hello hope you are doing well so yeah thought I just had to do a quick reaction video to the open AI spring update and the release of their GPT 40 models that was pretty interesting right uh if you watch it I think you just watch it in replay if you didn't catch it but basically uh they are introducing a new flagship model that can reason across audio vision and text in real time uh I thought it was pretty cool I'm a big fan of uh the audio part of chat GPT I use it a lot to try to learn stuff so so yeah pretty excited for me but of course we have to test it out first o STS for om a step towards much more natural human computer interaction H kind of what I found most interesting was kind of the low latency so this is an average of 320 milliseconds uh similar to a human response time in conversation so that is a big deal that is a big deal because yeah as you know on this channel we have been working on creating low latency uh

it would be really cool if we can get this kind of latency on a simple API call. I'm very excited for that, and they also show that we are getting 50% cheaper API cost for GPT 40. It's especially better at vision and audio understanding compared to existing models. So I went ahead, I wrote a script, so we actually going to test out the image functionality. We're going to put in some images, ask it about it, use GPD 40 to analyze the images, use GPT 40 to actually get a response from the images we put in this description, because we can't try out audio yet, because I had a look at the documentation here. You can see accepting text or images input and outputting text, so we don't have that available yet. That's a shame, because I really wanted to try that today, but hopefully soon. So it's supposed to be two times faster, yeah, 50% cheaper, and it has a I think it was 128k token context.

I think that's fine, for most use cases, right? And I wanted to just take a quick look at my notes here, so I took some V notes during the live stream. So yeah, we can see that there were some voice interruptions that looks to pre-working pretty good. Yeah, voice input and output, there were some emotions. So actually, in the live stream, they tried to change the emotions of the voice in real time. So I thought that looked pretty interesting. We're going to test that, of course, later. And latency looked pretty good, tone adjustment, yeah, that's kind of like testing emotions. And I think it actually could sing the output. So I thought that was interesting. And yeah, voice input choices. So I think you can kind of, if you say something in like a sad voice, I think it kind of picks up on that. We will have to test that out later, of course. And of course, images and video input, we're going to do a quick test on paint here. Just going to paint something and we're going to try it out.

I thought also the desktop app could be interesting to just have in the background right when we are actually when you're working on some code or something just have your the desktop app from open AI in the background you can just talk to it. Uh, I really like the app one on my iPhone, so I use that a lot. So yeah, some pretty interesting stuff here. I'm not going to spend too much time today, uh, going through every single feature. I guess a lot of other YouTube channels or something. I'm going to do that Twitter maybe, uh, but I think we're going to come back when we have more practical use cases. So that's probably going to be on Wednesday. I'm going to come up with something cool. And yeah, there's a bunch of evaluations here, uh, looks pretty strong overall. You can see it's pretty much on top here. And another thing I just wanted to mention is that they said that they going to bring this GPD 40 to all free users. I think that's a big deal, to be honest, because remember

when everyone gets access to this, I think that's a yeah. I think that's kind of more important now that I think about it, and during the live stream, because I was looking out for API stuff. But now that I think about it, if everyone gets access to this during the free version, that is a pretty big deal considering like, uh, yeah, Claude anthropic, and of course the other development. Now, if you have a free GPT-3, I don't know, so there's something there that's going to be interesting to hear about in the next few days, what people think about the free version of GPT-3, and what do you get when you pay. I didn't really catch anything around that, but like I said, I haven't done any research. Now, so I think we're just going to go into some testing, so that was kind of my first reactions. So yeah, I could be wrong on some things, but I haven't done. I just wanted to give my first, uh, reaction here. So, uh, I just wanted to show you kind

know the thing I created here. Yeah, I added some audio, but we can't use that. I found out, so we're not going to be able to test that. So let me just show you here. So, I put in these images from kind of my previous videos. So, these are the images from the mixture of models, right? The three architectures and kind of the intro slide. So, I think we're going to feed all of these slides into the image analyzer from G with GPT-40. And we're going to use a GPT-40 chat to kind of give us answers. So, let's just do image description. Feed in all the image descriptions. Explain the system for the image description, both structured format, and print out the final explanation. So, yeah, let's just run this. I think so. I guess there's no streaming now, but we can also check the speed now in kind of real-time here. I thought it'd be interesting. Yeah, that's pretty good. So, you can see analysis for 27.png.

this is kind of our first slide mixture of models looks pretty strong. We have the three architectures here. We have image 28 PNG. We get a summary of each image pants. Yeah, this looks pretty good. 29 components. So this is the DU poly architecture, and I guess the first one was the king architecture. Here is the Democracy architecture, and yeah, I think I think this is pretty quick. And here's kind of our final explanation system over you. Let's zoom in a bit. Mixture of models leverage has the wisdom of a crowd principle to achieve Optimal Solutions by using multiple AI models. Yeah, that's pretty good. Three different architectures perfect. And here kind of explains each architecture. So we have use query. Yeah, this looks strong, very strong. I just wanted to see if there was some summary here. Yeah, so I like the format. It kind of explains every single architecture here in kind of the same format. So that's pretty cool. The mixture of model system utilizes multiple AI models to generate diverse responses to user query.

depending on architect you use these respons are either refined by the leading model the king uh discussed by co-founder models the do poly and voted on by the models themselves the Democracy this approach ensur that the final response is well rounded and deeply considered benefiting from the strengths of multiple models yeah that's a good answer because remember this is not something it has seen before right so this is brand new stuff and I think it did a very good job in analyzing all images so we're not going to spend any more time on this now because of course we're going to do this in the future so I just wanted to run through a quick script so yeah seems to be working pretty good so we used base 64 here so all I had to do was to feed just put all my images in one folder and we just pointed to that folder PAAD that was pretty easy uh let's do some more testing so uh I am in a playground here you can see we have selected 40 uh

just wanted to do something uh so this is going to be a bit of a live live stream kind of so let's do like a triangle and let's do like six two and four right and let's try to screenshot this yeah that was an image of me and let's save this and upload it and ask can it do any calculations on this what math calculations can you perform on this image so what I was thinking is Pythagoras right so let's see what happens here so I guess we can also check out the latency now should be pretty pretty quick uh yeah that's that's pretty fast that's much faster than gbt 4 right uh here are some calculations we can perform verify the triangle in equality theorem okay uh check if it's a right triangle using pagor theorem okay that's good this is not a right triangle calculate the area yeah looks pretty cool uh fuel calc be performed so yeah I guess that was a pass uh let's just try something different let's check out the

speed a bit because now we have this latency down here right, and we can see the token count and um, yeah, the latency. So let me set something up here. So now you can see we kind of are comparing GPD 40 with GPT-4 Turbo, right? And we're going to go right to three paragraphs about the life in Paris in the 1800s. So let's check out the latency now. Wow, that's a big difference. Kind of reminds me almost of Gro. This, that was pretty cool. And let's do a calculation here. Uh, tokens per second. I think we're looking at something like 100 and oh, it's slow. God, we just have to wait. And then we can do it. 23 seconds. That's four times faster, at least almost five times faster. And it only and it brought less tokens. Let's do a quick calculation here. So let's just do it like this. So latency, latency, and tokens. Okay, so you can see GPT-40 110 tokens per second, GPT-4 Turbo 20. So that's over five times faster.

that's pretty impressive. That's kind of what I advertise, so I'm pretty impressed by that. Five times faster, that's a big difference. If we have the same reasoning capabilities, I think that's a huge deal, right? So yeah, pretty exciting. So 110 tokens per second. Yeah, that's good. Let's also do some small logical test. So let's just take some of the problems we had in the previous videos, okay? So this is kind of the marble problem. So this is a problem we borrowed from Matthew Burman, I guess. You've seen his YouTube channel. I just asked him if we can test it out. It's not his problem, but yeah, just nice to ask him. He's a great guy and a good channel, so go check him out, I guess you probably have heard. So this is assume the loss of physics on Earth. Small marble is put into a normal cup. The cup is placed upside down on a table. Someone takes the cup without changing its orientation and puts it inside the

microwave where is the marble now so of course the C the marble is uh left on the table right or something like that so let's compare this with the GT40 and gp4 turbo and see if it can solve this let's just check the settings here uh yeah they are the same okay so let's run it okay so GPD 40 says the marble is now on the floor of the microwave uh the marble remains inside upside down capup now sitting on the microwave tray so nope either gp4 o or gp4 Turbo got this one correct let's try another problem write 10 sentences that ends with the word apples so let's try that uh Apples Apples Party apples apples apples apples apples apples so it got nine out of 10 let's take GPD for Turbo H 1 2 3 4 5 6 yeah that got every single one correct so that one I have to give to GPD for Turbo I don't know I can't really say

much about the performance of GPT 40 now it's too soon for me so just fun to do some logical testing right but yeah I think that's actually it for me now uh of course we're going to follow up on Wednesday with a video on GPD 40 when I kind of had the time to look more into it but this was kind of just my first reaction so I thought it was pretty cool I like the stream uh I thought it showed up some pretty interesting things uh but of course we have to explore more uh there's a bunch of examples here you can take a look at I thought I was pretty cool too but yeah uh exciting I think so what do you think let me know in the comments uh I think it was pretty cool that if everyone gets this for free uh but what happens when we pay well let's just wait and see about that uh other than that yeah have a great evening and we speak again on Wednesday

OpenAI GPT-4o | First Impressions and Some Testing + API

Welcome to my channel All About AI =)

Website:
https://www.allabtai.com

How you can start to use Generative AI to help you with creative or other daily tasks.

So I aim to bring Generative AI to everyone.

- AI Engineering
- ChatGPT
- GPT-4
- Midjourney
- Stable Diffusion
- Python
- AI Automation

We will answer all your questions about ai, genereal questions about the future of ai and the ethics of artificial general intelligence.

Follow our channel if you want to get going in this space =)



OpenAI GPT-4o | First Impressions and Some Testing + API

OpenAI GPT-4o | First Impressions and Some Testing + API

스크립트