# OpenAI'S Q-STAR Has More SHOCKING LEAKED Details! (Open AI Q*)

## Метаданные

- **Канал:** TheAIGRID
- **YouTube:** https://www.youtube.com/watch?v=K0LTpemBu74
- **Дата:** 21.03.2024
- **Длительность:** 14:01
- **Просмотры:** 43,075

## Описание

✉️ Join My Weekly Newsletter - https://mailchi.mp/6cff54ad7e2e/theaigrid
🐤 Follow Me on Twitter https://twitter.com/TheAiGrid
🌐 Checkout My website - https://theaigrid.com/

Links From Todays Video:
https://twitter.com/tsarnick/status/1768087211881480480/video/1
https://twitter.com/tydsh/status/1770614875708166557
https://www.theinformation.com/articles/why-openais-breakthrough-isnt-a-harbinger-for-humanity-ending-agi?rc=0g0zvw
https://pastebin.com/RkBUQPLb

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=K0LTpemBu74) Segment 1 (00:00 - 05:00)

so ladies and gentlemen there has actually been a new qar leak although quite fishy but there's a lot of information floating around the internet that I would just rather put out here into a video and let me know what you all think about it because it's rather fascinating so let's take a look at the new Q star leak and all of the information surrounding it okay so this is a paste bin and it was tweeted by someone but I honestly cannot find the original tweet anymore I'm not sure if the person deleted their account but essentially what we have here is a tweet about opening qar system now I do want to say as is with any leak you always want to be a little bit skeptical because it isn't confirmed they didn't really say where this came from but like I said it's just been floating around on the internet but there is some actual real information that I will get to later on in the video with as to why people are giving this some kind of credibility due to other sources talking about this quite as much so let's take a look at this okay so it says qart is a dialogue system conceptualized by open AI designed to enhance the traditional dialogue generation approach through the implementation of an energy based model okay and we're going to get into that in a moment it says distinct from the prevalent Auto regressive token prediction methods qar aims to mimic a form of internal deliberation akin to human thought process during complex problem solving such as chess playing where a deeper analysis of potential moves leads to better decision- making compared to Rapid less considered responses this model shifts Focus towards the inference of latent variables reminiscent of construction probabilistic models and graphical models fundamentally altering how dialogue systems operate so I'm guessing that here and I will get onto this later on in the video but um you can see that they're basically changing how uh these llms think and I guess you could say that they're moving from token prediction methods like it says here to something that is similar to human thought that's basically what they are saying using these energy based models and I will get into the technicals of that in a moment but it says energy based model for dialog generation and it says at the core of qar is the energy based model which operates by assessing the combat ibility of an answer to a given prompt through a scaler output this output signifies the energy of the response where a lower value indicates higher compatibility which is essentially just a better answer and a higher value indicates a poorer answer this mechanism allows qar to evaluate potential responses holistically moving beyond the sequential predictions of tokens to understanding the relevance and appropriateness so essentially if you think of it like this these energy based models they assess the compatibility of an answer where the lower value indicates better answer and a higher value indicates a poor answer and then essentially they talk about uh optimization in abstract representation space the innovation in qar lies in its optimization processes conducted not within the space of possible text strings but in an abstract representation of space here thoughts or ideas are represented in a form that allows for the computational minimization of the ebm Scala output AK to finding the least path of resistance in a landscape and this process involves a gradient descent a method for finding the minimum of a function applied to iteratively refine these abstract representations towards those that yield the lowest energy in relation to the prompt it continues to go on here it says from abstract thought to textual response once an optimal abstract representation one that minimizes the ebms output is identifies qar employs an autoaggressive decoder to transform this abstract thought into a contextual coherent response this step Bridges the gap between non-linguistic conceptual understanding of the dialogue system to linguistic output required for human interaction now here's where they talk about the training of the system they says that the ebm within qstar is trained using pairs of prompts and responses adjusting the system parameters to minimize the energy for compatible pairs while ensuring that incompatible pairs result in higher energy levels and this training process can incorporate contrastive methods where the system learns to differentiate between compatible and incompatible Pairs and non-constructive methods which involve regularization techniques to control the distribution of low energy responses across all possible answers and they State here you know whatever this is okay um implications for dialog system Q stles approach leveraging ebms for dialog generation represents a significant departure from traditional language modeling techniques by optimizing over an abstract representation space and utilizing gradient based inference and qar introduces a more efficient reasoned and potentially more powerful method for generating dialog responses and this system not only promises improvements in the quality of generated text but also offers a blueprint for future advancements in ai's ability to engage in human-like reasoning and conversational interaction so essentially why a lot of people were talking about this is because essentially there was a paper put out by open AI I think it was around 2019 it was called implicit generation and modeling with energy based models and essentially like the paper said you know these energy based models are a system that has a really unique way about thinking of you know and solving problems and essentially when you ask them questions well when you try to figure out questions with the system they don't just give you the first answer that comes to mind instead of they you know they consider a lot of possible answers but they don't think about these answers in terms of final

### [5:00](https://www.youtube.com/watch?v=K0LTpemBu74&t=300s) Segment 2 (05:00 - 10:00)

Solutions right away instead they start out with like rough ideas or concepts of what the final answer could kind of be and then they have a special way of judging how good each one of these ideas are so you can kind of think of it like each idea has a certain amount of energy and the best ideas have the low energy while the not so good ideas have high energy and then of course you take the idea with the lowest energy you refine it and make it better until you can't reduce its energy anymore and then once they have this best idea then it's translated into words and that becomes the final answer to your question and you know the process of starting with many rough ideas judging them based on their energy and then of course refining the best one which make is essentially what makes these energy based models powerful and you know essentially a lot of people are saying that this might be the future of planning because it's considered very flexible and can handle complex situations where there are many solutions and you know traditional planning methods often struggle when there are too many possibilities to consider and they actually might take too long to find a solution or get stuck on a not so good answer and you know Sam Alman did actually talk about that in an interview he he did with Bill Gates say that you know and maybe this is actually referencing this kind of model he did state that you know um in fact if I can find the clip I will put it in here but he did say that there are 10,000 answers that uh you know track gbt can give you and when it gives you 10,000 answers one of those is going to be good and usually you'd want to get that one every single time and that is something that they're working on if I can find that clip right now it's not actually in the presentation but eventually video clearly people really want that we launched images and audio and had a much stronger response than we expected we'll be able to push that much further but maybe the most important areas of progress will be around reasoning ability right now gp4 can reason in only extremely limited ways and also reliability you know if you ask gp4 most questions 10,000 times one of those 10,000 is probably pretty good but it doesn't always know which one and you'd like to get the best response of 10,000 each time so that'll be that that increase in reliability will be important customizability and personalization will also be very important if I can find it I will add it but essentially what he's talking about is something that's eerily similar to this now that I think about the clip um and I'm guessing that you know whichever kind of system you do use um which other research you know energy based system what whatever system it is to get your final output um essentially they're just trying to optimize whatever it is for reasoning so um these energy based models of obviously do offer a promising approach to that but the paper that they did put out it wasn't actually for text based uh you know outputs this one was for graphics I think it was for video and images so um that's why people are actually you know talking about this cuz it was you know something that people were deciding to tweet about now the reason why people were also talking about this as well is because you know Yan Lun was on a podcast and he actually spoke about energy based models um and a lot of people were starting to say and if you don't know who Yan is uh well respected scientist and Mets fair so that's essentially the AI division um and he actually spoke about this on a podcast with leex Friedman um and I'm going to basically talk you guys through this because essentially why this leak is so skeptical is because this was on like the 9th of March or something and then that leak was on like the 11th either way it was 2 days after I'm not sure about the dates but it was 2 days after so some people are stating that it's just a summarization of what Yan leun said so let's just take a listen to that uh really what so basically the principle is like this you know the prompt is like a observed uh variables mhm and what you're what the model does is that it's basically a measure of it can measure to what extent an answer is a good answer for a prompt MH okay so think of it as some gantic neural net but it's got only one output and that output is a scalar number which is let's say zero if the answer is a good answer for the question and a large number if the answer is not a good answer for the question imagine you had this model if you had such a model you could use it to produce good answers the way you would do is you know produce the prompt and then search through the space of possible answers for one that minimizes that number um that's called an energy based model but that energy based model would need the model constructed by the llm well so uh really what you need to do would be to not uh search over possible strings of text that minimize that uh energy but what you would do it do this in abstract representation space so in sort of the space of abstract thoughts you would elaborate a thought right using this process of minimizing the output of your model okay which is just a scalar um it's an optimization process right so now the way the system produces its answer is through optimization um by you know minimizing an objective function basically right uh and this is we're talking about inference we're not talking about training right the system has been trained already so yeah that was uh Yan Lan's uh statement on energy based models you can watch the full podcast with Lex Freedman but like I said the reason why many people were this credit Ting this leak is because uh

### [10:00](https://www.youtube.com/watch?v=K0LTpemBu74&t=600s) Segment 3 (10:00 - 14:00)

someone also tweeted out you know a des someone actually just basically tweeted out a description of qstar and essentially what they said is that you know it's not actually description of qar at all rather it's an autogenerated explanation by Claude of yan Len's ebm project as you see it looks very similar indeed and I'd be very skeptical about these claims of opening eyes leaks as it seems to be just summarizing what Yan Lian said so essentially he's just saying that look uh this kind of text document it was just used by Claude opus to summarize the document so it it's stating that it is probably fake however I do want to sayate some other things cuz I I think it is pretty interesting number one is the fact that on the same podcast samman said this uh you know a couple of days ago I think it was 48 hours ago one can drink open AI is not a good company at keeping secrets it would be nice you know we're like been plagued by a lot of leaks and it would be nice if we were able to have something like that can you speak to what qar is we are not ready to talk about that see but an answer like that means there's something to talk about it's very mysterious so yeah that is uh samman's statement on qar which goes to show that even if this leak is fake which it probably is uh it does mean that you know of course there's something there which I'd say open AI clearly have some kind of breakthrough I don't know what kind of breakthrough it is but clearly they are working on something that is there now of course the original letter was pretty crazy it had numerous insane crazy claims which are you know I guess you could say on bar with you know threatening to humanity and the original claim as well you know another one of the claims as well and this is from the information article um you know this this core principle of you know if qar actually does utilize ebm principles or similar methodologies to evaluate and select the optimal solution for unseen math problems this approach could be theoretically possible and the model would effectively search through the space of possible solutions minimizing the energy or compatibility of its answer leveraging patterns and principles it's seen during question so I mean it's something that starts to get into the area which is uh definitely pretty speculative because we don't know if this is something that openi is actively working on we don't know if they're using this um and that last paper they hadn't updated it for quite some time so it will be string I hope there is more information about this but like I would say you know even if this is fake uh someone at Fair someone at you know Mets a division also did tweet that you know look like the qar rumors align with a series of works on our latent space planning search areas um and then he links to seven different research papers that referenced that and then he says uh you know also check my previous comments regarding Sora which is I's recent video model and he states that doing planning SAS search in a learnable latent space rather than original space has a unique Advantage which reduces compound error and planning Horizon and there is strong evidence that you know this representation actually really helps in these scenarios and he says that is uh the area that meta are looking into and of course you know this is something that um that we've seen with of course Yan Lun and now he's obviously confirming it so I'm guessing that this is something that they're working on and of course he also talks about this in Sora and he said you know one interesting component of Sora is that the video is not created by next frame prediction but by constructing and redefining latent token sequence first and then decoded them back and doing planning search in latent space rather than original space has a unique advantage and of course this is why they say um this is beyond traditional RL formulation which assumes fix State SL action representation that may hinder optimization efficiency so I'm guessing that they're trying to use a more effective methods to you know provide problems to the solution because uh video generation is one of those things that's really difficult um and opening eye as always have come up with a new method but it seems that meta's hot on the heels in terms of their energy based models as well so either way we're going to be getting more information about these ebms because it seems like meta is working on that with Yan Lun and his team confirming that um and even if opening I don't publish a research sometime soon we know that meta are going to publish there so either way qar whatever it is it's definitely kind of fascinating definitely kind of interesting um and all we can do right now is uh of course not speculate too much but we can just wait until we have more information

---
*Источник: https://ekstraktznaniy.ru/video/14448*