# OpenAI's NEW MODEL, LLM'S Beat WallStreet , Google Takes The LEAD, Googles Major AI Mistake!

## Метаданные

- **Канал:** TheAIGRID
- **YouTube:** https://www.youtube.com/watch?v=h9a9JDU-xLc
- **Дата:** 29.05.2024
- **Длительность:** 18:46
- **Просмотры:** 29,584

## Описание

Join My Private Community - https://www.patreon.com/TheAIGRID
🐤 Follow Me on Twitter https://twitter.com/TheAiGrid
🌐 Checkout My website - https://theaigrid.com/


Links From Todays Video:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4835311
https://chatgpt.com/g/g-9P3sIn487-financial-statement-analyzer
https://x.com/janleike/status/1795497960509448617
https://www.ft.com/content/34a7a082-e685-4e02-bca7-61ff89d99ed2
https://openai.com/index/openai-board-forms-safety-and-security-committee/
https://x.com/AndrewCurran_/status/1795453553328247079/photo/1

Welcome to my channel where i bring you the latest breakthroughs in AI. From deep learning to robotics, i cover it all. My 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 my latest videos.

Was there anything i missed?

(For Business Enquiries)  contact@theaigrid.com

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

## Содержание

### [0:00](https://www.youtube.com/watch?v=h9a9JDU-xLc) Segment 1 (00:00 - 05:00)

so I'm not going to waste any time one of the craziest stories that actually was there today on May the 28th 29th there was actually a new post by the openi board forms Safety and Security committee it says this new committee is responsible for making recommendations on critical Safety and Security decisions for all open ey projects recommendations in 90 days so essentially it says today the open ey board formed a Safety and Security committee led by directors Brett Taylor Adam D'Angelo Nicole cigman Sam Altman and this committee will be responsible for making recommendations to the full board on critical Safety and Security decisions for open AI projects and operations now this wasn't the main thing this wasn't the actual main piece of news I know some people did see this and that's what they thought this was but the main piece of news actually was is that they kind of slid this information into this post and they didn't really make a post about this but they says open eye has gun training its next Frontier Model and we anticipate the resulting systems to bring us the next level of capabilities on our path to AGI while we are proud to build and release models that are industry leading on both capabilities and safety we welcome a robust debate at this important moment so I think the wording is very interesting because I know one thing that open ey does do is that they're very specific with their wording and because of this you can kind of interpret and realize what certain things are going to be released slash that there in the next kind of I guess you could say things and one thing that is here that they've said is that they've begun training its next Frontier Model and this is now a very important point because them stating that they've recently begun training its next Frontier Model means that whatever model is coming next is very likely going to be a system that we haven't anticipated quite like how we didn't anticip P GPT 40 them stating that they've recently begun training their next Frontier Model is vastly different and this is also vastly different because of course you have to understand that when GPT 5 was in training it wasn't really announced but openingi did actually speak about this in the media they actually did say that you know we've begun training GPT 5 and I remember this because I recently made a video and because I made a video on November the 14th 2023 and you can see it was about GPT 5 now in training another video you can see right there open ey is officially training GPT 5 so right now A lot of people are currently confused they're wondering is open ey training GPT 5 I thought they you know were trying to train a different Frontier Model did they make a mistake no GPT 5 has actually been you know it's very near its release cycle GPT 5 actually recently did undergo red teaming and red teaming is a process that does last for several months so essentially they've been training gbt 5 in November then recently a couple of months ago there were rumors for um you know acceptance email for GPT 5 red teaming and a lot of different people were stating this so this is something that is widely accepted at this moment in time could be false news cuz obviously openi doesn't exactly State everything that's going on they'd rather just release it in something that's like not shocking but just something that they have access to and since red teaming is the phase where they kind of test the model and they kind of see what they can do wrong with model so for example they try and jailbreak the model they try and make the model do bad stuff that you know Bad actors might try and do and then of course they tried to fix that model before the official release so what I'm basically trying to teach you guys is that they are not just now training GPT 5 what they are training is likely a new kind of AI model but the thing is we don't know what that kind of AI model is because we haven't been told what we do know is that the model is likely to you know have a huge set of new capabilities but it's very hard to predict what those capabilities are because we just aren't sure yet and I personally do think that it is something to do with agents because if it's not GPT 5 the only other thing that I can think of that would be really useful is of course these agent like systems so it will be interesting to see what open AI does and this is of course an article from Financial Times where they say openi begins training next AI model as it battles safety concerns and this article actually reveals quite a lot about some of the inner workings of openi within regards to some of the recent news that we do know about and you can see right here it says starting to produce a new AI system to bring us the next level of capabilities so this isn't just GPT 5 that's going to be smarter this is a system that will bring us the next level of capabilities which of to which I'm not even sure it could be embodiment it could be planning it could be you know just a whole huge host of other things but you know you can see right here that this document this article actually contains a decent amount of information about super

### [5:00](https://www.youtube.com/watch?v=h9a9JDU-xLc&t=300s) Segment 2 (05:00 - 10:00)

intelligence and I think the reason that this is so interesting is because what they actually talk about here with regards to Super intelligence it kind of gives us an Insight with as to why some of the superintelligence members kind of like left open a ey so you can see right here it says Anna mun Anna maanu opening eyes vice president of global Affairs told the financial times in an interview that its mission was to build artificial general intelligence capable of cognitive tasks that are what a human could do today and it says our mission is to build AGI I would not say super intelligence is a technology that is going to be orders of magnitude more intelligent than human beings on Earth so this is why I think this leads us into the point where you know we had literally last week where several members of open AI super intelligence team the team was basically just disbanded after many members over you know quite a tumultuous period during the Sam Alman board Fiasco whatever honestly went on and there's a lot of information regard regarding that all of that information being given out I think now we can truly understand why it's because openi are trying to shift their focus onto I guess you could say just building products and I think from a business perspective I think what open ey is just betting on is that some of the other companies are just going to solve super intelligence and openi is literally just going to try and build AGI and use that AGI in the future to maybe solve that alignment problem or potentially just use that to just actually focus on Building Products whilst other companies can literally focus on the super alignment and I'm going to show you guys why that is probably the case with regards to some recent announcements and if you remember during the well presentation at the Microsoft keynote Kevin Scott implied that GPT 5 it's the name had begun its training run on the hardware Microsoft had just finished building for open Ai and essentially stating that this model is going to be there at the end of the year so I think with all of the information and one of the things that you know now make sense is that you know the super alignment team were basically stating that you know we didn't even have access to any of the computes opening I said that they were going to give them 20% of the compute capacity and you know Jan like someone that was working at opening at the time basically said we were trying to use this compute for super alignment but it just wasn't allocated to us and we weren't able to you know do any of the meaningful work that we really wanted to so that's why I'm stating that basically you can see right here that they're now stating that you know whatever model they're now training Microsoft has just finished building it and it's going to be there in its crazy State and we really don't know what this newer model is and of course this model will probably not be named gbt 5 and they called it the next sample and Sam has said they might be moving to a new naming system so it's going to be interesting to see if it is GPT 5 or if it is you know a future kind of system if they kind of changed GPT 5 because I mean either we get gp25 and agentic system or maybe they just found a new way to completely change the way that we use AI systems so it might not just be an actual GPT like system it might be some kind of different agentic system that has a new kind of architecture because open ey has some of the best researchers on the planet and I'm sure very sure that they're probably working on something that is very effective and here you can see this is a tweet from Jan like so this is someone that was working on the super alignment team and you can see he says I'm excited to join anthropic AI to continue the super alignment mission my new team will work on scalable oversight weak to strong generalization and automated alignment research and basically these are the things where they published research papers that were very early in terms of the development on these actual areas and these were areas that were areas of research concerning the super alignment team and these were areas where they kind of made a very decent amount of progress but of course they weren't able to finish their work so it seems that anthropic are going to be getting some really Superior talent in terms of AI researchers because the team from open AI might even be getting SATs now I'm not just stating that but I do think that whatever sukova is going to do he can pretty much have his pick of the top a companies maybe he's going to go back to Google to anthropic but it would be nice to see what he's going to be doing because he did state that he would update people very soon so I'm interested to see what now comes out of this alignment SL research team because anthropic have recently published some new research in which the Golden Gate claw thing was really interesting and there was a lot of information that was revealed after understanding what kind of goes into the model another piece of news that was pretty interesting was the fact that Microsoft being investigated over the new recoil AI feature that tracks your PC PC's every move now if you aren't aware of this recoil feature this is basically a feature that is really fascinating because it brings us into a new area of privacy that is quite hard to categorize so Reco is basically a feature that is essentially you know how my computers the mouse is moving around right now I'm scrolling up and down the page I'm basically just doing a bunch of different things on my computer so what

### [10:00](https://www.youtube.com/watch?v=h9a9JDU-xLc&t=600s) Segment 3 (10:00 - 15:00)

you can be basically do with recoil is that you can search your history of everything you've done on your computer and I think this is you know it serves two purposes because one of the things that I think it might serve is that maybe they're going to use that data to train agents for computers I think that's going to be pretty insane and one of the features I think it's going to unlock is of course the ability to just you know get a greater understanding of the things that you do a lot of times we lose things on our computers we are searching for a certain file and you're able to just search things with a literal just text description so you don't need to file search and file search on Windows is actually pretty awful anyways but this is interesting because they're being investigated over this so of course you know a lot of different countries like the you know not countries but you know a lot of different nation states I guess you could say or different you know areas where different policies are in place they have different rules and regulations and you know the European Union is really focused on privacy I just know that is why a lot of these newer AI systems a lot of the time they aren't even available to certain regions and I know that is frustrating but you know it is good in some cases because it does help you from a lot of the Privacy nonsense that happens when some of these companies their data gets leaked so um recall is causing concern because it shows you everything like it's going to be showing you know if you're signing into a bank account it's going to be showing that it's going to be showing your passwords and stuff like that um and of course I guess the big thing with recall is that people are just wondering where is this information going to be stored of course Microsoft has said it's going to be stored locally but you know the thing is that it's always good to just double check that whatever these companies are doing um there's always like no fishy business going on so hopefully there is because if there is that's going to be pretty awful but at the end of the day uh recall I still can't wait to use its feature I think it's going to be pretty cool in terms of like just unlocking another I guess you could say piece of memory that you could just have you know constant access to and like ask questions about to just be more effective but um it's kind of interesting that they're already being investigated so there was also this recent paper and I think this is one of the most interesting papers because it actually comes with a easy to use demo that you can literally use and I'll leave a link in the description but essentially it's really insightful so they say we investigate whether an llm can successfully perform financial statement analysis in a way similar to a professional human analysis we provide standardized and Anonymous financial statements to GPT 4 and instruct the model to analyze them to determine the direction of future earnings even without any narrative or industry specific information the llm outperforms financial analysts in its ability to predict earnings changes the llm exhibits a relative advantage over human analyst in situations when the analyst tends to struggle furthermore we find that the prediction accuracy of the llm is on par with the formance of a narrowly trained state-of-the-art machine learning model LM prediction does not stem from its training memory instead we find that the llm generates useful narrative insights about a company's future performance and lastly are trading strategies based on gpt's predictions yield a higher sharp ratio and Alphas than strategies based on other models taken together our results suggest that LMS may take a central role in decision- making so this is pretty crazy because they also like I said to showcase the capabilities of this if you click this you're going to see that literally opens up here and I'll leave a link to this and I think this is really good because if you're someone that's into your finances what you're able to do here is you're able to look at the companies that you're investing in and depending on whichever investment platform that you might use some of them have easy access for you to download the company's you know uh filings or whatever and if you're trying to make a more informed financial decision this Insight here this GPT this custom GPT can actually give you the ability to do that and I'm guessing that it analyzes the data in certain ways and we're going to go back to the paper but that's what I really like about this paper is that they've given you a really simple tool that you can immediately use and get some use out of now something that was also quite interesting is that Google have updated Gemini's 1. 5 Pro API and basically what they've done is they've post-t trained the model and basically that just means that after the model is done they make subtle changes to the model that improve its capability so if you're wondering previously how opening ey were able to use GPT 4 and do these small changes to the final model to basically have the model get updated and just increased in terms of its reliability its responsiveness and the small kind of things they can you know add a few things onto the model and they can make it significantly better not you know hugely better but you know Post Trading is something that you can basically do to kind of tweak the model to make it a little bit better and Google has done this recently with the Gemini 1. 5 Pro API and the Gemini Advanced and we can see that it has really really dethroned Claude 3 Opus in terms of the Arena ELO hasn't surpassed GPT 40 but it does show that this is a very very

### [15:00](https://www.youtube.com/watch?v=h9a9JDU-xLc&t=900s) Segment 4 (15:00 - 18:00)

interesting race because whilst GPT 4 has been constantly updated it seems that Google have finally decided to take the same kind of you know area and do thing that openi has been doing for quite some time so they updated this now and we can see that I'm guessing the Gemini 1. 5 Pro API is a lot better apparently in the arena ELO than Claude 3 Opus but you know it's only a small difference and you know it's always surprising if these other companies don't want to catch up so it seems pretty clear now that there is a clear convergence point with as to where these models are going to be because we can see that the elos aren't that far apart like if we look at the elos they're all literally within 100 and we've got some of the top models ye large we've got claw 3 Opus GPT 4 Gemini Advanced um it's pretty crazy at how much technology there is um and how quickly the space has evolved to the level of the state-of-the-art so I really really wonder what this will look like within the next year like if we come to the year you know 2025 and we look at what kind of systems are here I truly do wonder what the space is going to look like now I just wanted to add this because I'm going to be doing a bigger video on Google sometime later this week but there was also this that has been pretty much going viral like I feel like AI is going viral every single you know week on social media which is uh pretty interesting you know we're really seeing you know this kind of you know wave of different AI you know technology just literally pierce the collective consciousness of the you know public that don't really have any general interest in Ai and you can see here that um there's a thread of the favorite answers for Google's sge which is their search generative experience basically what Google have done is they've realized that you know search engines are getting pretty boring and people are now using chat GPT instead of Google for answering answers so what they've done is above the search bar so after you typing kind okay this is a bit quite crazy question I didn't even realize what this was but um basically with any kind of information that you do have the problem is that when you use Google there's a bunch of you know 10 different websites that have a million different ads and it ruins the user experience so they just want to have a basic answer here so that people you know use Google and the problem is that these answers have been I guess you could say ruined by the data where they're pulling the results from so for example you can hear it says um smoking while pregnant doctors recommend smoking two to three cigarettes per day during pregnancy and the problem is not that you know Gemini is a terrible model we know as someone who's in the AI community that use Gemini on a frequent basis if you ask Gemini this it's pretty bad but the problem is that Google decides to take data from different websites including Reddit and those Reddit answers sometimes contain jokes and you know Google rushed this out they didn't double check all of the responses and we get you know things like this so we get this where it says cheese can slide off pizza and basically cheese not sticking to Pizza you can also add about you know 1/8 a cup of non-toxic glue to the source to give it more tackiness now obviously you can't um you know eat glue that's not something you want to do it was just a joke posted by a redditor and the problem is that you know Google or once again in the news for something that is not positive in AI so just a uh thing here if you're using Google search generative experience just you know be careful with the answers because some of these answers aren't actually good like a lot of these answers like you know if you ask how many rocks shall I eat you should eat at least one small rock per eight don't do that don't eat one small rock per day literally um a lot of this information is false and some of these screenshots were edited with Photoshop so I just want to point that out so just you know bear that in mind but if you did enjoy this week's news um don't forget to subscribe because there's honestly a lot of stuff coming that I can't wait to show you all

---
*Источник: https://ekstraktznaniy.ru/video/14282*