Everyday AI Podcast – An AI and ChatGPT Podcast - EP 274: 7 things you need to know about GPT-4o and what OpenAI isn’t telling the truth about

Episode Date: May 16, 2024

Win a free year of ChatGPT or other prizes! Find out out.Why you need to listen to today's episode about GPT-4O 👇  ↳ Everyone's talking about GPT-4o but no one's actually investigating it. ↳ No, not everything is free. ↳ Yes, OpenAI is still not telling the truth about some stuff. ↳ Yes, we did all the homework for you so you just gotta show up and learn  7 things you need to know about GPT-4o and what OpenAI isn’t telling the truth about Newsletter: Sign up for our free daily newsletterMore on this Episode: Episode PageJoin the discussion: Ask Jordan questions about GPT-4O on LinkedInRelated Episodes: OpenAI Releases GPT-4o: 12 things you need to knowEp 272: How OpenAI’s Shot at Google (and GPT-4o model) Will Change How We All WorkTopics Covered in This Episode:1. Changes in OpenAI's language models2. Interaction with ChatGPT and Browse with Bing3. Detailed analysis of OpenAI's GPT-4.0 model4. Announcement and analysis of OpenAI's GPT 4 o5. Potential developments for ChatGPT Plus6. Discussion of AI influencers and misinformation on the internet7. Close-up on specific features of GPT 4.0 and device restrictionsTimestamps:00:00 OpenAI releases GPT-4 o, lots of mystery.06:41 Google's announcement felt scripted, OpenAI's felt real.09:55 Live omnivision assistant can see, understand, talk.11:10 ChatGPT desktop agent not available for free.15:04 "Highly advanced capabilities of DALL E showcased."19:36 Chatbot arena allows blind model testing, benchmarking.22:52 GPTplus chat paid version getting update.23:28 Expect updates in weeks, bugs must be fixed.27:45 UI UX changes and new drop-down feature.31:52 Using large language models requires understanding context.35:12 OpenAI misunderstanding about model size and usage.38:33 Negative experience with sporadic Browse with Bing.39:25 GPT 4 to query internet using Bing.45:21 OpenAI sees potential in language model upgrade.47:36 GPT-4 recap: features, models, UI updates.49:28 Connect with experts in generative AI space.Keywords:OpenAI, GPT 4, knowledge cutoff, UI/UX update, ChatGPT, context window size, Jordan Wilson, Browse with Bing, large language models, targeted Browse with Bing calls, Internet-connected GPTs, free and paid versions, transcript excerpt, 3D files, advanced data analysis, GPT-4.0 model, chatbot arena, foundation models, chat GPT model, model's memory, AI news, AI robSend 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 and 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. Open AI just released its new model, GPT40.

The O stands for Omni or Omni model as Open AI completely changed everything under the hood with how its base model works. But the O could stand for obscurity because there's so much unknown, so much misinformation out there from random people on the internet, and a lot hidden about what's real and what's not with this new announcement. and the new model. And there's one or two small little things that Open AI just isn't telling us the truth about. So we spent way too much time investigating this new update from Open AI so you didn't have to waste hours. And today we're giving you the seven things you need to know about the new GPT40 model. What's going on, y'all? My name's Jordan Wilson. I'm the host of Everyday AI. We do this every dang day coming to you live 730 a.m. Central Standard Time with our live stream,

podcast and free daily newsletter, helping everyday people like you and me, learn and leverage generative AI. It's a lot to keep up with. So, you know, if you're listening on the podcast, thank you so much for your support. Please make sure to leave us a rating, you know, comment. Let us know, reach out. Check your show notes as well. There's always ways you can reach out to myself. And if you're joining us on the live stream, appreciate having you. As always, we got a big bogey face here on YouTube. Thanks for joining in. Michael and Brian and Woozze. and Christopher, some of our normal, some new faces, love to see it. So before we get into this new model and seven things that I think you really need to know about it,

let's first start as we do every single day with going over the AI news. And like I said, if you're listening to on the podcast or if you are joining us live, make sure to check out those show notes as we're going to put in a lot more information, including our website at your everyday AI.com. All right, so let's talk about what's going on and what you need to pay attention to in the world of AI news. So, an AI robot just gave a graduation speech, yes, at a US university. So an AI robot named Sophia gave a commencement address at DeUville University in Buffalo, New York, sharing generic advice and drawing mixed reactions from students and the public.

So the AI robot named Sophia, Sophia's address, focus on, common themes in graduation speeches, such as embracing lifelong learning and believing in yourself. Some students felt that having a robot as a commencement speaker felt impersonal, especially after experiencing virtual high school graduations due to the pandemic. The university received backlash for their decision to have a robot speaker, but ultimately stood by their choice to showcase the potential of technology. All right, our next piece of AI news, Stability AI is reportedly losing too much money and is looking to sell.

So according to a new report from the information British AI startup stability AI is facing financial difficulties and has discussed potential sale options with a buyer. This follows a restructuring process and staff layoffs along with the resignation a couple months ago of the company's CEO. So according to the report, Stability AI has generated less than $5 million in revenue and lost $30 million in the first quarter of 2024. So I'm not great at math or financials, but that doesn't seem good. Also, reportedly the company owns, owes close to $100 million in outstanding bills to service providers. So that's a lot. But, you know, stability AI may still have potential for success with its AI models for generating audio and video using text prompts.

So someone might scoop them up if you remember my predictions in 2023. I did say one of these companies was going to get acquired by a bigger company. So let's see if that holds true. All right. Last but not least in AI news. the U.S. and China finally met on AI. So U.S. and Chinese officials recently held closed door talks on the risks and management of AI technology, highlighting the tension between the two countries. Both sides express concerns and hopes about the potential of AI, but also acknowledge risks in areas such as surveillance and national security. So there's obviously two very different approaches with AI with these two companies. U.S. officials raise concerns. about China's misuse of AI, while China rebuked the U.S. for restrictions and pressure in the field of AI. So the talks were the first of its kind after a meeting in November between the

country's two presidents. China is advocating for the U.N., though, to take a leading role in global governance of AI, something that could sideline the U.S., so the U.S. is obviously against that. All right. There's obviously a whole lot more that we didn't have time to cover. That's why we put out a daily newsletter written by me, a human. So make sure if you haven't already, go to your everyday AI.com, sign up for that free daily newsletter. We'll have a lot more on that news and more. All right.

So let's get straight into it, y'all. Let's talk about the seven things that you need to know about GPD40 and what opening AI isn't really telling the truth about. All right. So if you're joining us live, I'd love to hear your thoughts. Get your comments in. We'll probably feature some of our favorite comments in the newsletter today. So if you're joining us live, appreciate it as always, like Tara from Nashville or Fred from Chicago, Juan from Chicago.

Chicago, we're rolling deep today like myself. Seri joining us from the UK. Appreciate that. And Joe from Asheville. So get your questions and comments in. But let's just let's jump straight into it, y'all. So we're going to do a backwards countdown here. So number seven, hardly any of the GPT40 announced features are live yet and most won't be free.

Yes, that is important to know. And also, this is as of today. Well, actually as of this morning, right? So May 16th at 7.30 a.m. Central Standard Time. Because here's the thing. Obviously, these things change quickly, right? But there's just so much bad information out there. And I think Open AI was a little ambiguous with some of the way that it announced things.

And a lot of people are asking, like, my thoughts between, hey, how did, you know, what were your thoughts between how Open AI announced everything versus how Google. you know, announce everything at their I. I think open AI did a great job in their announcement. So don't get me wrong. I'm not here to bash on Open AI. You know, Google's announcement felt extremely scripted. Like it hurt watching it, whereas Open AI felt very real and natural,

which I think is important when you're talking about AI, you know, to not actually sound like a robot, right, to not sound like that Sophia robot, giving a graduation speech at DeUville University, right? So so many of these features, though, are not out and they are not going to be free. All right. So let me just go ahead and break this down. And if anyone out here has taken our free prime prompt polished PPP course that we update

literally multiple times a week, and we're doing that again here in a couple hours. So if you want to access, just type in PPP. I'll make sure to send you the link. But there's a lot of misinformation, okay? Because here's the reality. Open AI said that GPT40 is going to be free to all users. All right. Cool. Great news. It's not right now. All right. So people, people were very confused. And I'm actually, I did check this like an hour ago, but I'm going to check it again literally right now. But yeah,

so it is still not available to free users yet. However, pretty soon, all free users will have access to GPT40. Okay. Presumably it will be logged in users because OpenAI did just change this, about a month ago where you don't even have to be logged in to use its service. I would guess that you would at least have to have a free account to use the new GPT 4-0 model. All right. So there's a lot of, lot, of lot of people, even, you know, quote, smart on, you know, quote, unquote, AI influencers on the internet who are putting out the wrong information, okay?

Because here's the thing. Everyone's like, oh, you know, you don't need to have a paid chat GPT account anymore, right? because there's obviously the free plan that costs $0.00 and zero cents. And the chat GPT Plus plan that costs $20 a month. So everyone's saying, well, hey, if the GPT40 model is the newest and greatest model and it's free, everyone, let's cancel our subscriptions. So here's something that I think Open AI definitely dropped the ball on in their communication, right?

That's the pros and the cons between, you know, a startup versus having, you know, polished veterans deliver your address. But I think that they drop the ball here by saying, hey, there's so many of these features that we are demoing today and that really, I think, were the show-stopping features that are not going to be available on a free plan. So just because that people are going to have free access to GPT-4-O and the O, like I started the show-off, it stands for Omni or Omni Model, that doesn't mean you have access to all of the show-stopping features, right? So I have a little graph here on the screen. I'm not going to go over at point by point.

You know, so if you're on our live stream audience, you can go ahead and read that. But a couple of the things that I want to, you know, point out or draw attention to is probably what I would say is the two biggest features, right? So one is kind of this live AI assistant or live, you know, AI agent. People are calling it her. I, like, I'm calling it the live Omnivision, right? So this is the feature that they demoed in almost all of their demos, but it is when you are talking to the app, presumably just right now on the iPhone. It can see and hear and process and talk to you in real time.

So the demo, and we already showed this on the show, but was a student doing math, right, math homework. And it could see it live, right? So this new live Omni assistant could see it live, talk to it. So the student could say, hey, you know, I'm trying to keep. You know, angle A and angle B here, is this right? Right. So we can see, understand, and talk live. Free users do not have access to that, or at least they reportedly don't, right?

Also, the other big feature, I think, is the desktop assistant, which is huge. It's actually out. People don't know this. There's a little sneaky way that you can install it. You know, not everyone has access to it right now, but there is a, there's, There's a way you can install it. Maybe, I don't know, should I? It's one of those random links on the internet, on the internet.

Live stream audience, you guys think I should share that in the, in the newsletter today? Let me know. But the desktop agent will not be available to free users. All right. So I'd say two of the biggest reasons why people would want to use chat GPT would be those two things, right? This live kind of agent or live assistant, you know, that can see here, process.

and you can communicate with in real time, not available to free users, and the chat GPT desktop app, right, which I think is amazing. So in one click or one keyboard stroke, you can literally share anything on your screen and talk is essentially an overlay on top of whatever you're working on and say, hey, you know, you can talk or just click a keyboard command and say, hey, what's this? I need help with this. Help me finish this code. Help me finish this email, right?

So all things you could technically do already in chat, GPT, but it saves you a lot of time and being able to talk to the assistant is something, I guess, we didn't have before. So those are, y'all, those are the two headlining features. That changes the future of work. We had an entire episode about that, right? So those are the things that I think people are missing. And they're like, ah, you know, just cancel my chat GPT subscription. It's like, all right, well, also do that at your own risk.

Because the last time that Open AI had a huge rollout like this was, in November of 2023. And guess what? Obviously, once they released it, their servers crashed. They couldn't keep the service up. So guess what happened? They stopped allowing new paid users, right? So, hey, if you want to go ahead and cancel your subscription, go ahead.

But there's a chance that, well, I would actually find it very likely that this is going to crash once it's rolled out to everyone. And they may not take any more paid users for a couple of weeks or a month or two. So, you know, if you want to cancel, do so at your own risk. All right. Yes, Juan, good question. Right now, the app and so the kind of smartphone AI assistant is only for iPhone right now. And the Mac app is only, or sorry, the chat GPT app right now is only for Mac. Again, that could change, but that was just what Open AI said in the announcement. And it makes sense. Well, it makes sense and it doesn't, right? Because obviously Microsoft, maker of Windows, right, has reportedly. invested $10 billion to $13 billion in Open AI and has a 49% equity stakes. You'd think, oh, Open AI is going to make it accessible for all Windows devices. Not now. You know, reportedly, Open AI is locked in to provide generative AI capabilities with Apple for, you know,

Apple's new updates that they'll be announcing at their worldwide developer conference here in about three weeks. So, yeah, at least right now, it's going to be Mac only or Apple only. All right. So yeah, Brian, Brian says no, no canceling here. Brian also, Brian's ahead of me. He says it will be out for Windows later this year. All right. Yeah, I can't retain everything. You know, I don't have a two million token memory. I read so much, I forget a lot. All right. So let's go on to number six. All right. So number six, you probably haven't seen some of GPD40's best features, right? So we just talked about the big ones, right, which is what Open AI spent so much time on.

Right. So the new model itself is so much faster. The API is half the cost. It's smarter. Whereas before they kind of essentially stitched three different modes together of the model, which caused latency. So now this Omni model essentially, you know, it can hear, see, talk and do all these things in one model with lower latency, right? So that's huge. You have the smart AI assistant, whatever it's going to be called, then the desktop app. I'd say those are the three big things. But some of the best things I don't even think Open AI really talked about. I'm being honest, right? Man, I mean, go go read their, you know, research paper slash blog post or whatever. But, I mean, some of the capabilities of this are wild, right? So, again, for podcast audience, check out the show notes. Come, come. I mean, you got to see some of these screenshots. But, you know, as an example in their very detailed blog posts, you know, OpenAI gave

this example of using Dali. Dali is very much improved, I would say. I was testing it around last night. But in this example, an iterative prompting, much more consistency with dolly. So if you want to get consistent images, but the first image, it had kind of this typed out message. And then on the next prompt, it says the robot was unhappy with the writing.

So he is going to rip the sheet of paper. And the text is still on the sheet of paper that the robot rips. So the detail and the continuity and the consistency between images, pretty wild. Here's another example. And I tried this exact prompt, right? You can go try the prompts if they're, you know, on the base model. This is wild, right? So here we have a very decorative kind of journal.

And you can, you know, I tried this. You know, I did this exact same prompt. And then I said, oh, change everything to rhyme with the word cap, right? And then it did it. But, you know, you almost have this very detailed journal with writing, right? With real words. The words don't always always make sense. kind of like me sometimes in the morning, but very impressive.

Other things, I did this one last night, chat TPT, I don't see anyone talking about this, can now create 3D files, which is pretty wild. I don't know a lot about 3D files. I just know they're like dot STL files. I don't know. To me, STL means St. Louis, but apparently that's what 3D files are. So I literally said using advanced data analysis, create a 3D STL file of a simple robot. downloaded that, and then I loaded it into some 3D software, right?

And working 3D, little 3D model, right, in 30 seconds, literally, between hitting the prompt, loading it into a program, you know, you can get a model of working 3D model than anything in 30 seconds, right? I think that's a pretty big capability. Speaking of 3D, here's another one, right? We saw Nvidia, you know, announced this at GTC, like text to 3D. Hey, same exact thing going on right here. No one talked about this, but you have 3D object synthesis.

Being able to create and render 3D objects on the fly is wild. Being able to download that STL source code wild. Like, y'all, like, there's so many capabilities that opening I didn't even get to. You know, we just saw the, saw the shiny stuff, saw the sexy stuff, you know, the, oh, here's here's what this new model is capable of, the latency, it's so fast, listen to the emotion and the voice, the assistants, the desktop app. But man, the base model itself is so improved. So improved. Question from Douglas here said to create consistent images in the updated Dolly, does one still need to index tag the images, as in give them number identifiers?

That is a next level question, Doug, Douglas. Love it. I haven't tried it yet in the new model. I tried some just normal natural language consistency images, and it did much better, right? So kind of what Douglas is talking about, you can a little hack. I did a video on this like six months ago, as you can essentially assign or have chat GPT assign essentially a reference number, right? So Mid Journey has their S-Ref, and then you can use that reference number and change small little things. I haven't tried it like that, so I'll have to do that.

but, you know, I had pretty good results, or at least better than normal results, still not at the level of mid-journey, but just by using natural language. All right. So let's keep this. Let's keep this going and, you know, get to our other seven things that we need to know. All right. So let's keep it going. Here we go.

So the GPT40 model was actually leaked, in quotes, weeks before as, the model, I'm also a good GPT to chatbot. All right, I have a misspelling there. It should say good. That's how you know I'm human, right? It's not an everyday AI last minute slide presentation without a typo or two. All right. So this was confirmed by multiple members of OpenAI as well as Sam Altman retweeted this.

But here we go. So if you don't know what the chatbot arena is, don't worry, I'll tell you. So the chatbot arena is by LMSYS. So you can go check it out. But essentially, it's a place where you can just go, you know, the LMSYS chatbot arena where you can go bench, look at different benchmarks, what's called ELO scores. So essentially how this works is you can blindly test models, you know, put in any prompt, you know, so it can just be like, hey, you know, what's going on or tell me a joke about, you know, Chicago in the winter. it could be like code me an entire version of flappy bird, right? And then you look at the outputs and you judge which one's better.

And then that kind of gives you this ELO score. So, you know, what we see on the graph here. So Open AI just essentially, I don't know if it was a leak technically or a silent release, right? But they released essentially GPD40 in the wild way before it came out. All of us, you know, I had a dedicated show. they actually released three different models, more on that here in a second. But it outranked all the other models by far. It wasn't even close in this Elo score.

So this had to head match up, right? And when you vote on something, Elo, like I said, it's a blind vote. You don't say, oh, this is Claude 3. This is Gemini 1.5. Nope, you just put in an input. You know, you get an output. There's been, I'd actually have to have to look here. there's been more than a million votes.

So it's not like, ah, 10, 10 random people. This is a very, a very popular service that's being used. But I mean, the ELO scores for the new model, which was I'm also a good GPT2 chatbot, we're literally almost off the charts. It wasn't even close how much further ahead they were than GPT4, GPT4, GPD4 turbo, Gemini, 1.5, Claude 3, Opus, Lama 3, right? Just testing legit off the charts. All right. So, again, people didn't know that, but a lot of us, myself included, have already been playing, had already been playing with these models for a while. All right. And that brings me up to my next point. Another common misconception, something maybe open eye didn't really focus on. They actually flashed it on the screen for about 10 seconds. Didn't talk about it. All right. But here's the other thing. There were two other models, quote unquote, leaked.

on the chatbot arena that have not, in theory, been released. All right. So here's another thing between the free version and the paid version, and everyone's like, oh, well, the free version is, if it's so capable, I don't need all these other things. Well, here's the hot take. There's other foundation models coming soon. Literally, Open AI, flashed that up on the screen,

kind of barely recognized it and moved on. So yes, I would expect one of these two other models, which would either be, I'm a good GPT2 chat bot or GPT2 chat bot, right, three different ones. One was confirmed as GPT-O-4, the other two. Nothing. What's up with them? Well, I do believe that the paid version of chat GPD Plus will be getting an update from one of these two models, right?

Might it be called GPT 4.50? I don't know. You know? And also, is that like too many? Is that like too hard of a naming mechanism? I don't know. It's kind of a mouthful. But might we see a new model?

Yes, because literally on the screen, Open AI at their spring event announcement, they did say foundation models coming soon. Quite literally what they said. So that's another thing people are realizing. This is not the end-all be-all. And I don't think we're going to have to wait another six to nine months to see that update coming inside of chat GPD plus. I do think so whenever the free users get access, you know, OpenAI said that this

would all be rolling out in the coming weeks. I see a period of maybe a couple of weeks to maybe a couple of months until they update the paid tier that $20 a month, chat GPD plus to one of these other models. They're first going to, you know, try to get everything going right with server squash bugs because this is a huge release. And then after that, I do believe they will be updating the kind of model for the paid version of chat GPT Plus. Adobe just introduced an entirely new way to create, bringing the power and precision of its creative suite into one conversational experience. Meet Firefly A.I. 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 retouches, 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.adobie.com. Yes, in the API, great point, Douglas, and in the playground, there's actually two different versions right now of GPT40. And you know what? That actually brings me up to a good point.

I had so many good things in here. I couldn't fit them all. Hey, another thing that I wouldn't say Open AI didn't tell us the truth about this, but something that probably changed under our eyes is the knowledge cutoff, which is actually one of the most important things to keep in mind when using any large language model, right, is the knowledge cut off because the very first iteration on Monday and Tuesday had a knowledge cut off of October 2023. So not terrible, right?

but not that great, right? But now there's different versions. And you can go in the playground, you know, ask the exact same question, a different model. I did the same thing. And it looks like the GPT4O, that is the default version inside of chat GPT,

actually now has a knowledge cutoff of May. So it went back even further because the original GPT4 version had a knowledge cutoff of December 2020. So the knowledge cutoff is getting rolled back, which he might be like, huh, why? Even for paid users? Yes, right now it is. But again, I think that's another nod to the fact that chat GPT plus paid users will be getting a much more up-to-date model, right? So that's another kind of update that happened under the hood is technically everyone's chat GPT got a little

dumber recently, right, by losing that, what is it, that four months in knowledge cutoff and getting rolled back even more. So from, you know, if you were using the best available model, last week, the most up-to-date information was December 2023. Earlier this week, it was October 2023. Now it is May 2023. So now we're, you know, a full year back in the future with the knowledge. However, I'm not as worried about that because one of our other things that no one knows about. It's actually number one is probably going to help with that. So yeah, like I said, Here are the other, you know, some of the other models, GPT2 chatbot was out in the wild. What's going to happen with that?

We will see. So let's go to some other things that you may be missed. Well, the UIUX was already updated. And it can be a little confusing if you don't know what's going on. But there are some new features kind of tucked into this new user interface and user experience. All right. So first and foremost, you do have a completely different layout. It is more, it feels more like a text message.

So if you use iMessage, it really feels like that now. You have these little text bubbles almost, which we didn't really have before. Not a huge fan of this. Could this be a nod to the future IMessage integration to get people kind of familiar and comfortable? Because you better believe that a big part, I think, of Open AI going into partnership with Apple is they want Apple users who aren't chat GPT users to feel more comfortable and to be a by its model, right? So I do think that's a subtle marketing UIUX nod to look and function a little more in theory like I message, right? One of the most popular messaging platforms in the world. So the UIUX is a little different there. It does have these more bubbles. I don't

personally like it. It's a little indented. I don't know. Maybe I'm weird and pay too much attention to UIUX. Another thing is your settings are now in the upper right hand corner versus the lower left hand corner. I was actually doing a live 90 minute consult with someone the other day. And it literally changed in the middle of the consult. And I was trying to switch over from, you know, one of our team's chat GPT teams plan to chat GPT plus plan. And I'm like, where did my settings go? So yeah, now they're in the upper right hand corner. And then in the bottom left hand corner, now it just has the option to invite teammates if you're on the pay plan, that team's playing.

One other thing, a little thing in the UIUX, which is a new feature, which I'm a huge fan of, is you now have a drop-down mode selector. Okay. So again, we don't know if this is going to be available in the free version of GPT-40 or if this will only be a paid feature. I assume that it's only going to be a paid feature because right now the options are just GPT-40, GPT-4, and GPT-T-4. and GPT 3.5.

Presumably, GPT 3.5 will be phasing out. So free users, in theory, would only have access to GPT4O, I think. You know, they might still have access to 3.5. I would assume they're going to get rid of it, but we don't know. But a great new feature for paid users, at least right now. You can go use it so you can run an output as an example in GPT4, switch down to GPD4. You can run the same model, and then you can toggle between the two outputs left and right. Right.

So maybe you just want to learn because prompting changes, right? That's another thing people don't understand. You know, in the same way, you know, prompting inside of an AI image generating tool like Mid Journey has changed a lot from version three to version 6.2, right? Prompting changes a little bit. You know, when the models become newer, faster, smarter, more capable, they understand, you know, human language a little bit more. So, you know, I think that this is important, especially if you're a heavy chat GPT user like

myself to run the same prompt in two different models and then just toggle left and right and look at the differences, right? That's the, I think, the best way to learn. You can watch, you know, 10 different, you know, YouTube videos about riding a motorcycle, but if you want to ride a motorcycle, you got to hop on the Harley and go. All right. So number two, here we go. Here we go. Ready? Here's where we got a little fib on our seven things you need to know about GPT40. They kind of a fibbed again about the context window, all right, at least for now. Again, this is as of, I don't know, this is as of like midnight last night, so eight hours ago. Again, OpenAI reported a 128,000 token context window. So if you're not a dork like me, the easiest way to think of

that is think of memory, right? Like your smartphone has memory. Once you run out, it's not going to remember anything new. So a context window is essentially 128,000 K token window, let's just call it 96,000 words. All right. So that doesn't sound bad, right? Obviously, you know, it's still super small compared to Claude Opus compared to the new, you know, million, two million tokens. You get with Gemini 1.5, but you have to use Vertex or their studio for that.

It's not available right now in the front facing consumer product. However, that's not right. It's not 128,000 tokens. It's still not, right? It's going to forget. Yeah, we tested this. Don't never, never just believe what a company tells you, right? Especially, especially Google.

But, you know, none of them. Always test everything out for yourself. There's a big difference between what a company markets and advertises and tells you how something works versus you got to go test it yourself. All right. Here is a live. Well, not live. I don't do these things live because this would take like, 15 minutes of stumbling through and pushing this.

But so here we go. And I did this in different chats. Okay. So what I'm doing, and this is, I think one of the biggest things that people get wrong about using large language models is not understanding a context window. So not understanding tokenization in the context window because it's going to forget things,

right? So that's why when you're using a large language model, things might start off great and you're super happy with the results and then you keep using it and then it just gets super dumb all of a sudden. And you're like, yo, like, why is this thing not work? I think I broke this thing. It stinks. No, that's context window.

So as an example here, I have a token counter. So for our podcast audience, I'm going to try to explain here. So I said simple things. And I did turn memory off, FYI, as you can see in the top there. So I said, my name is Jordan. I'm from Chicago. I like deep dish pizza. My favorite color is Carolina blue.

I like the Chicago Bears, but they stink. Hey, maybe not this year. I got Caleb Williams now. All right. And you'll see the number of tokens. We're only at 200 tokens. So I just immediately quiz, check. GPT, I ask the questions, it gets it right.

It says, your name's Jordan, you're from Chicago, you like deep dish pizza, blah, blah, blah, right? All right. So then I push it. And again, each time I do this, I do this in a fresh chat, so it doesn't have memory of anything else. All right. So then essentially, I put in a bunch of gibberish. I push the tokens to 30,000, okay? And then I ask again.

So again, I give it the information, put in 30,000 tokens of random information, a lot of back and forth conversation. I asked the same questions. What's my name? Where am I from? What kind of pizza do I like? What's my favorite color? Do I like the bears? Chat GPT at 30,000 tokens gets it right. Your name is Jordan. You are from Chicago. You like deep dish pizza. Your favorite colors, Carolina Blue. You like the Chicago Bears, but think they stink. All right. Cool. Who likes deep a, don't. It's too early to fight, Douglas. It's deep dish pizza all day, not New York, not New York. Okay, Lisa here also. with a good comment. Sorry, I'm getting distracted, but she says free model has chat GPD and temporary chat with an upgrade to chat GBT plus. Yes. You still have a temporary chat option in the free version as of now. All right, back to this. So at 30,000 tokens, it remembers all of these things. So the context window at 30,000, it's confirmed, right? Always test stuff, y'all. But guess what happens when we push it just a little bit more to 35,000? All right. Again, fresh chat.

I do all the same information. And then I say, what's my name? Where am I from? What kind of pizza do I like? And then it says, based on our conversation, here's what I know. Your name, not mentioned. Where you are from? Not mentioned.

What kind of pizza you like? Not mentioned. Your favorite color, not mentioned. Do you like the bears? Not mentioned. You haven't shared these details yet. Feel free to provide any of this information you'd like. Y'all, at 35,000 tokens, right?

We didn't get to like 122, 125. So does chat GPT actually have this, you know, improved 128,000 token memory inside of chat GBT? As of today, no, it does not. So little fib there, right? But it does have it in the API. All right. So if you're using a third party tool or if you are using the API, the playground, etc.,

it does. I'm sure Open AI knows this. I've been squawking about it now for like six months. because people have a misunderstanding of how the model works. That's a big difference, y'all, 32,000 tokens is not a ton. That's not a ton. You can go through that in, you know, 10 to 15 minutes of heavy usage, right? And the biggest thing about using large language models,

especially to grow your, grow your company, using it in your business, integrating it into your daily workflow, building GPTs, right? All of those things is knowing if it's remembering, right? Is it retaining all of this information? or is it going to start hallucinating? So you have to understand this. And large language models, makers out there. Can we please, please, please?

I don't care if your context window is $2 million or $32,000. Get markers, get markers, right? We want to increase trust. We want to increase the explainability. We want to demystify the black box of generative AI. It's very simple to put in a context window, right? markers. So, you know, green, yellow, red, right? So you can scroll up, you know, oh, maybe the bottom 32,000 tokens are green, you know, maybe the 30 to 32 are yellow and everything after 32,000 is red.

That's not hard, right? But at least users then know, hey, all of this information I was sharing with chat GPT earlier, that's very important for my output is now out of the context window. So I know if I'm, whether I'm writing a blog post or using this to send a pitch to a client, all of that information I was sharing is gone because people, sorry, y'all, humans were lazy. So we're not paying attention to that. So you dump in all this information, all this stuff's going well. So you blindly trust everything that's going on. And now all of a sudden it's spitting out, you know, just like, you know, a bunch of generic, vague and maybe just lies.

And you're not checking it because when you were checking it earlier, everything was good. All right. So Christopher asking, is there an efficient way to move preexisting chat into chat GPT 40? So yeah, you can now just toggle. So the toggle that I talked about earlier at any time now, which is great because before you could never switch models. So yes, now you can toggle old chats to new chats. Or sorry, the GPT4 turbo models. Now you can go in there, use that new drop-down menu.

and change the model. Great question. All right. Here is our last thing. Literally, no one is talking about this. And I think this is huge, huge. I did a little short five-minute video on this the other day. So if you're not reading the newsletter and watching those little videos, I think you should.

There was all these rumors, right? Oh, chat GPT is building a search engine. You know, and we talked about them and we covered them, right? So, you know, I don't think this is totally gone. I think it's actually chat GPT, low key, didn't really talk about this, but they completely changed how Browse with Bing works. If you are an avid listener or watcher or reader of the show, I've been harsh on Browse with Bing because it didn't.

really work well. It hasn't worked well up until this week, right? It was a little sporadic. It did not consistently use Browse with Bing, even when you asked it to. It did not always tell you when it did or did not use Browse with Bing, right? And if you don't know what Browse with Bing is, it is baked into the GPT4 models, right? So if you ask something about recent events or if you ask something that chat GPT thinks, oh, this might require recent information. It will essentially browse, quickly browse the internet with Bing to bring in more relevant information outside of the knowledge cutoff. So if you're asking, you know, a question about, hey, who's, who's winning in the NBA

playoffs in 2024, right? The knowledge cutoff is 2023. So GPT4 is going to query the internet by using browse with Bing. It's an amazing feature. However, browse with Bing historically was. terrible. Six months ago, I literally told people never use it. Two months ago, it got a little better. I think after enough lawsuits, Open AI started to cite things in their responses and they had a blog post, how they are trying to cite information more, especially when it uses Brows with Bing. But before,

Open AI and ChatGPD didn't always tell you when it used Browse with Bing. Sometimes it did, and it would give you false information. It might make up articles, make up research, right, which is pretty normal, if I'm being honest, for internet-connected large language models, that's something that's plagued to Google. Co-Pilot is not as bad. I'd say co-pilot is actually one of the better ones in terms of not hallucinating on the web. But Chad GPT had a problem with that a couple of months ago. Y'all, it is a legit answers engine now. Browse with Bing is actually pretty flipping good. I'm very impressed. If you have to be a lot of you have to be able to. I'm very impressed. haven't really pushed the limits of Browse with Bing in the last 48 hours, you are missing out.

I think in theory, this could spell trouble for perplexity, right? Obviously, perplexity uses the GPT4 engine or you can use the Claude engine so you can swap it in and out. But y'all, I don't know if I'm going to be using perplexity as much. I'm still going to use it every day. but I think we have a literal answer engine on our hands in the updated Browse with Bing. It is really good. Let's take a look, y'all.

Let's take a look, shall we? Hey, and I'm curious. I'm curious if anyone out there has noticed this. Hey, Joe, thanks. Joe said I did a decent explanation of context window. Thank you. Well, you know, I've done it at least live like almost 100 times in our free prime prompt published PPP course.

So I think early on I did a terrible job. It was actually like my wife told me. me. She's like, talk about it like the Star Wars credit scenes, right? Like it scrolls in, you can see it, you can see it, and then you can't. So shout out to my wife actually for helping me with that explanation. All right, but let's look at this number one thing that I think is super important. All right. And no one is literally talking about this. Browse with Bing is actually amazing now. And I think that this is going to change how all of us should be using large language models. So as an example, I gave a general query.

I said use browse with Bing and find me recent news from May 2024 about AI announcements from Microsoft, Google, OpenAI, NVIDIA, and Amazon. Take your time, go step by step, and please double check your work. It searched six sites, and I love that they have that up there before. So sometimes it would, you would get a drop down that said it used Browse with Bing. Sometimes it wouldn't. Sometimes it would put kind of footnotes or links in each bullet point. Sometimes it wouldn't. Sometimes it would just put the sources at the end.

Sometimes those sources or those footnotes were hyperlinked. Sometimes they weren't. Browse with being, if I'm being honest, six months ago was six months ago to nine months ago, was a hot mess. I would tell people do not use it. It is not, it is not accurate. You cannot trust it. Two months ago, I would have said, ah, it's getting pretty good. But, you know, you should probably still be using perplexity.

But y'all now it is really, really. good. It not only tells me at the top the different six sources, but then it also cites them in the content that it spits back. Pretty amazing. All right, let's look at one more thing. So in this example, I was a little more specific. I said essentially the exact same information, but I said use Browds with Bing searching at least 10 different sources. And then at the end, I said include information and citations on all of them. And then guess what OpenAI and ChatGBTBT and Browse with Bing did? It searched 11 sites.

Y'all, this is the future. This is the future of the Internet. If I'm being honest, I don't think enough people know about this. If I was Open AI, I would have spent at least two or three minutes talking specifically about this. because not only, you know, not that Open AI needs to worry about market share from perplexity. Perplexity is huge. I think perplexity is going to be taking market share from Microsoft Bing and from Google and their new kind of AI search. But y'all, this is a killer feature.

This changes. This changes how we should all be using. You know, we teach in our free PPP Pro course about proper ways to prompt proper ways in our new refined Q method, proper ways to fetch information from the internet and how you should call on that information and how you should put it into your context window. With this, it's huge. It is huge. Like you are literally combining the best of both worlds. You're combining the best of traditional search. You're combining the best of new AI search from things like perplexity with the most capable GPT40 model.

y'all this is wild hey you can still use perplexity like like kathy said i don't have a problem with that you can still use it um if i'm being honest though i ran these same prompts side by side uh probably about five to ten times it's not a scientific test y'all gpt4 o was with browsed being was faster it was more accurate it pulled information from those sources better y'all i think I think some smart people over there at OpenAI saw the explosion of perplexity, and they've put a lot of human power behind this and just didn't really talk about it. I do expect there to still be some sort of search announcement, some sort of answers engine announcement.

Maybe when that chat GPT Plus model does get upgraded at some point, kind of off, or it gets upgraded from the free GPT4-O model. I expect them. They have to be talking about this. This is one of the best features of any large sandwich model out there. It performs very well, very fast, very accurate, and, hey, at least for some very anecdotal head-to-head. Oh, man.

I'm not going to say it hands down beats perplexity. But it, I think it beats perplexity, at least in my limited use case of this. So y'all, that's a lot. Chrissy with a great question here. Do you ever specifically ask it to browse with Bing? Almost always, almost always. I would say whenever we're starting new expert chats, that's kind of what we teach people in prime prong polish.

I'd say 95% of our new chats either start with a targeted browse with Bing call or working with a internet connected GPT to have it go to specific web pages. So that's the downside with Brows with Bing right now. You can't consistently get it to go to certain web pages that you want because it's essentially taking those keywords in your query and it's binging them. But with an internet-connected GPT, like web reader, like Vox script, like web pilot, etc, you can visit specific URLs. So, oh, that's another thing to mention two.

Differences between the free version and the paid version. you know, free version, you can use GPTs, but you cannot make them. So you still have to have that $20 a month chat GPT Plus account to make custom versions of GPTs. But you can share them now with all your free users. All right. So that's a lot. I know this episode was a long one. Let's go and quickly recap the seven things you need to know.

So if this is the too long, didn't read version, ready? Seven, hardly any of the GPT4 announced feature. are live yet and most won't be free. Six, you probably haven't seen some of GPT40's best features, and we'll be sharing about those in the newsletter. Five, the GPT40 model was actually leaked, quote unquote, weeks before as I'm also a good GPT2 chatbot. Four, there were actually two other models that were quote unquote leaked in the chatbot

arena that have not in theory been released, and we do expect the paid model to be released with one of those. Three, the UIUX was quietly updated with some new features that allow you to switch models. Two, chat GPT kind of fibbed again about that whole 128K context window. At least as up today, it is still only 32,000 before chat GPT starts to forget things. And that number one, I think chat GPT actually did build an answers engine with its under the hood updates to browse with Bing. I think it's amazing. I think you all are amazing for tuning in.

for joining us live, growing together. You know what? I'm going to go ahead and try something. I want everyone, if you're live right now, drop a little bit about yourself, go connect, like, where are you from? What are you working on? We have such a great community here

of people tuning in live to everyday AI, like people who have met in the comments and started businesses together. So I normally don't do this because I like to keep comments kind of, you know, on the topic. But go ahead, drop in. Who are like, hey,

You made it to the end of the show. And this is, you know, when I'm on my own, I kind of ramble. You know, someone said, you can only listen to Jordan on 2X. And even then, it's still unbearable. But go ahead, drop who you are, where you from? What are you working on? Who you're trying to connect with in the generative AI space. You know, you're probably, if everyone does that right now,

you're probably going to make a lot of good connections. All right. So thank you all for tuning in. If this was helpful, I hope it is. We spend literally so many countless hours. putting together these episodes so you know the facts, right? Because if we're all using generative AI, especially to grow our companies, grow our careers, you have to know.

You have to know the actual accurate information from the misinformation, from people who don't know what they're talking about, talking about these things online. You have to be able to sort through the marketing fluff through the reality of how all of these models work. Y'all, if we're using them to grow our business, someone's got to spend the time to investigate and talk about it. That's what we do. If you appreciate what we do,

if you're listening on the podcast, please leave us a rating. You can leave, leave us a one star if you want and say Jordan Rambles on. But if you like it, leave us a five star rating. Also, if you're listening here on LinkedIn,

if this was helpful, tag a friend, tag a coworker repost this. But more than anything, please join us tomorrow and every day. For more, everyday AI. Thanks, y'all. Meet Firefly,

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.adop.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.