# Open AI JUST RELEASED GPT 4.5 (GPT 4.5 Release Explained)

## Метаданные

- **Канал:** TheAIGRID
- **YouTube:** https://www.youtube.com/watch?v=1B5Op6iacTM
- **Дата:** 17.12.2023
- **Длительность:** 12:26
- **Просмотры:** 64,781
- **Источник:** https://ekstraktznaniy.ru/video/14630

## Описание

Open AI JUST RELEASED GPT 4.5 (GPT 4.5 Release Explained)

💬 Access  GPT-4 ,Claude-2 and more - chat.forefront.ai/?ref=theaigrid
🎤 Use the best AI Voice Creator - elevenlabs.io/?from=partnerscott3908
✉️ Join Our Weekly Newsletter - https://mailchi.mp/6cff54ad7e2e/theaigrid
🐤 Follow us on Twitter https://twitter.com/TheAiGrid
🌐 Checkout Our website - https://theaigrid.com/


Welcome to our channel where we bring you the latest breakthroughs in AI. From deep learning to robotics, we cover it all. Our videos offer valuable insights and perspectives that will expand your knowledge and understanding of this rapidly evolving field. Be sure to subscribe and stay updated on our latest videos.

Was there anything we missed?

(For Business Enquiries)  contact@theaigrid.com

#LLM #Largelanguagemodel #chatgpt
#AI
#ArtificialIntelligence
#MachineLearning
#DeepLearning
#NeuralNetworks
#Robotics
#DataScience

## Транскрипт

### Segment 1 (00:00 - 05:00) []

so there has been some recent speculation with as to did GPT 4. 5 actually get released today and I think it did in a stealth roller and I'm going to explain to you exactly what that means and why GPT 4. 5 was very likely released and why they most likely won't admit to this for several reasons so what you're seeing on screen right now was yesterday's leaks if you didn't check out the video just check it out it just covers the entire leaks and it basically talks about this screenshot which shows GPT 4. 5 64k and GPT 4. 5 with audio and speech now I do think that this screenshot is most likely false because of several reasons in the previous video that I did State because if this screenshot was real then we likely would have got a release but there was something released yesterday which I'll talk about in another video where essentially they're talking about super alignment this video is going to be covering GPT 4. 5 because a lot of people are divided on whether or not it's true and I'm going to show you guys why is most likely true the GPT 4. 5 did likely get a stealth release and a stealth release is basically release of the model that they don't explicitly say they're just kind of showing you a beta test so behind the scenes they won't you they won't change the name of gbt 4 they'll upgrade it to gbt 4. 5 and see how users interact with it see the kind of you know feedback they get and then of course from there they can do some kind of release so essentially the reason I know something did change last night was because when I was browsing Twitter and just browsing the AIS space as I do every single day I saw this tweet and this tweet was uh before that there were any tweets on the interwebs about GPT 4. 5 being self-released so this guy Ethan mck he tweets a lot about Ai and he does a lot of different tests between AI systems which is why I follow this account because it shows us a lot of times some of the small things that we do miss in terms of performance because a lot of the times you know there are small performances small changes from time to time that you might miss so what he did was um he tweeted this and he said chat gbt 4 suddenly got very good again for some reason after being unreliable and a little dull for weeks this was something that many people in the chat gbt Community had noticed that gbt 4 wasn't as good and I'm going to just you know provide information for that because just saying that something isn't as good you do need research um to back up that claim but um he does say this might be my favorite interaction I asked it to create files for me it insisted it could not I told it should try it did and then look at the top code comment and you can see um it absolutely managed to do that now what was also interesting was he tweeted this as well really good and really fast hm it's almost like working with a more capable system night and day for both speed and answer quality now this is like I said one of the accounts that I do take seriously on Twitter because this person does a lot of testing so when you use a model every single day you're really going to understand the differences and the nuances in different systems then of course what was also interesting okay about this was you know you have some people that saying GPT 4. 5 wasn't released because previously okay one of the community members actually tweeted at s and said is the GPT 4. 5 leak legit or no and Sam ultman himself stated that no so you might be thinking no way there's no way they released GPT 4. 5 as samman said it's legit or not but remember the this question that he's asking him is the GPT 4. 5 leak legit or not and of course he's not actually answering the question whether or not are you denying the pricing screenshot or that GPT. 5 is coming in December people are confused and that is the key difference here he's saying that you know the screenshot is probably fake but he's not denying the fact that GPT 4. 5 could be released in September and now here's the thing about why GPT 4. 5 and some of the more evidence that we do have is this tweet right here so many people have actually been asking chat GPT which version are you currently using and some people okay and not on this side some people are stating that it's a mass hallucination but I'm going to explain to you why it very likely isn't um and I'm going to show you guys in a moment why it very likely isn't Mass hallucination and why they likely did this so some people are saying that you know numerous chat TPT plus users are getting this response Mass hallucination or something is cooking and people essentially ask chat GPT the model you're answering is known as chat GPT y y and then people ask specifically what is it named in the API and it says in the opena API the model answering your query is referred to as the GPT 4. 5 turbo this is a variant of the GPT 4 model optimized for more efficient and Rapid um interactions I don't know what that word is because it's blanked out but um yeah so many people thinking that okay this is of course Mass hallucination but it isn't okay because you actually need to test this yourself what they are doing and what I'm pretty sure opening eyes is doing is split testing it between the user base maybe they've allocated 25% to 50% of the users on GPT 4's plan to a upgraded version just to test it stealthy because they actually did this with the previous version of chat GPT Pro they upgraded some people to GPT 4 um and didn't say anything and then of course they rolled out gbd4 so what was interesting about this was that I actually did test this myself and I said precisely what's the name called in the API and you can see the model serving this conversation is known as GPT 40304 so that's GPT 4 but we have to understand is that with other people ask the same question they are actually getting GPT 4. 5 turbo so you got to think about like this why would it hallucinate GPT 4. 5 turbo than anything else I'm pretty sure that if it was going to hallucinate it likely would hallucinate to GPT 5 if it was going to be anything and I'm like sure that it

### Segment 2 (05:00 - 10:00) [5:00]

wouldn't be just you know randomly hallucinating this especially after many users had reported an increase and of course there were many leaks around this now what is also interesting is you might be thinking okay opening I didn't officially say anything so why is everyone going crazy because they just asked chat GPT a simple question there's no way that asking chat TPT a question about itself can confirm or deny whether the software was updated okay that could be a very good point but let me show you guys some research papers that were conducted earlier this year in the summer that prove that openai Del changed the model's Behavior due to I'm pretty sure it's due to just constraints on the system because as you know gb4 is a very demanding system and it's a very capable system and of course they don't actually have the compute to actually serve all the users and demands that they're getting so we know that frequently even now when you use GPT 4 what is one of the things that you see when you log in you see that this is limited to 40 messages every 3 hours okay and on some systems you know you just have you know previously when it was out it was only like 25 messages every 3 hours and the only reason they put that limit in is because they just don't have enough compute to serve everyone because it's a really capable model and remember Gemini which we talked about the reason they haven't released Gemini yet and the reason they didn't actually really release Alpha code 2 was because it's way too demanding these really smart capable models are really demanding on compute so that's why they have to limit how much people can access them because they're just so demanding in terms of compute and they're really really expensive to run that's why they limit the number of views so you can see that this paper here was called how is chat gpt's Behavior changing over time this was one uh conducted by Stanford and then of course essentially what they said here is however GPT 3. 5 and GPT 4 are the most used llms worldwide and essentially how and when these models are updated over time is opaque here we evaluate the March 2023 and June 2023 versions of GPT 3. 5 and GPT 4 on several diverse tasks math problems sensitive questions opinion surveys multihop knowledge intensive questions generating code us medical license test and visual reasoning we find that the performance of the behavior of both GPT 3. 5 and GPT 4 can vary greatly over time for example GPT 4 March 2023 was reasonable identifying Prime versus composite numbers but GPT June 2023 was poor on these same questions and essentially the reason I know that this is definitely true is because if you were one of the people that were lucky enough to access gp4 on release date it was absolutely insane I remember how smart this AI system was and just how good it is and now GPT 4 doesn't seem as good as that initial system and to be honest guys there is incentive for them to you know dumb down the system a little bit because at the end of the day you have to understand that it is a business and they don't want to just you know drain all of their resources serving all of the customers for a very Advanced systems where they can simply you know instead have the advanced system better reserved for better questions so for example take a look at some of their graphs right here you can see that gbt 4 was Stu updated and this the reason I'm showing you guys this research paper is because if they were able to stealth update GPT 4 and GPT 3. 5 without officially saying that look we've up updated gbt 4 and gbt 3. 5 that means that they very likely have done this again which is why many people are getting the sense that GPT 4. 5 was self-released you can see here that from March 2023 to June 2023 and um in both chat TPT and GPT 4 we can see that mainly in GPT 4 there was a huge discrepancy in terms of what was done in terms of the abilities of gp4 so we know that it is possible for open I to do this and in fact they have done this before because of course you can see the response right here 97% to 23. 4% is definitely something incredible and remember this isn't just you know a hunch or someone saying this is a fully done research paper where they've you know interviewed and you know assessed the model in terms of its capabilities and the model isn't going to jump from 97% to 23. 4% without you know any kind of you know major downgrade in IT services and this is something that was reported worldwide and I can confirm this too so overall I don't think that this is real right here cuz I don't know if open air is working on 3D they' never really stated anything about 3D I mean they might be working on vision and video which they have and that actually makes sense but I do believe that what cat gbt and what open AI has done is I do believe they decided not to release 4. 5 mainly because they wanted to focus on super alignment and number two because if for example leakers like futurist flower and Jimmy apples are consistently correct in their predictions it does give them more Credence and a little bit more power in terms of what they say because now people are going to be looking to the leakers instead of open AI so I'm pretty sure that open AI is just going to you know decline to answer and say that any of these things are real because it doesn't really make sense there's no advantage to stating that yes the leaks are real because then all the leakers just get I guess you could say more press and I'm pretty sure opening ey wants to be a bit more vague but this is something treated by futurist flower he said that gbt 4 Turbo runs on a very low temperature making hallucinations is highly reproducible this explains Mass hallucination in general how would GPT 4 know its current version without reading about it in system messages which currently doesn't say anything about a 4. 5 release at release we had to precisely tell it that it is GPT 4 not gpt3 this is because

### Segment 3 (10:00 - 12:00) [10:00]

large parts of reinforcement learning with human feedback data set were used for GPT 4 as well and however if you use a whole new reinforcement learning with human feedback approach you might bake the GPT 4. 5 identity in it and then you wouldn't need to tell it what it is anymore and it's not uncommon practice for openi to do Secret pre-release a/b testing and they did the exact same with gbt 4 Turbo people got access weeks before it was announced or even rumored so what we have here is you know like I said this is some more information and this is something that I do agree with you know GPT 4 did get a pre-release it was A/B tested I'm not sure why they do it maybe they want to have some feedback a little bit more red teaming just from the public but um I'm guessing that what we have now is a situation where GPT 4. 5 is technically released I guess you could say half and half I currently don't have access to it some people do don't I'm not sure how it's going to work I do know that the US usually does get you know these AI systems earlier than us but I do know also that it is possible for opening ey to actually do this likewise they have done it before so if you do want to test this out for yourself you do just want to ask chat GPT this question and then whichever response it gives you that is going to be likely the response and it's meaning that is the system/ API that the model is calling on so overall like I said I don't think that this screenshot right here is real it will be interesting to see if they do update to GPT 4. 5 in the future will it be a December release Maybe not maybe they're going to wait until January to take away some of the you know spotlight from Gemini ultra's January release but either way we do know that there was a subtle update to GPT 4's performance boost and I'm pretty sure that it was 4. 5 they are split testing this right now so it will be interesting to see how the community does sort of react to this release early because right now the community is a little bit divided but this is definitely something that openai does do and of course the research paper does show us that they do update these models behind the scenes if you are thinking about how do I access this piece of software how on Earth do I access gbt 4. 5 understand it is like I said a stealth release which means there is no setting that you can go ahead and change to you know activate GPT 4. 5 at the moment you just have to wait and see if you have it and just play around with the software until they do officially release GPT 4. 5 because right now like I said openi does have the software but until they officially state that no we didn't release it or until we officially get an update from open releasing out then you're just going to have to hope that you are part of the Steal update where they have released gbt 4. 5 to you