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

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 uh I'm very excited for that and they also show that we are getting 50% cheaper API cost for GPT 40 uh and it's especially better at vision and audio understanding compared to existing models so I went ahead i wrot a l 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 uh 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 to today but uh 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 uh for most use cases right and uh 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 uh there was some voice interruptions that looks to pre working pretty good uh yeah voice input and output uh there were some emotions so actually in the live stream they try to change the emotions of the the voice in real time so I thought that looked pretty interesting we going to test that of course later and latency looked pretty good tone adjustment yeah that's kind like testem motions 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 uh on paint here just going to paint something and we going to try it out and 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 get 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 and but now that I think about it if everyone get 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 40 llama 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 GPD 40 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 just going to go in 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 the thing I created here uh 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 uh I put in these images from kind of my previous videos so this is the images from the mixture of models right uh 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 shat to kind of give us answer so let's just do uh image description feed in all the image description 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 set check the speed now in kind of real time here uh I thought i' B interesting uh yeah that's pretty good so you can see analysis for uh 27.png so 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 uh 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 uh let's zoom in a bit mixture of models uh 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 explain every single architecture here in kind of the same format so that's pretty cool the mixture of model system utilizes multiple AI model 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 I 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 are comparing GPD 40 with gp4 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 toen 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 so that's pretty impressive that's kind of what I advertise so I'm pretty impressed by that five times f faster that's a big difference if we have the same reasoning capabilities I think that's a huge 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 uh in the previous videos okay so this is kind of the marble problem so this is a problem we uh borrowed from um Matthew Burman I guess You' 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 uh assume the loss of physics uh on Earth small marble is put into a normal cup the cup is placed upside down on a table uh 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

Channel: All About AI

Views: 57,964

Keywords: gpt4o, gpt-4o, gpt 4 o, gpt4 o, openai, openai gpt4o, openai gpt-4o, chatgpt

Id: 701JD6mDMJw

Length: 13min 12sec (792 seconds)

Published: Mon May 13 2024