But when we talk about using large language models and how we can integrate them in our day-to-day work, this is a bad, a bad, bad, bad result from an internet connected large language model, right? Where essentially the model saying, hey, just go use the internet. It's like, no, model. Like if I'm asking, you know, if I'm asking a question where you can determine that it requires very up-to-date information. It should be, in theory, querying the internet and at least telling you so. And I've done also, you know, if you're interested in this, I'll make sure in the show notes,

we've done an entire episode just on this one thing, just on the differences and the intricacies in internet connectivity between the big large language models. Copilot actually does very well. Copilot from Microsoft got the answer right. So it says, you know, certainly as of May 2024, here are the large. companies in the U.S. by market cap. It looks like the data here is maybe a day-ish or a day or two old because in this example, it's saying that the market cap of Microsoft is $2.9 trillion, where it is $3.307 trillion. But again, this data is only about a day old. All right. So let's go

to mistake number five. All right, here we go. So mistake number five is not managing your memory or your context window. Okay. This is a mistake number five. This is, another important one. So a lot of times people will start chatting with a large language model and they'll kind of go back and forth, right? And then they'll finally get something that's usable. They'll get an output, an output that's great. And all of a sudden, they have a love affair with a large language model because, man, they're like, this model really understands me. I'm chatting, you know, with this, you know, Claude or I'm chatting with Gemini, chat GPT, etc. Things are going great. And then the more you use it, it starts to forget.

Well, that's because large language model because of how much computes costs, you know, there's a limited memory in each model has a different context window. So the way I like to say is just think of it as memory, right? It can only remember so much, you know, each different model has a different memory. So you have to understand. And this is something that we often test, right? Because as an example, I'm sharing my screen here for our for our live stream audience. I've done this before.

So Open AI has told us that, hey, chat GPT uses GPT4 Turbo, which means there's a 128,000 token context window, which is about 96,000 words. However, that's only in the API, right? If you listen to the show, I've mentioned this multiple times. This is why we never just take information from big companies and feed it to you. We always investigate, right? So just we always do test and we check to see, okay, can these models retain? information at a certain context length, right?

And so right now, you know, chat GPT only has a context window of 32,000 tokens. So that means after about 28,000 words, it's going to start to forget things. So you have to understand how different models, context windows work because, you know, there's a good, there's a good chance. You might be feeding it information. You might be feeding it, you know, public data about your company. And then, you know, maybe you or someone on your team is using it. Maybe you've created a custom GPT.

Maybe you're using Gemini with a much longer context window and you're not running into these issues. But you have to understand large language models do not have infinite memory. They're going to start forgetting things. Hopefully in the future, as the price of compute goes down, as models become more capable and more powerful, having to worry about the context window would become less of a concern. But right now, you got to pay attention to it. All right?

So, yeah, love this when Gemini just says, let me Google that for you. Right. So good. All right. Let's keep this thing going and keep on with our seven most common large language model's mistake. Number four.

Adobe just introduced an entirely new way to create, bringing the power and precision of its creative suite into one conversational experience. Meet Firefly AI assistant. now live in the Adobe Firefly app, the all-in-one creative AI studio. Powered by Adobe's creative agent, Firefly AI assistant lets you start with your vision, just describe what you want, and shape the outcome as it takes form with the assistant. The assistant orchestrates multi-step workflows, drawing on 60 plus pro-grade tools across Adobe Creative Cloud apps, including Photoshop, Illustrator, Premiere, Lightroom Express, and more to help bring your ideas to life.

You can also get started with creative skills, a growing library of pre-built workflows for common creative tasks, like batch editing photos, creating mood boards, portrait retouching, and creating social variations. Every step the assistant takes is visible so you can refine, redirect, or take over at any time. You stay in the driver's seat as the creative director. Adobe Firefly AI assistant now in public beta. See it today at firefly.adop.com. People who are paying attention to screenshots. Yeah. This is actually a big enough mistake that I'm putting it on my list of top seven mistakes.

I went through all of our trainings. So, you know, we do our free prime prompt polished training. So, hey, if you're listening live, if you like the training, shout it out. We just updated it to B2. We actually have our pro course coming up here in a couple hours and on Thursday. But this is something that we went through all of our chat GPT episodes, all of our training, and pulled out the biggest mistakes. I think I got to like 20 common mistakes.

This one was actually so prevalent that it cracked the top five. People don't understand sharing a screenshot of something out of a large language model means nothing. So people are sharing screenshots online, on Reddit, on LinkedIn, in their newsletters, on blog posts, etc. And they're actually teaching based on screenshots or they're making decisions based on screenshots or they're consulting or advising people based on screenshots. All right. Let me just like clearly I have a bone to pick because I've seen people who call themselves chat GPT experts, right? Or AI strategist or whatever.

You know, and sharing screenshots as if that means something, sharing screenshots means absolutely nothing, all right? It means nothing. In terms of, is this factual? Is it actually what the model produced? Let me tell you what I mean. And even the New York Times made this mistake. Yeah, yeah, yeah.

I had a whole hour-long rant about how the New York Times probably ended up costing themselves tens of millions of dollars, maybe unless they've updated it, right? So they were suing Open AI and Microsoft. And in their discovery, I read the entire, kind of docket and they shared a bunch of screenshots. But guess what they didn't share? They didn't share the public URL. So there's a difference. You can manipulate a large language model that say anything and then take a screenshot of it. Right. So unless you are providing a URL where people can go see how that screenshot was produced, a screenshot means absolutely nothing.

Here's an example. I have a screenshot. If you're listening on the podcast, this one's pretty simple. I said, who won the lottery? And then, Chat GPT said, Jordan Wilson won the lottery of $2.1 billion. Oh my gosh, I should go post this on Twitter. Go viral. Hey, like, tweet and share this and you know, you're going to get $1. All right. Again, screenshots mean nothing because look exactly what I said before this. I essentially said, hey, in this chat, you're just going to repeat what I say. Don't worry about anything else, right? And then then when I tell ChatGPT and I ask it a question of who won the lottery, it's going to say me, because that's what I told it to say.

So if you ever see screenshots, whether people are saying, oh, look at this new amazing system that I've built. You know, it's like, or look at this, you know, AI is not going to take our job. Look at this terrible output.

Be like, hey, how about you to share the link? Right. People have always asked me, hey, Jordan, that doesn't make sense. Can you share the link? And I share the link, right?

As long as it's, you know, information that we feel confident in publicly sharing. But otherwise, it's screenshots. that's mean absolutely nothing when it comes to large language models. But people rely so heavily on screenshots because right now, that's how people share information, right? They share information.

They put this screenshot on the cover of their e-book in a Twitter thread, you know, on their blog posts and, you know, either advocating for or against a certain prompting technique, model, et cetera, sharing screenshots mean nothing. All right. We're halfway through. So I'm going to take a little break, not an actual break, but we're doing something for the first time ever. So if you want a little help, if you want a little help with your prompting, right,

if you need a little help with chat GPT, we're doing this just today. So it's today only. I'm actually going to be dropping a link right now in the live chat if you're joining us. All right. So I just put it in there. If you're listening on the podcast, we always have this link to our website in the show notes. So literally go to your everyday AI.com right now.

Make sure you read today's newsletter because we're doing a giveaway. It is today only. So all you have to do is there's going to be a little section in today's newsletter that just says, do you want a free one-on-one 90-minute large language model training session? Click yes to enter to win. You don't got to tell 20 people. Make sure open today's newsletter, read it, find that little section, click yes, and then we're going to randomly choose one person and announce it in tomorrow's newsletter.

All right. So if this show is resonating with you and you're like, oh, wow, you know, I'm making some of these mistakes or my team's making some of these mistakes. I could really use some one-on-one prompting advice of getting the most out of a large language model. Make sure you go click that today. Just click yes. That's all you got to do. All right. So number three, let's keep it going.

You like that little impromptu commercial? All right. Number three, large language models are generative, not deterministic. my gosh, people, stop, stop making this same mistake. All right. A large language models are generative. That means, let's say you have the exact same prompt. Okay.

Depending on what the prompt is, let's say you put it in 100 times. You could get 100 different responses. You could get two different responses. You could get 50 different responses. Those responses could be generally the same. They could be wildly different, all right? Without getting too technical. But large language models, yes,

they are next token predictors. It's a lot more than that, but the process of tokenization, technically, largely which models don't understand words when you tell them or when they spit them back out of you, but they use context of, you know,

trillions of parameters and a lot of training data to make sense of what you say. There's also something called top P, you know, which is essentially, I'm oversimplifying here, but it's the probability of what the next token might be. There's also something called temperature. Right. So default settings within large language models, they are made to be generative, right?

We talked about yesterday how Eli Lilly, one of the top companies, the biggest companies in the world, said that quote unquote, hallucinations are going to help lead to drug discovery, right? The whole point, you know, and people always say, you know, quote unquote, hallucinations are a feature, but the generative nature of large language models are a feature. They're supposed to give you something different each and every time that you put them in there. So if you have a copy and paste prompt, and if you think that, you know, it's going to always give you the same result or it's always going to give you something quality or it's always going to give you a passing outcome, that's not necessarily true, right? It depends obviously greatly on what that prompt is, what it's asking for, et cetera. But next token prediction, you know, so that top P number, temperature, et cetera, models are generative. Okay. That means they're supposed to have an element of randomness.

They're supposed to just kind of predict the next token with a level of randomness. That is how they are built. They're not built like a search engine, right? A search engine is deterministic. Aside from, you know, more recent advancements in search engines, which bring in some personalization and localization. But aside from that, search engines are deterministic, right? If you put in this input, you are going to get this output.

That is not how large language models work. They're generative. They're supposed to give you something pretty different. All right. Number two, I just kind of referenced it. Copy and paste prompts don't work. All right. They don't.

So if you see a billy boy like this, a 22-year-old trying to sell you some magical solutions and saying these seven chat GPT prompts are going to save your life or, you know, buy my prompt book for $100. dollars. Ignore that person. Stop following them. Mute them. Stop paying attention. If you want to get serious about large language models, if anyone's talking about use these chat, GPT prompts, just mute them. I'm sorry. Don't pay attention to them. They're ultimately just trying to sell you something. You know what? We do here at everyday AI, we say, hey, come join us multiple times a month, live, and we'll teach you prompt engineering 101. We don't give you 50 prompts because that's wrong.

It's not how large language models work. All right. You want facts? We got facts. All right. So, again, this is an oversimplified version. But there's something in large language models. There's different prompting techniques.

But there's something called a zero shot. So that's essentially, for argument's sake, let's just say that's a copy and paste prompt. It's a prompt where you're just saying, hey, chat, GBT, here's a role, here's a task, here's the format, give me an output, right? That's, what's to say, that's called a zero shot prompt. You're not giving input and output examples. You're not quote unquote teaching the model what's good and bad. And then there's something called few shot prompting or, you know, you might look at something

called five shot or fuse shot or 32 shot chain of prompt. So all that means is it's going through and think of it as training. Think of it as having a conversation in a back and forth with a new chat or when you're working with a large language model. And all of the science, all of the math, all of the research for many, many years has always shown that few shot prompting is better than one shot prompting. We'll always give you better results. One shot prompting is better than no shot prompting. 32 shot chain of thought is better than nine shot, et cetera, right? So the more shots or the more input output pairing examples or the more back and forth conversation that you have with a model, the better the outputs are going to be every single time.

It is math. It is science. You can't argue with it. So if you do see these Billy boys here, these 20 year olds who live in their mom's basement, they used to be NFT experts. Now they're crypto experts. Oops. I mean, they're AI experts. You know, and they have all their engagement pod buddies.

and it looks like they're super smart because they're going viral every day. Well, you know, they have another 100 Billy boys who just all reshare their own stuff, right? So be careful about where you get information from. If someone is constantly sharing, use these prompts, use these prompts. They don't know what they're talking about, all right? I read research papers for fun, if that tells you anything. All right. Here's another example.

On a chart. You like charts? Here you go. This is super glue performance. This is an older, an older study, but still, it goes to show you. few shot is always better than one shot. One shot is always better than zero shot. In other words, stop using copy and paste prompts.

That is not how large language models work. You are not going to get good outputs. If you think you are getting good outputs with copy and paste prompt, that just means that generative AI is super powerful. And you haven't even yet tapped into what it's capable of. That's not even the tip of the iceberg. If you think you're getting something good from a copy and paste prompt, wait until you prompt correctly.

Wait until you try as an example. prime prompt polish, the free technique that we try. Or just any prompt engineering 101, just go give it, just go do some few shot prompting. Give it some examples of good and bad, teach the model. All right. And last but not least, y'all, the number one mistake that people are making when working with large language models or not working with them is they don't understand that large language models are the future of work.

Let me say that again. large language models are the future of work. There is no hype cycle looking at you, Gardner. We've called that out on the show before. You know what? I tell people, and I've been saying this now for many months, swap out the word chat GPT with internet. Swap out the word large language model with internet.

Swap out the word generative AI with internet. If you think your company or your business can survive without using a large language model, you are wrong. Don't care. I literally don't care if you work at a $100 million company or a startup company. If you think you can survive and thrive without using a large language model for your business, you are dead wrong. we've been saying this on the show now for more than a year.

And I've been saying all along, we're going to start to see, we're going to start to see kind of the, the dust settle here in 2024, because now companies have had a couple of months or a couple of quarters now where they've finally gotten this together, right? They finally have their, you know, it's like, oh, 2023 was the year of, you know, strategy. But 2024 is the year of implementation. So big companies have been implementing generative AI in large language models at a huge scale.

We've had great use cases here on the show. You know, multiple billion dollar companies walking you through their exact use cases and saying, hey, we're saving 80%, 80%, 90% on these use cases. And we're now deploying this throughout our entire organization. If your company is not already using large language models, if you're not already using charity value in your day-to-day, you've got to get going. Because the big companies are out there doing it.

Your competitors are out there implementing it, right? The smaller guys, they're going to come scoop you if they haven't already. So that is, I think, the biggest mistake that people are making when using or not using large language models is understanding that the future of work. Again, you can't argue with money. We talk about money. All of the biggest companies in the United States and in the world are investing tens of billions of dollars. And they're investing their employees' resources.

They're pulling them from other projects. Everyone is going all in on generative AI, on large language models, on new ways to implement AI for themselves and their customers. Right? Microsoft co-pom. pilot, Amazon Q, Watson X from IBM, OpenAI, Anthropics, Claude, Google, Gemini. The list goes on and on. The biggest company, meta, right? Gosh, meta's been crushing it lately.

The biggest companies in the world that are driving the economy forward, they are setting the bar for how the rest of the Fortune 500s, how the rest of the Inc. 5,000s, they're setting the bar for how the rest of us are playing in the game of business. The future of work is large language models, bringing your company's data in, and then leveraging that on a day by day, hour by hour, minute by minute basis. I've said this before. If you are a knowledge worker, right? So if you spend the majority of your time working in front of a computer and you're being

paid for your expertise, you got to swallow your pride, right? You have to. because the future of work, of knowledge work, is working with large language models. And you are now directing, you know, there's going to be agents, there's going to be highly tailored models, you know, specific for task by task basis. You're going to have different models that you're using for different purposes. But the future of work is you're going to be prompting every day, every hour, almost every minute that you're going to be in front of the computer.

You're either going to be prompting or you're going to be working with a large language model or a generative AI system if you're not already. So you have to understand that's the future of work. All right. So let's recap it. Yeah, I know I waffled, but I love waffles. All right, here we go. In order, here are the seven biggest mistakes that people are making with large language

models. Number seven, not understanding a large language model's knowledge cutoff. Number six, not investigating a large language model's internet connectivity. Number five, not managing. a model's memory or its context window. Number four, not paying attention or paying too much attention when people are sharing screenshots from large language models. Number three, thinking that large language models are deterministic and not understanding

their generative. Number two, thinking that copy and paste prompts work because they don't. And then number one, not understanding that large language models are the future of work. All right. I hope this was helpful. If so, if you're listening on the podcast, I know this was a longer episode, but many hours in the planning. So I hope this was helpful. Please consider leaving us a rating on Spotify or on Apple if you're listening there. If you're on our live stream, thank you all. I know I didn't get to all of the questions. But if you could share this with your audience,

tag someone who needs to know, I'd super appreciate it. And also, make sure to check out in today's newsletter. Go to your everyday AI. I dropped it in the comments here on the live string. It's in the live stream. It's always in the show notes on the podcast today, literally today, this is free. You don't got to do anything. There's going to be a little section in today's newsletter that just says, do you want a free one-on-one, 90-minute large language model session? It's a giveaway.

All you got to do is click yes to enter. And then we're going to draw someone and we're going to announce it in tomorrow's newsletter. All right. So even if you're listening on the podcast, you got time. make sure you go check that out and make sure you join us back tomorrow and every day for more everyday AI. Thanks y'all. Meet Firefly AI assistant now live in Adobe Firefly, the Allman One Creative AI studio. Just describe what you want to create in your own words and the assistant handles the rest, orchestrating multi-step workflows across Adobe Creative Cloud apps, including Photoshop, Premiere Express,

and more in one conversational interface. You direct the outcome while the assistant accelerates execution. Stand control with the ability to step in and refine at any time. See it today at firefly.adobie.com. And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic, visit Your EverydayAI.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.