Ep 261: GPT2-chatbot – a new version of ChatGPT? (Large Language Mystery explained)

Episode Categories:

GPT2-Chatbot: Unraveling the Mystery

Chatbots are having their own moment in the digital world with the emergence of a chatbot leaderboard, where these revolutionary digital assistants are scored based on their performance. Imagine a ranking system similar to chess scoring, but for artificial intelligence (AI). The exciting world of AI just got a lot more competitive, raising the bar and sparking quality improvements across the board.

However, there's an air of uncertainty surrounding gpt2-chatbot, notable for its absences on ranking chart, throttling due to free usage, and slower performance. Doubling the perplexity is the mystery of its origin and ownership.

A New GPT2-Chatbot

Emerging from nowhere is a fine-tuned version of the original 2019 GPT-2 model. Its improved parameters management promises enhanced performance, creating ripples across the AI industry. Today, exploring chatbots and AI isn't just about addressing customer service concerns. This advancement can broaden the horizon of AI usage by leaps and bounds.

NVIDIA Strikes a Hefty AI Deal

Keeping pace with the dynamic AI environment, NVIDIA recently announced a whopping $700,000,000 acquisition of Run AI, an AI infrastructure management company. This move holds considerable significance since Run AI streamlines AI hardware infrastructure management, rent, and compute power. Furthermore, NVIDIA's acquisition of Run AI will open up this technology to a broader customer base, including Fortune 500 Companies.

Smaller AI Models Packing a Punch

Don't be fooled by the size - a smaller generative AI model can still be a formidable contender. They're trending right now, and for all the right reasons. These condensed packages of AI brilliance promise a robust impact on everyday AI usage.

Top Performing Models and Their Features

Among the most capable chatbot models in the AI landscape, GPT-4 Turbo, Claude 3 Opus, and Gemini from Google 1.5 take the podium. However, 'room for improvement' is a phrase that resonates well within the tech world, and these AI models are no exception.

Plenty of large language models are on the horizon, hailing from tech giants like Microsoft, Google, OpenAI, and Meta. They're designed to act as agents and drive the next wave of AI advancement.

GPT2-Chatbot: Speculations and Expectations

This new model, gpt2-chatbot, is built on the architecture of OpenAI's GPT 4. However, it promises more than just a hive mind of previously established AI capabilities. Its name and origin remain a mystery, there's compelling evidence which suggests that it could be a greatly revised, entry-level version of OpenAI's pioneering GPT technology.

Anticipated to be a multimodal, this mysterious bot could potentially boast a larger parameter size and demonstrate superior problem-solving abilities. Watch this space for more on the gpt2-chatbot, promising yet another AI revolution.

Word for the Wise Business Owner

In a rapidly transforming digital landscape, the ability to solve complex riddles, technical diagrams, niche recommendations, and geometric problem-solving with ease is a unique advantage. Bolster your business prowess with the rapidly evolving AI technology and stay one step ahead of your competitors. Engage with customers, provide solutions, and seize opportunities in business growth and expansion with the emerging generation of AI chatbots.

Topics Covered in This Episode

1. GPT2- Chatbot functionality and limitations
2. Origin of the GPT2-Chatbot
3. Speculation around ChatGPT and GPT2- Chatbot
4. Forecasts for Future AI Models

Podcast Transcript

Jordan Wilson [00:00:16]:
There's a brand new GPT model out right now. You can go use it, play with it, explore, discover, test it. But here's the thing, no one really knows what it is. Is this the next version of chat GPT? Are we looking at GPT 4.5, GPT 5, or is this something else? Well, we're gonna be talking about this new GP 2 chatbot today and more on everyday AI. What's going on y'all? Thanks for tuning in. My name is Jordan Wilson. I'm the host of everyday AI and this is for you. It is your guide to learning and leveraging generative AI to grow your company and grow your career.

Jordan Wilson [00:00:58]:
So, we do this every single weekday on the podcast. So if you're listening on the podcast, thank you. We appreciate it. As always, check your show notes. Make sure, you check those out for more information. If you're on a live stream, thanks for joining us as well, and make sure you check out our free daily newsletter if you haven't already at your everydayai.com. Alright. So we are gonna get into a good amount of depth into this new mystery gpt2-chatbot model.

Jordan Wilson [00:01:28]:
I played with it myself a little, maybe like 2 hours between last night and this morning. So I have some observations and some observations that I don't even think have been shared yet. But before we get into that, let's do as we always do. Let's start off with going over the AI news. And hey, as a reminder, do make sure we are gonna be giving away the meta ray bans today, so make sure to check out the, the newsletter where we're going to announce the winners. So yeah, make sure you do that. Alright. Let's talk about AI news.

Jordan Wilson [00:01:58]:
So Google just unveiled a new model for healthcare purposes. So Google just released Med Gemini, which it hopes will revolutionize multimodal healthcare. So Med Gemini is built on Gemini 1 and 1.5, and can be easily be adapted to new medical modalities with custom encoders showcasing promise and accurate multimodal, multimodal dialogues, particularly in radiology and dermatology images. So Medigemi's outputs are preferred, overdrafts in their, initial testing. They're preferred overdrafts from clinicians for tasks like simplifying or summarizing lengthy medical notes, drafting referral letters, or demonstrating practical applications beyond benchmarking performance. So the introduction of Med Gemini promises more accurate multimodal conversations regarding medical images, surgical videos, just a lot of things. So pretty pretty exciting, news there from Google Gemini and their new medical large language model. Alright.

Jordan Wilson [00:02:59]:
Speaking of Google and large language models, Apple has reportedly been, just pulling from Google staff to build its AI team. So according to reports, Apple has been has trying to significantly expand its AI team and resources, particularly targeting employees from Google and also establishing a secretive European laboratory. So Apple is focused on deploying generative AI on its next mobile device, which we talked about here on the show a lot. But it is facing challenges and not just utilizing the technology due to data and memory limitations, but also finding the right people, to help them build it. So according to reports, Apple has poached dozens of AI experts from Google and has also established a somewhat secretive laboratory in Zurich to expand its global AI and machine learning team. So Apple is particularly interested in deploying generative AI on its mobile devices, but faces limitations such as data and memory constraints. Also, as if you follow along on the newsletter, we share this every day. Especially in the last couple of months, Apple has acquired a lot of AI startups and has been investing heavily in AI research and development for over a decade.

Jordan Wilson [00:04:14]:
And like we talked about yesterday in our weekly Monday News That Matters show, Apple's kind of been flip flopping. If if they're going to partner up, with with Google Gemini for their next iPhone and iOS or if they're gonna be working, with OpenAI. So we're not sure, but, hey, they've been poaching a lot of people from, from Google apparently. Alright. Last but not least in the AI news, Drake has removed an AI generated disc track that featured the late Tupac and also Snoop Dogg. So Drake, the very popular artist who sings and raps. I'm not a fan, but some people are. But Drake took down his AI generated disc track, Taylor Made, after receiving a legal threat from Tupac Shakur's estate for unauthorized use of the late Tupac Shakur's voice.

Jordan Wilson [00:05:00]:
So the disc track drew attention for its deep fake manipulation of voices and sparked legal action. It actually sounded pretty realistic. I was surprised. Right? We've we've we've heard some of these, AI generated, tracks before and this one was actually pretty good, you know. Who knows? Maybe for me to like a Drake song, he has to, clone Tupac's voice. But Drake's, Drake's track was actually those seen as hypocritical, because he had previously condemned AI deepfakes of his own voice and had actually taken, legal action against unauthorized use. Snoop Dogg, though though reacted humorously to the situation, posting an Instagram video expressing bewilderment at the events. Alright.

Jordan Wilson [00:05:44]:
So we're gonna have a lot more, AI news, so don't you worry. Every single day we break down the conversation as well as, you know, different news, fresh finds from across the Internet. So make sure to go to your everydayai.com and sign up for that daily newsletter. And, hey, thank you for our livestream audience joining us. As always, hey, I'm curious. Has anyone out there used this new, g p, what it's it's the gpt2-chatbot. So we're gonna get into some some of the naming. We're gonna get into the rumors.

Jordan Wilson [00:06:15]:
We're gonna get into what this means. I'm gonna give you some of my first impressions, but I'm I'm curious if anyone, out there listening on our livestream has, has already explored this. So whether it's, you know, Tara joining us from Nashville, doctor Harvey Castro. Hey, doctor Harvey Castro. What do you think of the new, the new, Gemini? Med. Right? That should be good. Juan joining us from Chicago. Cecilia joining us from Columbia.

Jordan Wilson [00:06:40]:
Love it. Woozy. And, Ross, thank you all for joining us. So, a a couple things I'm gonna put out there. If I sound a little weird, if I look a little weird, apologies. I just just had a root canal actually. So, my mouth is feels a little weird from talking. I might be puffed up.

Jordan Wilson [00:06:55]:
So, you know, in case you're wondering why why is Jordan seem off today? That's that's probably why. Alright. But, what also seems off is this new gpt2-chatbot model out in the wild. Alright. So let's let's talk a little bit about this. So you are not gonna find this new mystery model by logging into chat g p t. Actually, the only place that you can find this model right now is the, chatbot arena. So, if if you haven't used this, before, let me just tell you real quick what this is and I highly highly recommend it.

Jordan Wilson [00:07:31]:
So if you ever are on Hugging Face, they link to this, but this is actually on a site. It's just called chat.lmsys.org. Okay? So this is what is commonly referred to as the LMSYS chatbot arena. It's where you can benchmark large language models in the wild. You can play, you you know, with a variety of different models. So right now, it's it's the only place that you can use this as far as, you know, at least as of last night. So I I was lucky enough to get to use it. It got super super busy and crowded because, you know, essentially, you know, Reddit and Twitter and even LinkedIn, every everything was blowing up saying like, oh, this is the new GPT 5, the next version of of chat g p t.

Jordan Wilson [00:08:15]:
I'll start here. I don't think it is, if I'm being honest. I don't think it is. I do have some thoughts on what this may be. But if you do want to give it a try, you're just gonna go to the chatbot arena, and then click on direct chat. And then you can always and you you know what? If you have never heard of the chat, Moderatorina, I should probably do a dedicated episode because it's something I use a lot. You can essentially put in a prompt and you can get back, 2 different, responses from 2 different large language models. So it's not gonna tell you which one it is until you vote on which one is better.

Jordan Wilson [00:08:51]:
And then from there, that's actually where they have a a leader board. So a lot of times we talk about, you know, a chatbot leader board. That's what this is. So, you know, it's from these blind scores. They're called like Elo scores. So if you're familiar with chess, that's kinda what it is. So it's it's kind of 2 competitors side by side and who wins and it's average people who don't know which, you know, whether it's Claude 3 Opus or you know, Google Gemini or you know, GPT 4 turbo or maybe it's, you you know, Command r or, you know, Mistral. Right? So you're gonna see just 2 head, like, side by side responses to whatever input that you put in for a prompt.

Jordan Wilson [00:09:30]:
Or you can just go and click the direct chat and use this model, by itself. Right? It is much slower, right, because it's it's free to use. You don't even have to have an account. So, the downside is, you know, even if you're using as an example, GPT 4 Turbo or Gemini or Cloud Opus, which on their, you know, on their actual platforms are very fast. Right? So on the, chatbot arena, they're not. So, that's probably how they keep it free. It is it is throttled. You know, it's it's not really meant for you to, you know, go actually use this output.

Jordan Wilson [00:10:01]:
It's it's meant for you to actually go play around with models and kind of score them as well. So with that out of the way, that is the only place that you can get it or use it right now. It is not available for download, and not a lot is actually known about this new, chatbot or sorry, gpt2dashchatbot model. No one's sure who put it up there. We don't know if it was a leak. We don't know if it's authorized. We don't know if even if this is OpenAI's model. We we assume it is, and I'm gonna get into that here in a minute.

Jordan Wilson [00:10:31]:
We we don't know a lot. All we do know is so far, it has been performing very, very well. Also interesting, normally, when models come out, within a couple of hours to a day, they show up on the leaderboard rankings. Right? We talked about that because when we covered, the new llama 3, we talked about how pretty quickly, at least especially on the English. Right? So you can you can sort, the models by performance or by language. And by English, you know, llama right away within a day or 2 was already, number 2 English model. It's it's number 3 as of today. But usually within a couple hours to a day, these different models are going to be showing up on the ranking charts.

Jordan Wilson [00:11:16]:
So for whatever reason, at least as of the time of this, livestream, this new gpt2-chatbot is not even showing up on the arena leaderboard, which is weird, because you can go in right now and vote for it. So, make sure, check out today's newsletter. Maybe it'll change within a couple of hours, but right now it hasn't. Alright. Here's the other thing. If any of you know this, I we don't really cover a lot of rumors on this show. Right? Because literally hey, and this is technically hot take Tuesday, so let me just get into hot take here. I can't stand people on on Twitter and people on Reddit, because, you know, lit it's I'm not exaggerating.

Jordan Wilson [00:11:56]:
Multiple times a week for the past couple of months. You know, it's always, oh, gbt fives out. Gbt fives out. Gbt fives out. You know, Sam Altman just gave an interview. You you know, the CEO of of OpenAI, he just said this about the future of of, you know, chat gpt. So that means CHET GPT 5 out is is out. It's it's not.

Jordan Wilson [00:12:15]:
Right? So let me be clear. This is not, as as as far as I'm concerned, my personal opinion is this is not, GPT 5. Could this be GPT 4.5? Maybe. Could it be a, slimmed down version of, GPD 4? Possibly. And I'm gonna get to that here in a second. But, normally, like I said, normally, we do not cover rumors because it's every single day that there's rumors. However, this model is live, at least on the, chatbot arena, which usually only happens when a model is officially released. So this is kind of the first time that I've seen, right, and I use this, chatbot arena leaderboard sites, pretty much every day, at least a couple of times a week.

Jordan Wilson [00:13:01]:
This is the first time I've seen what appears to be a very capable model that does not have some sort of release notes tied to it, that does not have, you know, some that that that does not exist anywhere else. Right? So, if you go on the normal, you know, repositories where you can download open models, it's not there. So don't think this is an open model. Maybe it could be. We're not sure. We're not sure how big it is. We're not sure if this is, you know, a 165,000,000,000 parameters, 2,000,000,000 parameters, or a a trillion parameters. We're not sure how this model was trained.

Jordan Wilson [00:13:32]:
We don't know a lot about this model, but here's what we do know. It's probably more than an average rumor, maybe, because CEO Sam Altman did at least acknowledge, this existence. So on Twitter, this was last night. He did say, I do have a soft spot for GPT 2. We know how hard it is to learn AI and how to actually leverage it, but that's what our newsletter is for. What's going on? This is Jordan, the host of Everyday AI. And I've got a real quickly just not just plug the newsletter, but a giveaway that we just launched for some cool AI powered prizes. So if you're hearing this little mid roll ad, our giveaway is still going on.

Jordan Wilson [00:14:19]:
So make sure you go to your everydayai.com, and sign up for our free daily newsletter. This giveaway runs just until April 29th, so you just don't have a lot of time. So if you're not already signed up for our newsletter, make sure you go do that today. So we're giving away not just the new Meta Ray Ban Smart Glasses, but some other very cool AI prizes you don't wanna miss that are gonna help you grow your company and grow your career. So if you're already signed up for our daily newsletter, just open it today and you'll see info. If not, go to your everyday ai.com. Sign up for that free daily newsletter, and you'll receive a welcome email with more information on entering this giveaway. So let's go learn AI together in the newsletter and grow our companies and grow our careers together.

Jordan Wilson [00:15:11]:
Alright. So a couple other things to even just think about, the naming mechanism here. Okay? I noticed this right away, and I'm like, oh, okay. Am I crazy? But I saw a lot of other people were talking about this, but even in Sam Ullman's tweet. So if you're listening on the podcast, you probably don't see this. But for our livestream audience, you'll know that there's no dash. There's no dash in GPT 2. So, obviously, there was a GPT 2 that was released, many years ago.

Jordan Wilson [00:15:38]:
Right? So it was actually released in 2019. So a lot of people, you know, some of the the the rumors or some of the thought is, okay. Is this just a very fine tuned version of that original model from 2019. It could be. It could be. Right? It could be an extremely fine tuned version that OpenAI has maybe cracked, the compute code and maybe this is something that they're showing off is that they can, you know, fine tune a very old, very small model. GPT 2 was 1,500,000,000 parameters. So so that could be a thing.

Jordan Wilson [00:16:15]:
Right? Because over the last couple of weeks, the small model game has exploded. Right? Exploded. So with Microsoft PHY 3, if you follow small models, with Llama 3 from Meta. So you have Meta and Microsoft, you you know, 2 of the 4 biggest, players in the generative AI space right now. So Meta and Microsoft over the last 3 weeks have both released very small, very capable models that are only a couple billion of parameters. And that really, I think, kind of changed, not only what is what we can, imagine to be possible with generative AI, but also maybe how models are built and used in the future. Right? Because the the tremendous upside of these smaller more capable models that are only a, you know, couple billion parameters like meta three's smaller model, like this new Microsoft PHY, well, is they can be run locally. Right? And when you can run models locally, that changes what is capable.

Jordan Wilson [00:17:16]:
Right? Like, that that changes what even society like like like, what we are capable to do with large language models. Mainly the reason is when you're running a reported 1.8 trillion parameter model like gbt 4 turbo, the latest model from OpenAI, that requires a lot of compute. Right? Like, so it requires so much compute that Sam Altman, the CEO, is out there trying to raise $7,000,000,000,000 to try to, you know, build or or buy more chips. You know, they're out there. They they created a partnership with Microsoft called Stargate, you know, 100 plus $1,000,000,000 data center. So these very, very large models, there's a downside to them. They require a lot of energy, like more energy than we probably have access to right now, which is why, you know, sometimes we get into, you know, the toll on on on the environment that these very large large language models play. So over the last couple of weeks when we've seen Microsoft, Microsoft's PHY 3, Meta's LAMA 3, very capable models that are only a couple billion of parameters, whereas g p d four was reportedly 1.8 trillion.

Jordan Wilson [00:18:24]:
So that does change it. So, you know, part of me thinks that, okay, this is an official model from, from OpenAI. You know? There there isn't even a claim aside from this tweet. Right? And this really set it off, but there's there's no dash. Right? The the GPT 2 model itself released in February of 2019 is GPT dash 2. Sam Altman put a little cryptic tweet up last night saying, I do have a soft spot for GPT 2. No dash. So a lot of people are saying, okay.

Jordan Wilson [00:18:56]:
Maybe this is just a new naming mechanism. Right? Everything we've talked about before was just GPT with a dash, right, and then a a version of it and this is GPT 2. So maybe we get into this new thing where it's GPT 2 dash 1 or, you know, GPT 2 dash and then giving it, like, a name like you do with operating systems. No one knows. All we know is this model that came out. It is small. It is very capable. Alright.

Jordan Wilson [00:19:21]:
So let's let's get into that a little bit here. And and, hey, if you do, have any questions or comments, let me know. So yeah. Yeah. Tara's saying she hasn't tinkered with it yet, but she's excited to. And we are gonna have all the links, in the newsletter so you can just go click and, and use it. And here's the thing. Literally, the rest of the Internet is is stuck on this because, when this model, when you use it in the direct chat, so like I referenced, you will only get 8 chats, and then it's timed out.

Jordan Wilson [00:19:53]:
And, also, even as of right now, or when I looked about 30 minutes ago, it was down. But I found a secret way that you can still use it. So, make sure to check out the newsletter, and I'll tell you that. Alright? So let's just go straight into some of the things that I did find out that I didn't see a lot of other people talking about. So one in well, one thing that, kind of caught my attention is I did ask this new gpt2-chatbot. I said, please, please be short and tell me what your training data cutoff is. Right? When is your knowledge cutoff? And it said November 2023. I should put this out there.

Jordan Wilson [00:20:32]:
Asking a large language model specific questions about itself is never the best idea. Right? You you can sometimes ask, as an example, a llama, like, hey. What are you trained off of? Or, you know, tell me about what model you are. And sometimes it might say, oh, I'm GPT. Right? Which is not true. So asking a model, certain questions and and, you know, saying, hey. This is truth. This is fact.

Jordan Wilson [00:20:54]:
Not always, the best route. However, I did confirm this. I ran this same version in, you you know, multiple, instances of this new gpt2-chatbot, and I got the same response over and over. So the November if if this GPT 2, chatbot is from November 2023 now, which cut off, that tells me a couple of things. One, it's maybe a newer model. Right? Because, again, the actual GPT 2, was from 2019, and I believe it had a knowledge cutoff of 2017. But it also tells me it's fairly recent. Right? Whether this is an actual model from OpenAI, whether, you know, a researcher leaked it, you know, we did see a couple researchers, you you know, kind of get let go or fired from OpenAI, about a month ago after there was, some leaks.

Jordan Wilson [00:21:43]:
So could this be a leak? We don't know. But the November 2023 date, I think, is important because, the most up to date model right now from OpenAI is December 2023. So, presumably, if this is, in theory, a GPT model from OpenAI, you can make the assumption that it's a very recent model, and it's not necessarily, you know, just a fine tuned version of the GPT 2 model from 2019. Alright? One other thing here. One other thing that I saw, and again, I'm gonna I'm gonna make my screen a little bigger here for our livestream audience. You can't like, I'm telling you all, take this with a grain of salt. Right? It's it's hot take Tuesday, so I'm just coming in here. I'm telling you guys exactly what I think, what I think this is, what I think this isn't.

Jordan Wilson [00:22:32]:
But I did ask, chat g p t or or sorry. See, I even made the mistake. G p t to chatbot. I said, what are you built on? Tell me a little bit more about the model. And it said, I'm built on OpenAI's GPT 4 architecture. Right? And then it goes in to tell me a little bit more that it's built off of GPT, and then I do have other screenshots that we're gonna be sharing in the newsletter saying that it was, essentially built on GPT 4 and saying that multiple times. Again, you can't take anything that you ask even from the most capable models. You can't take it as absolute truth.

Jordan Wilson [00:23:11]:
Alright. Let's talk about a couple other things. So some capabilities and comparisons. So, again, across the Internet, a lot of people were talking on forums and social media with speculation that this was GPT 4.5 or GPT 5. Also, the system prompt was allegedly leaked via prompt and jump, injection, hinting that it might be a GPG 4 variance. So we got a, smaller, version a smaller version of the system, of the system prompt. We were able to kind of extract it, out of this, model. We didn't get the the the very long one, although some other people online did, so we will share to those things and just like I'm saying, I'm putting this out there.

Jordan Wilson [00:23:51]:
None of this is confirmed, but it is very interesting because, hey, if this does turn out to be gpt4.5 or a very early version of GPT 5, then this is probably your first look at, technology that is going to change the way the world operates. Right? However, let's let's talk about the community and the speculative response. So, one thing that is for sure is this new version of gpt2-chatbot was extremely impressive. Extremely impressive. Right? I kind of have some some use case and examples that, you know, I always throw out there that normally can stump a large language model, whether they're, some some, you know, like, logic questions, whether they're some math questions, coding. Right? It was passing. A lot of the ones I tried and a lot of the other ones that people, you know, on Twitter and Reddit were trying as well. So a lot of people there's there's these common, kind of types of problems or, types of tasks that generally large language models, even the most capable ones.

Jordan Wilson [00:24:59]:
Right? Gemini Ultra, you know, or Gemini Pro 1.5, Claude 3 Opus, GPT 4 Turbo. There's these common questions that you can ask a large language model that it will usually struggle with. And, at least early testing shows that this new model did fairly well. Right? It did fairly well for us not knowing what the heck it is. So it did demonstrate some superior problem solving abilities, particularly in complex riddles and designing technical diagrams. Alright. And we are gonna be sharing, some of those examples in in the, in the newsletter today as well. Next, it did excel in generating niche recommendations and solving intricate geometric problems without code interpreters.

Jordan Wilson [00:25:43]:
You know, another thing that I, really was taken aback by was its ability to almost think like an agent. So something that, some things that I like to test models on is I give it a actual, like, sometimes a very complex task or sometimes a simple task. And I ask it to break it down in each step as if you were assigning this to team members. Right? And right now, even the most capable models, which the most capable models, in my opinion, in order are GPT 4 Turbo, right, which is crazy because it's an older model, Claude 3 Opus, and then Gemini, from Google 1.5. And even these models sometimes have problems almost, kind of taking on this agent sort of, role. Right? But that is the future of large language models is being agents. Right? And and, you you know, they are going to be completing tasks and they're, you know, these these, you know, generative AI agents are gonna be working together and and talking to each other and with each other. Right? So if if any of you are a little dorky like me, you know, Langchain is, you know, one of the more popular agents.

Jordan Wilson [00:26:51]:
We actually just, you know, shared about this, last week on our show. Will AI take your jobs? Yes. They will. You know, all the biggest companies, you know, Microsoft, Google, OpenAI, and and Meta, they're all working on agents. So, you know, another thing that I noticed on this, on this new model, this new gpt2-chatbot was kind of its way to think and respond in a way that you would want an agent to respond the way that it could, you know, essentially reverse engineer a solution to something and break down complex tasks step by step was pretty impressive. Again, my very informal, you know, studying of this model so far, it did a better job than, you know, some of the most capable models that are out right now. Alright. Let's talk a little bit about the technical and development, aspects.

Jordan Wilson [00:27:41]:
So technically, we don't know a lot about this. We talked about this. This whether it's a leak or whether it's authorized, it could have originated from a researcher, or developer with access to the model. In theory, it could have been a former employee gone gone wild. It could be the result of a hack, or like I said, it could be very official and could be authorized. It could be a direct play. It could be from OpenAI. It could be from no one.

Jordan Wilson [00:28:06]:
We're gonna get to that here in a second. So some users, online speculated that it could be a test version of a larger 400,000,000,000 parameter model, such as LAMA or an early 4 point, GPD 4.5 version. Also, it's important to distinguish, like we talked about, from a technical capacity. It is technically named a little different. GPT 2 dash chatbot is different than GPT dash 2. Alright? And again, we talked that was the original model, released in 2019 developed by OpenAI with 1,500,000,000 parameters. Legal issues and this is another one. Right? So, yes, it does seem like no matter what you ask of this new chatbot, it does respond and say it is a model from OpenAI.

Jordan Wilson [00:28:55]:
It is based on the GPT framework. It is, you know, derived from GPT 4. Right? No matter which way you ask it, you'll generally get a response, something along those lines. However, it is important to know, OpenAI was denied the GPT trademark, allowing others to use it freely. So again, just because we are asking this model, hey, what are you based off of? And for the most part, it's saying OpenAI, it's saying, GPT 4. There is no naming mechanism right now that says that is for true. Like like, that says that that is for certain, because the GPT name is not trademarked. Right? Technically, other people can use the GPT name at least right now.

Jordan Wilson [00:29:43]:
You know, that may change, here in the future, but that's important, to keep in mind as well. Alright. So great great question here from Douglas. So as we, you you know, wrap up this this episode, Douglas asking, is this only text to text or is it multimodal? It's a great question, Douglas. So the complete system prompt, which we will be sharing, in our newsletter, wanted to properly, you know, cite and source it and give credit to the person who was able to, extract this, does give a nod to the fact that this is multimodal input and output. At least right now, how you can go test it out. It's obviously only, text, right, because it's in the chatbot arena. But the, the rumors are or the kind of reporting out there right now from people who, you you know, have been giving it a a hard push is that it is multimodal.

Jordan Wilson [00:30:46]:
Yes. So, another question here. Wouldn't we be talking about something that would be terabytes in size? If it was a leak, I would suspect that OpenAI would be able to tell who downloaded that much data. Yes. In theory, if this is a, you know, like a 1.8 or a 2,000,000,000,000, parameter model, it would be extremely large. However, you know, like like I talked about, you know, midway through the show is this this could be, again, this could be OpenAI's, kind of attempt or retort to, these small models. Right? And these small models are obviously much easier to to download and to use and maybe less traceable if it is a leak or if it is authorized. Right? So, these these, models that are only, you know, a couple billion, parameters such as, you know, Microsoft's PHY 3, Meta Llama 3.

Jordan Wilson [00:31:42]:
Those are much easier, right, to download and to use. Whereas, you know, something like a GPT 4 or something with, you know, Gemini Ultra when there's trillions of parameters, it you can't just, like, download that easily or or, you know, you know, quote unquote, you know, leak it without a trace because it is huge. Right? Those kind of, models with the size they are leave a gigantic footprint where some of these smaller models, maybe not so much. Alright. So let's just get into it. Hey. It's Hot Take Tuesday. Here is predictions.

Jordan Wilson [00:32:14]:
Alright. So predictions. I don't think this is a GPT 4.5. I don't think this is GPT 5. I do think that this is from OpenAI, whether it's authorized or not. I do actually think this. So if if you all remember, you know, especially if you're an iPhone user, there was a new type of phone that Apple came out with, in 2016. I I I believe it was sandwiched, somewhere between the iPhone 6 and the iPhone 7.

Jordan Wilson [00:32:49]:
So what this iPhone was called was the iPhone SE. Okay? And essentially what this was is it was a, it was an updated from an old model, and they essentially just put some new guts in it. Right? So they essentially said, hey. Here's a phone. We're not gonna update a lot of it. We're gonna keep a lot of it kind of the same in in old, so it's a little more lightweight and a little cheaper. Right? So they came out with this iPhone SE, which is essentially a, you know, you could call it an entry level, cheaper version of the much more expensive and much more capable, you know, flagship iPhone of the time. If I had to make a prediction right now, I would say that's what this is.

Jordan Wilson [00:33:34]:
I do think that this is probably this g p, g p t 2 dash chatbot. I do think this is from OpenAI. Right? I've used, the GPT technology since late 2020, thousands of hours. I can usually tell even just by looking at an output, you know, even in a chatbot arena. I'm not usually good at that because I can usually tell which one is, you know, from GPT, like a GPT based model or an OpenAI model just because they kind of, you know, use similar words, they use similar formatting. They use, you know, similar structure, you know, similar hang ups, similar strengths, similar weaknesses. So so from my again, I I was only able to play with this for maybe about an hour and a half total between when it came out, yesterday and, you know, to, this morning's show. This feels like almost like the iPhone SE to me.

Jordan Wilson [00:34:26]:
I if if if I had to guess, I don't think this is gpt 45. I don't think that this is GPT 5. What I think this is is a GPT what do we wanna call it? Like, a GPT 4 Lite. I do think that whenever, you know, GPT 4 turbo gets replaced, whether that is with a GPT 4.5 or a GPT 5, I do think that whatever this model is right now, this, gpt2-chatbot, I think this is actually gonna be the free version. I think you you know what? It's it's it's hot take Tuesday, so I'm coming in with some takes. That's what I think this is. I think this is the eventual, replacement of, the free version of, Chat gbt. Right? So right now, it's still running on, gbt 3.5 turbo, which is, you know, pretty old.

Jordan Wilson [00:35:18]:
It's not that great. So if if if I had to guess, I would say it's it's one of 2 things. Either it is that, either it is the eventual free model, the eventual replacement, or it is OpenAI's first kind of small model. Right? So so whether this is, you you know, has a connection, to this reported, potential partnership with Apple. Right? That's that's a real capability as well. Maybe this is, you know, my, OpenAI's kind of answer to to PHY 3 or to Meta's llama 3, their small, 7,000,000,000 parameter version. Maybe this isn't an answer to Google's Gemma. Right? So maybe this is a small, version of GPT 4, just a fine tuned, much smaller version.

Jordan Wilson [00:36:08]:
Obviously, OpenAI has some of the best, engineers in the world, and maybe they've figured out, like a lot of other big companies have, how to still get, the most out of a model without it being enormous. So I would say if I had to make a prediction now, it is one of those two things. It is either, the eventual free replacement, not the flagship, quote, unquote, model. So when you think of the iPhone SE, that comparison, it's gonna be the free version of once the next version is launched, or it is a potential model that other, companies may be using when it comes to running Edge AI on device AI. So could this be the next, large language model for our iPhones? Could it be, the next large language model that we can download and run locally on our Macs, the first, from OpenAI? Maybe. If I had to guess, it is hot take Tuesday. I would guess one of those 2 things. Alright, y'all.

Jordan Wilson [00:37:04]:
That is it. I know this was normally, we don't cover rumors. I think this one was important to talk about because if you haven't seen it already, we actually did get this in the newsletter yesterday, but it came out after the live show, and right before we, sent out the newsletter. So make sure you're reading the actual newsletter. I think we were the first, the first people out there to get it in their newsletter because we got it, I don't know, within minutes after, after it was first reported, online. So, make sure, if you haven't already, go to your everyday ai.com, and we are going to be in our newsletter today, announcing who our winners are, not just of the Meta AI Ray Bans, which is gonna go to the person who had the most referrals celebrating our 1 year anniversary. But we are also giving away to 2 other people, random people, not people in 2nd and or 3rd place. These ones are random.

Jordan Wilson [00:37:56]:
We're gonna be giving away 2 90 minute generative AI, strategy sessions. These are something we don't even really advertise this, I don't think, but, sometimes companies hire us. You know, big companies, small companies start up, and they say, hey. We can't figure generative AI, generative AI generative AI out. I can't even speak because of this, this tooth pain here. And they hire us, and we sit down. We answer their questions and kind of at least get get them to a good next step. So we're gonna be giving away 2 of those consoles that we normally, charge a couple $100 for.

Jordan Wilson [00:38:27]:
So make sure, if you haven't already, go sign up for today's newsletter. Make sure you read it. Make sure you open it, and make sure you join us tomorrow for more everyday AI. Thanks, y'all.

Gain Extra Insights With Our Newsletter

Sign up for our newsletter to get more in-depth content on AI