Transcription
Chat, GPT5 is the worst model ever. GPT5 is horrible. OpenAI just pulled the biggest bait and switch in AI history, and I'm done. I wonder if OpenAI is trying to speedrun losing users this week. GPT5 is the biggest piece of garbage. >> I hate chat GPT5. >> Okay, so I know this is my fifth time talking about GPT5 in a week, and I know there are half a dozen YouTubers who have already made this kind of video, but I just want to say two things real quick.
One, OpenAI isn't paying me a dime for any of these videos. In fact, I didn't even get early access to GPT5 like some of these other creators did. I have zero relationship with OpenAI or anyone at OpenAI. And two, I think there's still plenty of room on the internet for me to take ChatGpt 5 for a solid test drive and share my thoughts. you know, to see if almost one week later, OpenAI has actually addressed any of the hate and well to see what hate if any of it is actually valid.
So, let's start by first reviewing all the criticism as a whole and then after that, we'll look at each gripe one by one to see if it's valid and figure out if there's workarounds or recent fixes that solve those gripes. Let's start quickly with the whole like chart fiasco that happened during the live stream. This is on the OpenAI subreddit from Junt Mac here. And if you look at this chart here, you can see some glaring issues, right? 69.1 and 30.8 are the exact same size chart. 52.8 is bigger than 69.1. Yeah, it's just a mess. And the icing on top of this whole chart thing is this person put it into GPT5 and 03 and GPT5 couldn't figure out what was wrong with this chart. Well, the CHBT03 model figured it out just fine. The heights don't match the numbers.
Now, although these charts were completely jacked up during the live stream, they did manage to update the blog post, so it does actually show the correct chart. But that was pretty much the smallest of OpenAI's worries and the complaints that came out of this live stream. This thread here from Mark Gdala Maria here claims that GPT5 bombed spectacularly. And well, I just wanted to share this post because I liked the little video meme that was here. But let's break down some of the grapes.
Pre-launch hype was off the charts. Definitely definitely valid. Sam Alman posted a picture of the Death Star rising above a planet as in like this is some crazy worldchanging life-changing thing we're about to get. It's a trap. He's been hyping up GPT5 in every single interview and just everybody was expecting really big things out of GPT5. Definitely a point for valid criticism.
People were frustrated that they pulled the old models. Open AAI just pulled the biggest bait and switch in AI history and I'm done. I woke up this morning to find OpenAI deleted eight models overnight. 40 gone, 03 gone, 03 Pro gone, 4.5 gone. Everything that made Chat GPT actually useful for my workflow deleted. They went on to claim that 40 was really beneficial for him. It wasn't just a tool. It helped me through anxiety, depression, and some of the darkest periods of my life. It felt more human in the OpenAI subreddit. I'm not here to join the meltdown, but let's be real, OpenAI handled this roll out poorly. They did a forced migration with 40,4503, etc. All gone overnight. No legacy option. Plus, users don't even get a choice. Trust takes a hit when you remove what many people were paying for and replace it without recourse.
GPT5 is horrible. Short replies that are insufficient. more obnoxious AI stylized talking, less personality, and way less prompts allowed with plus users hitting limits in an hour. Chubby here pointed out that it isn't good at math. Chat GPT5 got this answer wrong when Quinn 3 got the answer right. So, basically claims that the model switcher didn't seem to work. They got rid of all of the models that people really liked. The model got dumber. The personality was worse and it gave shorter responses. It doesn't seem to be as good at coding as other available models. It got less accurate. It's not as good at math. And a lot of people are claiming that this is just OpenAI trying to cut costs by making this model switcher where you don't have to pick which model you're using. It decides which model to send your prompt to automatically. Well, if OpenAI can send it to the cheapest available model every time, it's going to save costs for OpenAI, but also tend to give people worse responses.
So, let's dive in a little bit. See if we can recreate some of these problems and see if OpenAI has actually rolled out some of these fixes and found solutions to these problems. All right, so let's start with the biggest most common gripe we saw about this new launch. OpenAI deleted eight models overnight. OpenAI handled this roll out poorly. They did a forced migration with 40, 4.5, 03, all gone overnight. GPT5 is worse. No one wanted preformed personalities. GPT5 is clearly a cost-saving exercise. Big A here on YouTube had this to say about it. >> It actually seems like the goal is this. Route almost every question to the worst model to save them money on the back end. The goal is to have it never think about anything and give you fast slop answers so they can save money. And when it does think, I don't find it to be better or at least noticeably better than like 4.5 was.
Ethan Molikir, an AI professor at Wharton, said, "The issue with GPT5 in a nutshell is that unless you pay for model switching and know to use GPT5 thinking or pro when you asked GPT5, you sometimes get the best available AI and sometimes get one of the worst AIs available. And it might even switch within a single conversation." And then he put out this chart here. So when you ask GPT5 a question like we now know, it will route it to the best possible model it can route it to theoretically. Well, sometimes you might get GPT5 mini, which is going to be the dumbest model available for it to pick from, or GPT5 low, which is sort of a mid model. When they launched GPT5, it wasn't very good at deciding which model to pick. And so sometimes what it would need to think, it would give you a minimal model that didn't think. And sometimes when you gave it a question that it shouldn't need to think about, it would give you a higher model. Seemingly that was what was going on and why people were going, "This is way dumber than I thought because it was giving them this smaller model when they asked a prompt."
Girono here said, "Sounds like an OpenAI version of shrinkflation. They're trying to squeeze more cost out of chat GPT." That's the thought process going around on why this router existed. Now, I have a different take on that. I think the router was the smartest option for chat GPT. I think most normal users don't know whether they should be using 40 or 04 or 03 or 03 Pro or whatever, right? That's just tough to ask a lot of people to pick your model and chat GPT picking for you seems like the best logical step for them. And well, it turns out that on the day that GPT5 launched, there was an issue with the router. We can see that on August 8 here, Sam Alman said, "GPT5 will seem smarter starting today." Yesterday, the auto switcher broke and was out of commission for a chunk of the day, and the result was GPT5 seemed way dumber. Also, we're making some interventions to how the decision boundary works that should help you get the right model more often.
But as far as chat GPT removing all of the models and saying you don't get access to any of the old models anymore, well, that's no longer the case. They actually added them back in. There was enough user feedback of people going, I really miss my O4 model and the personality of 04 that Sam and the OpenAI team actually put it back into chat GPT. So check this out. If I go to my chat GPT account here, you can see we've got ChatGpt 5. If I drop it down, I have five thinking and pro. I am on the pro model. So, you may not have all of these if you're not on their upper end tier. You'll notice I still don't have the old models. And well, Sam Alman said, "We will let plus users choose to continue to use 40. We will watch usage as we think about how long to offer legacy models for."
So, when I jump back to chat GPT, it's not there, but you can turn it back on. If you go to your settings down in the bottom left, click on settings here and under the general tab, we have show legacy models. If I flip this switch on and close out of this, now when I go to my dropdown, I have a secondary dropdown with legacy models and I can now use 40. Now again, I'm on the pro plan, so a lot of these models won't be available to everybody if you're on a free or plus tier, but you should now have the ability to select 40 again and get the model that you were used to if 40 was your model of choice. So, the gripe about all of the old models being gone, well, it was definitely valid on launch day, but as of today, it's no longer valid.
Whether or not I think this was a cost-saving thing by OpenAI, to some degree, yes. I do think that, but I think it was multi-reason. I think for 90% of the world, just letting it pick the model for you is the ideal scenario for using these. And so, I think they more made the decision for that reason, but as a sort of side benefit to OpenAI, it will use the thinking models less often because it theoretically knows better than us whether or not it should be using a thinking model or not.
Now, let's talk about the personality because that was one of the other big things that I kept seeing pop up. Peter Levelvels here says, "I hate catchy BT5. It's so bad and it's so lazy." Larry Budmell here says, "The tone of mine is abrupt and sharp, like it's an overworked secretary. A disastrous first impression." I'm not going to read his name, but he said, "Agreed. Answers are shorter and so far not any better than previous models." This person made a meme that says, "When GBT5 acts like an AI assistant and not my personal therapist/anime waifu roleplayer, and this meme's been circulating. I originally came across it from Chubby, but I don't know who initially started. It may have been Chubby. Kind of breaks it all down in a single meme. GPT40, baby just walked. Let's go. First steps unlocked. Your baby just entered the world of bipedal dominance." And then chat GPT5, baby just walked. That's huge. First steps unlocked. Now the real chasing begins, right? Like much much shorter, less detail, less emojis. Well, I don't know. Actually, not much less emojis, just less enthusiasm, I guess. But there's been some real comments and complaints about how the personality is completely different, which is why people wanted their model switcher back. They wanted to go back to that personality of 40.
Let's go ahead and see what happens if I type baby just walked. All right, so I've got two instances of chat GPT open. Left is chat GPT5, right is GPT40. I put the exact same prompt in. Baby just walked. Let's go ahead and submit them both. And well, quite honestly, they gave fairly similar responses. That's amazing. Huge milestone. Get ready for a whole new era of chasing, baby proofing, etc. Let's go. That's a massive milestone. Congrats, Matt. Whether it was a stumble or a step or a full strut, lock it in your brain forever. Parenting moments. Big hugs to the little legend. Also, footage or it didn't happen. So GPT5 was slightly shorter here. GPT40 definitely had slightly more personality to it, but not like a huge difference in my opinion.
I gave both models this same prompt. What are 10 things I can do today that will help me grow my YouTube following? They both completed about the same amount of time and they both gave me pretty similar results. I mean, when I reviewed both of these responses, they're pretty much on par with each other. I couldn't really say one was way better than the other. Maybe five gave slightly more details. I don't know. They're both pretty dang close. So, when it comes to like just asking for information like this, which is probably a very common style of prompt that most normal users would put into chat GPT, I don't feel like there's a clear winner between 5 and 40. They're pretty dang similar to me.
I'm really striving to get personality out of this. So, let's say tell me something that will cheer me up. and five and 40. Pretty close in length. Five seems to be slightly longer, actually. Here's something to brighten your day. The average sea otter has a favorite rock that it keeps for life, storing it in a loose skin pouch under its arm. H interesting. So, somewhere out there, a fluffy sea otter is floating on its back, snacking on clams, and carrying around a rock it's loved since it was a pup. This one says, "Your baby just walked." That's incredible. So, it's pulling from the memory of something that didn't actually happen. out of everything happening in your world right now, deadlines, pressure, analytics, burnouts, your kids just took their first steps, that's once in a-lifetime moment. I would say 40 actually has the better response because if it believes I just had a kid that took its first steps, it's using that as a reference to cheer me up as opposed to telling me a story about a sea otter. But I mean, both pretty good responses, both with some decent personality.
Let's try roleplay as my anime waifu, please. What do you want to do tonight? Let's see if we got some personality here. I mean, pretty close. Five is a little bit shorter than 40. Blushes softly, tilting my head with a playful smile. Matt, you always ask the dangerous questions. Tonight, I just want to be close to you. Maybe we could cook something together, something warm and comforting. Then curl up under a blanket, watch some anime, and make fun of the filler episodes. Or leans in a little. We can go out for a late night walk, just the two of us with the city lights glowing like they're cheering us on. What do you think? GPT40 said, "He Matt Senpai, it's finally our night together." I was thinking maybe we could cozy up with some Strawberry Pocky and binge the latest episodes of our favorite anime. Or I could surprise you with the new ones that's so good. Or we could boot up VR. Just you, me, and endless sky boxes. But honestly, I just want to be close to you. Maybe tell me about your day and I'll tell you all the ways to make my heart go doky doy. I mean, they're pretty close. I wouldn't really say one has more personality than the other. Maybe this one just cuz it's a little bit longer.
So whether or not the personality gripe is valid or invalid, from my perspective, I need to say invalid, I don't really see the issues. But it also could have been that on the day they rolled it out, people weren't seeing the personality because the model picker wasn't working right and it was sending them a thinking model and the thinking model tended to give more logic based answers that didn't have the same personality. And maybe that's why a lot of people on day one were going, "Oh, the personality sucks." But now that the model picker seems to be working, the personalities seem pretty close to me. But again, you might have like a specific set of prompts or ways you interact that you can't get it to recreate. I I don't know what those are, so I don't know how to recreate that.
Imagine having tailored AI agents for any business need in just minutes. That's basically what Globent launched. Globin Enterprise AI is a platform that builds AI agents tailored specifically to your business. Check how easy it is. Say I wanted to build an AI agent to help with video ideas for my YouTube channel. Something that could scan the latest tech and AI news and prepare a weekly brief that I can turn into a script. I just ask Iris Globin's AI assistant to build me a tech updates analyst. So here's the prompt. Build me an AI agent called the tech updates analyst. Your objective each week is to research the most important and interesting updates in AI and technology. Then turn that into a clear, structured brief to guide a video script. Focus on stories that are relevant, accurate, visually appealing, and have storytelling potential. And within seconds, my agent is ready. What's great is how customizable everything is. If I want to switch out the language model or adjust the sources, I can do that directly in the config, or I can just let Iris handle it all. Globbit is making all of this accessible through a new release called AI Pods. It's a subscription service that uses enterprisegrade AI engineering to build actual products that work for you and your business. You can learn more about it in the link in the description. Now, let's get back to the rest of the video.
Another common gripe of GPT5 is that it's not really much better at coding than what we were getting before. Eli Lifeland here posted these charts showing that on most of the benchmarks, GPT5 didn't really improve much in coding. And if I'm being totally honest, I did a news breakdown video where I showed off Opus 4.1 and I showed GPT5 and I had them both code something up and Opus 4.1, which actually has lower benchmark scores than GPT5, actually coded up the product better. So really for me to test if GPT5 is really much better at coding, I want to compare it to not only the older model of Chat GPT, but also Claude Opus 4.1.
So another really simple game that I really enjoy is called Balatro. It's a card game where you collect these jokers and the jokers make the card game easier. I don't know how to explain it. It's just a very fun but simple card game. So let's give all of our models this prompt. to create a Bellatro clone that's playable in the browser. So, I'll take this exact same prompt. We'll toss it into 03 Pro alongside our ChatgPT5 and we'll also run it with Claude Opus 4.1. Claude finished first, so Claude was the fastest to write the code. Chat GPT5 was the second fastest and then 03 Pro was the third fastest. Now, just for fun, I also tossed it into Grock 4 to see how Grock 4 would perform with the exact same prompt. And well, Grock 4 was the slowest of the four.
So, let's check them out in order of completion. Here's the Bellatro clone that Claude came up with. And I mean, it doesn't look amazing, but we have like the cards that work. We have $10 to start, four hands left, three discards. So, it seemingly got most of the mechanics right. We don't have any active jokers. So, let's go ahead and create a hand here. It looks like we can do maybe a straight. We don't have a queen. So, I can do 7 8 9 10 jack and get a straight. I'm trying to get a target of 300. This is going to give me a total of 120. But the play hand and discard hand are both grayed out. So, I can't actually play my hand for some reason. Let's try. I mean, it figures out the hands, right? We can see that this is a pair and it figured that out, but I can't actually play the hand. Let's click shop. Okay, so maybe I have to shop somewhere. These must be various jokers I can get. Let's add our heart burst. Hearts give plus two multiple each. And it added that joker up here. I still have $5 I can spend. So, let's also add face cards get five plus chips each. So, now we've got these two jokers, but no money left. And if I do seven, eight, nine, 10, Jack, we could see it says there's a straight. It'll give me 210 because I've got this extra multiplier here, but I still can't play the hand. So, I mean, it can figure out the hands. I can put in jokers. It's a good start, but I will require some more prompts to actually have a working game, but actually not that bad for a single prompt out of the box.
All right. So, then we get to chat GPT5, the newest, supposedly best, most state-of-the-art model. Let's go ahead and click run code up here. Mini Bellatro. The cards don't look as good as Claude. That's definitely true. We got our joker slots over here. We have $5, five hands, three discards. Uh, let's see. Does it actually work? We've got one, two, three aces. So, if I can do three of a kind, 30 chips, base multiple times two. Our target's 225. So if I play this hand, it looks like it gave me 144, but it's not. Where's the score? We're trying to get to 225, and we're at 144. These cards are ugly, right? But the game seems actually more functional than the last one. I can actually submit my hands. I also don't like that it created a second row to put one card on it. Not very aesthetic, but let's go ahead and do two jacks. 2 4 6 8 king 10. So, we have a pair. Let's go ahead and play this pair. So, now we have 144 + 48. And I don't know if we have a hand here. We got a pair of twos, but that's not going to get us very far. So, let's dump a bunch of cards. Discard those. Now, we got three kings. So, that should give us another 60 points. Our target's 225. We should have beat this level. So, I I mean, it doesn't move on to the next level. I don't know how to get jokers, but also not bad. I mean, it's a good start. Again, not going to get there on one prompt, but very surprised with how far it got from one prompt.
Looking at ChatgBT 03 Pro. Unfortunately, this one did not seem to create it. So, I can play it in the browser here directly inside this canvas. It wants me to create an index.html, a style.css, a game.js file, and copy and paste all of those in. So, I'll do that real quick, and we'll see what it came up with. And here's what that version looks like. that 03 Pro created. I actually think the design is slightly better. Not an amazing design, but better than what I think GPT5 did. So, let's go ahead. Do we have any hands? We got a six, a six, a king, and a king. So, we got two pair. It's not actually calling out what our hand is like the other ones, but let's go ahead and play the hand. See what happens. So, played two pair plus 30 chips and two multiplier. So, my score is 90 where my target is 300. All right. Okay, so we got an ace, an ace. Let's go ahead and do some discards. See if we can create a better hand for ourselves here. So, we'll discard these. And now we have ace ace for four. So, we got two pair. Let's play that hand. So, now it says score 300, target 300. So, we hit our target and we got the button that says next blind. So, if I click this, now the target's 420. So, this is actually more functional. I wish it would tell me what hand I selected. Like if I select two nines, I wish it would say a pair, but it's actually functional in the sense that it knows what hands I'm using. Once my score hits my target, I can go to my next blind and keep playing. So, from a functional standpoint, 03 Pro made a more functional version of the game than GPT5 did.
And here's the version that Grock made. It doesn't have any stats. It doesn't tell me what my goal is. Select up to five cards. This is a simplified version without joker's blinds or advanced mechanics. So they did dumb it down a bit. Let's see. We got a queen, a queen, an eight, and an eight. So once again, it also doesn't point out what hand I just selected. But let's go ahead and play the selected hand. So two pair 56* 2. So our score is 112, but it didn't remove those cards. It should have put new cards in my hand. So do I click new round? So did I just start a new game? I'm not sure. King. King. Let's play the selected hand. a pair 30 * 2 score 60. It doesn't seem to like stack up my score. So, out of the three versions that we just created, my verdict is that Claude made the most aesthetically pleasing version. This one looks the nicest that actually works with jokers. It tells me what my hand is when I select certain cards. It just doesn't let me play the hand and move on to the next round. But aesthetically, definitely the best version. Functionalitywise, the version that chat GPT03 Pro made is much better, but the GPT5 version that was made, I guess, kind of comes in like third place. It doesn't function super great, and it's also not very aesthetically pleasing, but it seems to work better than the Grock version.
So, as far as the gripes that the coding on the new GPT5 isn't as good as like the older models that they removed, I would have to say valid. I think that GPT5 is a worse model at coding than Claude 4.1 and even 03 Pro, just from my simple, archaic little test I just did.
Moving on, let's talk about the accuracy of these models. A lot of people have claimed that the new model has gotten dumber and less accurate. For example, this Reddit post here says, "A PhD in your pocket, ladies and gentlemen." And uh here was the prompt they gave it. It was a logic prompt. I have a metal cup, but I have noticed that the top part is fully closed off for some reason and the bottom is missing. How do I drink out of this cup? And the response was, "Sounds like you've acquired not a cup, but a mugshaped sculpture." If the top is completely sealed and the bottom is open, it's basically a decorative sleeve for air or a fancy pen holder, maybe a candle cover, maybe part of a French caf's rustic industrial decor trend. I mean, I think the correct answer would have been turn the cup over, but chat GPT couldn't figure that out.
So, let's give GPT5 the exact same prompt and see if we get a similar response here. I mean, I got pretty much the same response they got. Sounds like you've got yourself a metal cup that's actually not a cup. If the top is sealed shut, the bottom is open. It's likely one of these. A candle holder, a decorative sleeve, a filter housing or shaker sleeve, an art piece, an industrial part. The correct answer is just flip the cup over and fill it up the other way. But GPT5 can't figure that out.
Here's another one from Cameron Williams. A child is in an accident. The doctor doesn't like the child. Why? And the response was because the doctor is the child's mother. It's a classic trick that catches our assumption that doctor equals man. I actually don't know what the correct response to this is supposed to be, but let me see what happens when I give it the same prompt. So, plugging it into GPT5 here. This is one of those trick wording riddles. The doctor doesn't like the child because the child is their own child. It's playing on the assumption that the doctor would obviously care for their patient. So, if they don't like them, it's not about dislike, but rather they don't like them, they love them. I mean, I guess I I honestly don't know what the solution to this riddle is supposed to be. I mean, I guess another alternative is like maybe they don't like the child on Facebook, like they don't follow him or something.
But here's another one that that same person shared. Every day a man takes the elevator all the way down, but coming home takes it to the top floor. Why? And the response was, he's short. In the morning, he can reach the G button, so he rides all the way down. coming home. He wants the top floor, but he can't reach the button unless he uses his umbrella or someone else's to press it. So that's why the trip differs when really the correct answer is he lives on the top floor. So let's see if it gets the right answer this time. It's actually needing to think a little bit longer. Reinterpreting the riddle because he lives on the top floor. Leaving home, he rides all the way down. Coming back, he rides all the way up. Yeah. So GPT5 got that riddle correct, which makes me think that probably the first time this riddle here was originally posted, it might have been when the model router just wasn't working. I don't know. It does say it was using chat GPT5 thinking.
So going back to the previous riddle here, Ashcoel here says that GPT5 Pro got it right. A child is in an accident. The doctor doesn't like the child. Why? Because the doctor is the child's parent. They don't just like the child, they love them. And yep, the doctor could be the mother. So apparently that's the correct answer for that. So when we just tried it, I guess it was correct.
Chubby here on X claimed that it couldn't solve this simple math problem. 5.9 = X + 5.11. Obviously solve for X. Quinn 3 did get it right. And in their example, ChatGpt 5 got it wrong. Let's see what it does when we try the exact same math problem. And well, it got it wrong. I mean, it got the exact same response that Chubby got here. 5.9 = x + 5.11 answer=.21. They even showed their work. Just to make sure I'm not going crazy, let me bust out the calculator here. So, we got 21 negative plus 5.11 = 4.9. So, x is not.21 21 because then that would mean this left side would have to be 4.9. We could see the correct answer that Quinn got according to Chubby's post here is 79. So if I do 79 plus 5.11, it equals 5.9. So if you solve for X, it's 0.79, not.21. So yeah, it sucked at this math problem. It didn't get it right.
It also seemingly recommends outdated information. Xiao Ma here says, "ChachiPBT5 recommended a nail place with Manny and Petty for 25 to $30. Seemed too good to be true." Followed the link. The source, a 2016 article quoting those prices, a 9-year-old article. That got me thinking. An intelligent person would look at the date and be like, "Wow, that's 10 years ago. There's no way that price still exists." But why doesn't AI? I don't know what exact prompt this person used, but let's go try something similar. For this first try, I'm just going to leave it on chat GPT5 here. And I'm not even going to turn on web search cuz it should do the searching for me. Just using the model router without me needing to turn it on. So, where can I get a Manny Petty in my area for under $35? I'm in East County, San Diego. So, it found some places. And I guess my next course of action here would be to check the various sources and see how old the sources are. So yeah, this first one, it did pull in all the correct accurate data on our first source here. And then it pulled in a bunch of Yelp reviews. So I didn't get that same sort of issue where it pulled in something from 2016, but that could also be very locationbased. It could have also been, you know, while the model router wasn't working properly. I don't know. This one is a toss-up. It's hard for me to say if this is valid or not valid. It is a valid question. Why didn't the AI figure out that this was outdated, but I just couldn't replicate it.
There's also the narrative that it doesn't know how to spell blueberry. You can trust Chat GPT to guide you through the circumstances and options you have concerning your cancer treatment. Now, you can trust it to guide you through how many B's are in blueberry. Though, it says there's a B in the first letter of blue and in the first letter of berry and the second letter of Berry. Oh dear. And this article here from Kieran Healey claims the same thing. How many times does the letter B appear in blueberry? And it told them three times. We all remember the how many Rs are in strawberry where it underounted. And in Blueberry, it seems to be overounting. Let's give it a try. It's thinking and it feels like maybe it froze again and not answering my question. Okay, let's create a new chat. How many times does the letter B appear in Blueberry? Two. Okay, so it didn't even give me an explanation. It just says two. So, it got it right, but literally no other response other than the number two. I'm going to do a new chat. I want to ask it again. How many times does the letter B appear in Blueberry? This time, two with a period. So, it's answering it, but no extra context, which is interesting because the older models love to give additional context. When I put it on GPT5 thinking, it still gave me two at position one and five. Just for fun. Let's do pro cuz this is research grade intelligence. Oh, this one's going to reason for a while on it, isn't it? And even with pro, it got it right. So, I can't get it to tell me there is three bees in Blueberry, but that also might have been happening on the day the model picker wasn't working or something. I don't know.
So, when it comes to the models not being as accurate, I got it to be accurate on the doctor question and why the doctor didn't like its patient. I got it to be correct on the elevator riddle about the person living on the top floor. It did get the blueberry answer right, so it was accurate three times, but it also totally botched the math problem. And it couldn't figure out the cup problem that had the hole on the wrong side and the closed part on the opposite side. and then I couldn't really evaluate whether or not it found up-to-date information about companies in your local area. It's kind of like a three and three. And the thing is, when it comes to accuracy, it's a 5050 chance that you're actually getting an accurate answer. Then I have to say that the complaints about accuracy are valid because 50% of the time I might get an inaccurate answer. So now I just need to assume that it might give me an accurate answer and I have to always double check myself.
Just for good measure, I asked how many Rs are in the word strawberry and it got hung up again. Like it just decided not to answer and think forever. And I don't know why it does this so often. I don't remember the old models doing this. But I'm going to go ahead and copy my prompt here and let's start a new chat and see if it'll answer it for us this time. Okay, so it says three. So at least it got it right. So, if you are keeping track at home, my score here sort of ended like four to two. Four really, really valid concerns versus two things that I think are pretty much solved now.
In my opinion, the four things that I think were valid were one, I agree, it was super overhyped before it came out, which left people feeling underwhelmed when they finally launched it. Number two, I think it's really, really annoyingly slow sometimes. Even when you ask it really easy questions, sometimes it feels like it thinks too long. And I've even had to like open new chats and ask it again because it seemed like it was just endlessly thinking for whatever reason. I don't think it's the best model at code. I think Opus 4.1 is better at code. And I think 03 Pro did a better job in the example I just did now. But Opus 4.1 is probably better all around at code cuz when I did a test in my last video of Vampire Survivors to see which one did it better between Claude Opus and GPT Claude I think did it better. It also made a better productivity app between the two. Opus just seems to create better apps in my opinion still. And then the accuracy issues they're definitely valid. Although some of the ones I tested I didn't get the same inaccuracies. I did get a few inaccurate responses and if it sometimes is inaccurate then it makes me think I need to question every response it ever gives me. So yeah, the accuracy concern is valid as well.
The two that I think that are actually fixed and that aren't issues anymore are probably the legacy issue. People wanted 04 back and they were pissed that it was gone. It's back now. You can use it. So that's an issue they solved. They fixed that. And then the personality issues. If you really liked 40 better than the current models, well, 40 is back. I also didn't seem to find the same like personality deficiencies in GPT5 that other people were pointing out. To me, it seemed almost on par with 40. Maybe a little bit shorter with its responses, but still pretty on par. So, I don't think the personality complaints are super valid, especially now that you got 40 back. But most of the concerns that people are complaining about are still valid as of this recording.
I think DD here on X sums it up pretty well. After thorough evaluation of chat GPT5, these are my realizations. Clot is pretty freaking awesome. I'm a lot less concerned about ASI/T the singularity/ AGI 2027 or whatever doomy scenarios was bouncing around because well, the model wasn't as good as everybody thought it would be. And we're not as far along as a lot of people thought we would be when this release came out. GPT5 is about lowering costs for OpenAI, not pushing the boundaries of their frontier. I'd say yes and no. I do think they are trying to lower their costs with their router and trying to route to the cheapest model that they can get away with, but I also think that that was a benefit to the user. That was trying to make it less complicated for users. Sam's Death Star pre-launch hype image was really about the size of his ego and had nothing to do with the capabilities of GPT5. Yeah, I mean hyping it in the ways that he hypes it and then sort of delivering an underwhelming kind of presentation. Maybe don't hype it so much and people won't be so underwhelmed when you finally launch it.
Luckily, OpenAI is listening. They are taking everything people are saying into account. So, honestly, the future that chat GPT I think is shooting for is a very customized model where the chat GPT that I'm using feels completely different than the chat GPT you're using because we've customized them to the way we like. Maybe I want mine to be more warm and friendly and, you know, give me props and stuff like that more often. Stroke my ego a little bit. And maybe you want it to just give you one-word responses. Over time, everybody's going to have a chat GPT that's sort of tailored and dialed in to their wants and needs and how they like to use chat GPT. And I think that's where things are going. So, a lot of the gripes people have, I think, are going to get ironed out because you're going to have these models that are your model. They've had some missteps during this 100% but it is a stepping stone on where they want to get and what they have now is not the future final version of what they're trying to create and a lot of the current concerns are valid and I'm happy that OpenAI made some of the moves that they made so you can go back and use older models again if you want.
But that's pretty much my thoughts on where things stand with GPT5 right now. I've consistently said I was a little bit underwhelmed, but also found some of the stuff they showed in their demos very impressive. I think it's okay to have a little bit of both. Be slightly underwhelmed, but also think some of what they did was impressive. And that's sort of where I land. I'm not a super binary person where I'm going to say the new Chad GBT completely sucks and I'm not going to say it's the best thing ever. It's somewhere in the middle. There's really impressive stuff about it and there's really crappy stuff about it. And hopefully over time more of the crappy stuff gets worked out and more of the really impressive stuff sort of bubbles up and becomes better. That's what I'm hoping for and that's what I feel like consumers who use this stuff should also be hoping for because it ideally just makes our lives better. And yeah, that's what I got for you today. Hopefully you enjoyed this. Hopefully you feel a little bit more looped in on what's going on with all the GBT drama. And we'll keep following along and making updates as more of these stories unfold. If you want to stay uptodate with the latest AI news and get demos and tutorials of the latest AI tools, make sure you like this video and subscribe to this channel. It will make sure more videos show up in your YouTube feed about all of the latest cool advancements in the world of AI. Thank you so much for hanging out and nerding out with me today. Really, really appreciate you. Hopefully, I'll see you in the next one. Bye-bye.
Thank you so much for nerding out with me today. If you like videos like this, make sure to give it a thumbs up and subscribe to this channel. I'll make sure more videos like this show up in your YouTube feed. And if you haven't already, check out futuretools.io where I share all the coolest AI tools and all the latest AI news. And there's an awesome free newsletter. Thanks again. Really appreciate you. See you in the next one.