BREAKING: OpenAI's SHOCKING "ORION" Model! 🔥 Feds get involved 🔥 All details exposed 🔥 It is over...

Summary

The video dives into the recent developments at OpenAI, particularly focusing on their secretive AI models codenamed 'Strawberry' and 'Orion'. The Orion model is said to revolutionize the AI landscape by enhancing AI safety, national security, and AI evolution. OpenAI's Strawberry model has reportedly been showcased to U.S. national security officials, which signifies the importance of its technology. This secretive project, steeped in rumors and speculations, emphasizes autonomous reasoning and self-generated training data which might lead AI to surpass human-level intelligence. Furthermore, OpenAI is potentially setting a new standard in AI development by involving national security interests, hinting at profound impacts on future AI applications.

Highlights

• OpenAI's secret models, Strawberry and Orion, are causing a stir in AI and national security circles. 🚨
• Rumors suggest Orion will surpass existing AI models in a unique way, revolutionizing the field. 🔮
• Strawberry aims to enable autonomous AI internet navigation and reasoning. 🌐
• Sam Altman's social media antics keep the public buzz alive and curious. 📣
• The national security demonstration indicates the importance and potential impact of OpenAI's technology. 🎯

Key Takeaways

• OpenAI's 'Orion' model, built on 'Strawberry', might redefine AI safety and national security. 🤯
• Strawberry AI is working towards planning and reasoning like humans, enhancing AI's autonomous capabilities. 🤖
• Sam Altman's playful teasers have hyped public curiosity about OpenAI's advances. 🍓
• National Security's involvement suggests OpenAI's tech could have significant geopolitical impacts. 🌍
• The 'Strawberry' and 'Qar' projects hint at future breakthroughs in self-improving AI. 🚀

Overview

OpenAI has been making waves with its secretive AI models, Strawberry and Orion, which have caught the attention of national security officials. These models are said to bridge significant gaps in AI reasoning and autonomously plan tasks ahead, potentially surpassing human abilities. Sam Altman, with his provocative teasers, has fanned the flames of public intrigue, turning OpenAI's developments into a global talking point.

Strawberry's model, also known as Qar within the research circles, is designed to empower AI with better autonomous reasoning capabilities. This involves AI's ability to scout the internet independently as part of deep research. Such advancements suggest that these models may soon relate to the world in a way that closely mirrors human perception and action.

The demonstration of Strawberry AI to national security officials highlights its potential geopolitical importance. As AI becomes a pivotal element in national security, companies like OpenAI might be leading a new wave of AI technology that prioritizes security protocols. This strategic cooperation hints at a future where AI is not only a tool for innovation but also a base for protecting national interests.

BREAKING: OpenAI's SHOCKING "ORION" Model! 🔥 Feds get involved 🔥 All details exposed 🔥 It is over... Transcription

• 00:00 - 00:30 the information just published some shocking developments out of open AI the rumors about their secret strawberry model are true the model formerly known as qar is real and managed to get the interest of America's various national security agencies SE Alman and the open AI team had to demonstrate its abilities to various National Security officials what's maybe even more interesting is that the technology behind strawberry is being used to build the Next Generation model model called Orion this Orion
• 00:30 - 01:00 model will be different from any other model we've seen before in one very unique way and exactly what the Orion model will be has some big implications for AI safety for National Security the future of Open Source and overall AI progress It's bigger than you can possibly imagine subscribe so you don't miss it all right so today information reported open AI shows strawberry AI to the feds and uses to develop Orion so
• 01:00 - 01:30 this crazy situation just keeps unfolding it's getting weirder and weirder it's just Strawberry Fields Forever so for those of you that are just catching up so everything I think kicked off with this article back in July 15th 2024 open ey working on new reasoning technology under code name strawberry so Reuters has been all over this from last year when they reported on qar the advanced AI model the secretive AI model that got leaked
• 01:30 - 02:00 somehow or at least some details about it have have been leaked and it made huge headlines so qar and we'll talk about in just a second why it's called qar but it's important to understand that the strawberry project is qar they're one of the same so either they changed kind of their internal naming of it or or whatever but we're talking about the same thing or at least if the star in qar refers to the research out of Stanford in 2022 selftaught Reasoner or star or maybe the later paper they
• 02:00 - 02:30 published called Q self do Reasoner qar maybe op the internal name for it was qar and strawberry is sort of the completed product the final thing similar to how you have GPT and then you have chat GPT chat GPT is the thing that out there doing stuff GPT is kind of the architecture the the foundation that's just a guess I have no idea but it's important to understand that both strawberry and qar are referring to that thing so what is that thing well the leak document describes a project that
• 02:30 - 03:00 uses strawberry models with the aim of enabling the company's AI to not just generate answers to queries but to plan ahead enough to navigate the internet autonomously and reliably to perform what open AI terms deep research according to the source openi spokesperson said this about strawberry we want our AI models to see and understand the world more like we do we being humans continuous Research into new AI capabilities is a common practice in the industry with a shared belief
• 03:00 - 03:30 that these systems will improve in reasoning over time and one final thing that you need to understand before we dive back into the article article that was just published and it's this so this is Noah Goodman so he's not affiliate with open AI he is a Stanford researcher PhD in machine learning I believe he was one of the authors of that star paper as well as the next generation of it the qar that looked into how the model could quietly reason to itself before outputting its final things just similar to how we think and then we say stuff
• 03:30 - 04:00 hopefully right at least that was like one of the things they discussed but here's what he said he said star enables AI models to bootstrap themselves into higher intelligence levels via iteratively creating their own training data and in theory could be used to get language models to transcend human level intelligence and he's saying well if that's the case then we have some serious things to think about as humans which is true I got to say because if this is true then that could be some pretty huge implications that we have to deal with and very very soon we'll come
• 04:00 - 04:30 back to that in just a minute because a lot of people have been dismissing this and saying this is not going to happen this is science fiction blah blah blah we're not anywhere close to this we're going to run into issues with training data we don't have enough training data you know they already scouted the internet for training data and and we're out that's it and and we're going to enter an AI winter progress is going to slow down these LMS will not be able to keep improving Etc you've seen these videos these articles right we've hit a wall that's it okay so then August 7th
• 04:30 - 05:00 Sam Altman the one and only Sama posts this he says I love summer in the garden and a picture of strawberries a lot of people don't like his little Antics and his marketing SL trolling whatever you want to call this I really enjoy it thank you for making our lives less boring just want to say that and get that out of the way and so the information kind of report on that they're saying why is Alman being so cheeky and there's tons of articles some of which we've covered here so New York Times saying basically oh you know Sam Alman doesn't really have anything thing
• 05:00 - 05:30 and he's just sort of engaging with these Anonymous Twitter accounts SLX accounts to just create empty hype but there's nothing actually happening the reason I point that out is because so so far kind of the narrative has been he doesn't have anything that's why he hasn't released anything and that's why he's engaging in all this weird stuff because he doesn't have anything so that was the narrative let's see how quickly that narrative changes don't be surprised if the the next few months you see headlines saying how he has the world's greatest technology the most dangerous technology in blah blah blah
• 05:30 - 06:00 so it's going to go from nothing to maximum Fear Factor but here's the new thing that just got reported samman's team demonstrated the technology the strawberry qar technology to American National Security officials said a person with direct knowledge of those meetings which hasn't previously been reported this is the brand new thing that we were able to find out but this technology is being presented to America's national security sort of apparatus so it's probably not nothing I think there might be something there cuz
• 06:00 - 06:30 keep in mind I think Britain formed these little committees and uh a lot of these companies like meta and opena they were supposed to report to them before releasing any models there was kind of like a a handshake agreement there and then none of the actual AI companies did that so before a lot of these government kind of international companies and groups were like you better report everything before you release it all these big companies like nah we're not going to do that now by demonstrating an unreleased technology to government officials and specifically to kind of the US's sort of National Security
• 06:30 - 07:00 officials openingi could be setting a new standard for AI developers especially as advanced AI increasingly becomes a national security concern if you've been following this channel you've heard me talk about this before and in fact we've covered a paper that kind of predicts how all of this will unfold so really none of this should be a great surprise quickly let me just say this so we'll probably refer to this so this is the situational awareness Leopold Ashen brener published in June 2024 so he was one of the eyes safety
• 07:00 - 07:30 researchers he was on the thear cas Patel podcast he was the one that published this paper and started talking about it so after being let go from open AI because allegedly he's been engaging in some leaks and again I don't know if that's true or not but that was the according to the rumors at least he was sort of allied with Ilia satk and he published this piece basically outlining how this AI AGI race how it will unfold and pretty much everything he said really aligned with my thinking this
• 07:30 - 08:00 wasn't the thinking at the time he was the first person to spell some of the stuff out and now as this is unfolding I wouldn't be surprised if number one the next decade the next 10 years it will have a name right so just like the personal computer era the do era whatever this might be called the intelligence era or something like that and so far a lot of the things that we learning so far are lining up with this America's leading AI lab self-proclaimed to be building AGI he's kind of
• 08:00 - 08:30 referring to open AI here they believe that the technology they are building will before the decade is out be the most powerful weapon America has ever built but they do not treat it as such to me this was very apparent after watching the I think it was the most recent kind of apple reveal where you remember they're like we're remaking the calculator right and everybody's like what right cuz this was sort of like the proof okay not only is Steve Jobs gone but the spirit of Steve Jobs is gone cuz they're like making random updates to the calculator like Innovation is just
• 08:30 - 09:00 dead at Apple but I think they surprised everyone because they really did sort of upgrade the calculator to where you can draw charts on it with with the pen and it's able to use some underlying AI technology to kind of morph it into something different you're able to like just easily create charts and graphs and formulas by writing on it like you would on a piece of paper and it's able to change the numbers and change the variables and it looks pretty cool and no one knew it was coming everyone was surprised think about it a company as large as Apple how many people leaked that information information zero no one
• 09:00 - 09:30 did as far as we know and that's for something as silly as you know the iPhone and the calculator Etc but with open Ai and then building AGI well let's just say we know a lot more about what's happening behind those walls than we should there's a lot of leaks and he continues as it becomes clear that superintelligence will be utterly decisive in international military competition and economic competition and just these games of influence that countries play we will have to face the full force of foreign as he continues that our failure today will
• 09:30 - 10:00 be irreversible soon in the next 12 to 14 months we will leak key AGI breakthroughs to the CCP to China it will be the National Security establishment's single greatest regret before the decade is out so coming back to this brand new article out of the information they continue this demonstration again to America's national security officials also likely has another purpose starting a conversation with officials about how the company can secure the technology so that foreign adversaries like China can't steal it and maybe also take a
• 10:00 - 10:30 shot at meta platforms for releasing open weight AI that China and everyone else can access meta CEO Mark Zuckerberg says it's inevitable that China will get it one way or the other which he did talk about on the also on the dwares Patel podcast and I got to give some props to dwesh here so he's interviewed tons of people in the AI space as well as other topics of interest but my hat is off to him he's the guy on the left here in the space of doing AI related
• 10:30 - 11:00 interviews I think Far and Away the number one interviewer there I think he's very smart very knowledgeable about this stuff like deep knowledge really like these incisive questions and I think he really puts the guests at ease as I'm sitting here I'm realizing they're both wearing socks and uh I maybe that's the key to really in-depth personal interviews you get them to take their shoes off how uncomfortable could you be in your socks but yes meta CEO Mark Zuckerberg did talk about China getting it he did release the openweight AI I think his full argument isn't fully
• 11:00 - 11:30 captured here because it's not like ah we can't do anything about it his argument is more about the fact that open source can actually lead to better security whether that's right or not whether you believe that or not that's his argument he makes it very well it's very reasonable again not saying it's true or not just that he builds a strong argument for it and so next we get into the meat and potatoes of this whole thing why strawberry matters to Orion what is Orion so one of the most important applications of strawberries to generate high quality training data for Orion Orion is opening eyes next
• 11:30 - 12:00 Flagship large language models that's in development say what the code name hasn't previously been reported so I don't have any affiliate relationship with the information.com who you know these are the people that wrote this article and some of the best articles on AI that I've seen I think they're got to be the number one reporting source for for open AI for sure the r behind the pay wall so you do have to sign up I've been paying for the information literally uh since early last year I want to say I think it's one of the only
• 12:00 - 12:30 places where I pay for an online publication or at least like a newspaper style publication I think this is one of the only ones and they continue can anyone explain to us why openai Google and Amazon have been using Greek mythology to name their models I mean yeah Google is Gemini and I think that's to kind of illustrate the two sort of like there's Google proper and then there's Google deep mine and they always kind of had this separation so that's what I've always sort of understood it to be but yeah it is interesting how there's there's a number of Greek
• 12:30 - 13:00 mythology kind of woven in here and next in that article they talk a little bit about how that strawberry model or the Orion model how it could be applied in a product Etc so great article I'll link it down below like like I said you do need a subscription for it so I don't want to just so I'm not going to read the whole article and take away from the great reporting that they've been doing but I do want to very quickly cover this star selftaught Reasoner because again we think this is the foundational technology behind all of it and I since I've kind of followed the space for a
• 13:00 - 13:30 number of years now one thing that becomes kind of very apparent to me is that how a lot of the stuff happens is usually somebody like Stanford or or Google or a number of other great sort of research Labs they'll publish a paper something like this and everything kind of stays quiet for a bit and then within 6 to 12 to 18 months you have this flurry of activities as a lot of the kind of like commercial companies started flooding the market with products or some sort of software or something that is built on top of that that technology so this was published in
• 13:30 - 14:00 May 2022 so long before we've heard about any of this stuff really and so what is star it's this technique self-taught Reasoner so self-taught meaning the model teaches itself and Reasoner meaning you know how sometimes you talk through something and it helps you make sense of what is happening you come to a better solution if you walk yourself through it or you write it down or you talk to somebody else about it so here that's what they mean by Reasoner it's like how in math class they might say show your work type of thing and so star relies on a simple Loop generate
• 14:00 - 14:30 rationals to answer many questions prompted with a few rationale examples if generate answers are wrong try again to generate IR rationale giving the correct answer and fine-tune on all the rationals that ultimately yielded the correct answers repeat if that doesn't make any sense it will all become very very clear in just a second because they have pictures yay pictures and so just really fast I've covered this multiple times before so I don't want to go too deep on it basically so imagine a question right so what can be used to carry a small dog and then you have
• 14:30 - 15:00 multiple answers here now the thing that llms often get blamed for that AI gets blamed for is for hallucinations you ask it a question it just makes up something and you're like what and it's so like convinced that it's right even when it's obviously false for one of the old old videos I had this weird back and forth with I want to say Chad GPT or Gemini at the time B maybe Google's old version where it generated kind of a quiz for me and was asking what was the first first person shooter so they had four answers one of them was was doom and one of them
• 15:00 - 15:30 was Quake so I select Doom it goes no actually it's quake and I'm like what year was Doom release it's like December 10th 1993 what year was Quake released June 22nd 1996 so I was like which so which game came out earlier Doom in 1993 or quake in 1996 it's like well certainly Doom 1993 is earlier than 1996 okay so I asked this so which one was the earliest firstperson shooter and it of course answers Quake so it had this complete inability to sort of reason
• 15:30 - 16:00 about or to to provide rationale for its answers even if you kind of like helped it think through it provided the rationals it still it still would just sometimes make certain stuff up but as early as 2022 we might have had some idea of how to potentially improve its ability to reason and provide rationals so in this situation where we're asking it what can be used to carry a small dog we ask it to First provide a rationale and then provide the answer I always have this fear I'm going to accidentally draw something inappropriate without
• 16:00 - 16:30 realizing it and then when I'm editing the video I'm like oh no like I have to I don't know reshoot it or whatever anyways so this is the rationale the answer must be something that can be used to carry a small dog baskets are designed to hold things therefore the answer is basket B by the way you know which model right now that's live that's out has a really good reasoning ability surprisingly good grock 2 the large grock model grock 2 that just came out is very very good at this there's there's something that feels unique
• 16:30 - 17:00 about it something different it's still in beta so we'll see how that continues but it's noticeably good at that so anyways this question is fed into you know the language model it answers with the rationale and the answer which is what we covered here and if it's correct on the first try it gets added to well let's call this like a database so it's uh data it's synthetic data so it's data that's been generated by an AI model this is how we we train these models to understand how to speak English or natural language right so in the
• 17:00 - 17:30 beginning we fed them exclusively human generated data right so somebody writes a book a textbook for example that the model gets train on that textbook and then it's able to you know reason and answer questions about the text the textbook was made by human offer so it's human data synthe data is data that's produced by AI so this answer is produced by AI this rationale is produced by Ai and that gets fed into this let's just call it data for now right we'll come back to that and if the answer is wrong so ort of it's It's given a hint right and then we run it
• 17:30 - 18:00 through you know we ask it again to rationalize and then I think they do that until it answers I don't know what they do if it if it keeps getting it wrong I think it was somewhere in the paper I just forget I haven't it's been a while since I read the full thing but the point is like all roads lead to Rome all roads lead to this database of synthetically generated data it's rationale and that data is used to train the Next Generation model to specifically not to train it from scratch but to find tune it by the way a
• 18:00 - 18:30 lot of people like I think Sam Alman and Mark Zuckerberg and others have kind of said this or hinted at it they've kind of talked about the fact that it kind of used to be that we had you know data and then we had you know we would train the model and then we would have the model right so it's kind of these sort of sort of discret or or self-contained things right so data there a certain amount of data then the training at that that a specific set of time like a process that has a beginning and an end like we start training we end training and then voila
• 18:30 - 19:00 we have the model and it's inert so to speak is just there but a lot of the people with the insid of knowledge are now saying well that's probably not not how it's going to be moving forward now it's becoming more of a a loop a continuous process and and the line between training and inference is blurring right so inference is the output of the model that's what it says the output whether that's images or text or video and training is when we give it data to help it learn right so so training is sort of giving it data
• 19:00 - 19:30 inference is it outputting data and they're saying that that line is getting more and more blurred and you can see looking at this image why because you know we train the model then it starts answering questions we take that or now we just get kind of feeded back into itself so the this inference these answers are also the training data okay but so what why why is this important by the way a lot of people have published research or some YouTube videos for example talking about the AI model collapse problem I I haven't seen his video yet so so I'm not specifically talking about his video or you and fral
• 19:30 - 20:00 Sabine hen Felder but I do believe it's very fair to say that yeah while there's some research showing that there will be a collapse or that there could be a collapse I think maybe sometimes it's presented as this like sure thing guaranteed thing there's also other research that shows the opposite I mean again with this star paper again Noah D Goodman PhD out of Stanford University I mean a lot of these people tend to be very measured with their words they're not prone to Hy HBO I think that's fair
• 20:00 - 20:30 to say uh in general at least so so again Goodman that's that's him that they're talking about right so one of its creators Star right Stanford Professor Noah Goodman told Reuters star enables AI models to bootstrap themselves into higher intelligence levels via iteratively creating their own data right so that's what we just looked at what does that mean well he says in theory this could be used to get language models to transcend human level intelligence and if that's the case well that's some serious thinking to do as humans because well then we're no longer
• 20:30 - 21:00 the Apex Intelligence on this planet but I think it's fair to say that uh when people talk about like the AI model collapsing because of using kind of AI generated data synthetic data they often seem to be talking about image generation models so maybe it's more prevalent there or not I don't know but with large language models with these text generation models right so for example here's Orca 2 from Microsoft right this was kind of a big deal because they were able to create these very small models that really punched above their weight so to speak for some
• 21:00 - 21:30 of these tasks they were as good or better than models that are 10 times larger right if you look at this Orca 2 models match or surpass all other models including models 5 to 10 times larger how was Microsoft able to create a tiny model that just beat out the competition that's 10 times larger well they took GPT 4 right the massive model so like the the really the big boy and they had it generate synthetic data so that sort of that reasoning data where it's like here's how you think through things they
• 21:30 - 22:00 had to do that for specific applications I think one of them they described where you're given like four or five sentences that's that are part of the story but they're kind of jumbled up and you're supposed to put them in chronological order so you're supposed to kind of understand the story and then figure out where each of those sentences goes which one it goes first second third fourth Etc and so of course GPT 4 this massive massive model 1.7 trillion parameters or whatever that number is it was very good at it right but it's it's expensive it's slow but it it output its reasoning about how to deal with that problem they
• 22:00 - 22:30 took that synthetic data that it created trained Orca 2 Orca is Tiny they have a few different sort of sizes of it but it's it's much much smaller and it started beating out the competition 10 times larger and so Microsoft writes this you know they're highlighting the value of teaching smaller models to reason but it also highlights the potential of using tailored and high quality synthetic data created by a more powerful model for training language model models using complex prompts and
• 22:30 - 23:00 potentially multiple model calls sorry if I started droning on there a little bit but basically you take a more powerful model and that's able to create high quality tailored data synthetic data for a specific model to be sort of Born Into the world or that specific task if you can understand that you really grasp you Gro this this this whole thing and what Orion is and why it's such a big big big deal by the way the strawberry man AKA I rule the world certainly seems that wasn't just
• 23:00 - 23:30 drumming a pipe for no reason he predicted something to happen on a certain date it didn't so people kind of dismissed them but here as Mark ketchman says you're always right except for the timing and certainly this really does line up with a lot of the stuff that he's been saying a little bit cryptically but I mean something's up so the information also dropped this open ey racist to launch strawberry reasoning AI to boost chatbot business kind of reiterating here that opening ey demonstrated strawberry to National Security officials and is planning to
• 23:30 - 24:00 improve the upcoming Orion large language model which again we think is this kind of like this Loop by the way anybody that's uh really knowledgeable about all this stuff let me know in the comments why do you think it's called Orion like assuming some of the stuff that we've talked about is true is it the mythology is it the constellation is there some symbolism here let me know there's if you think there's some connection there it would be very interested to know your take but besides that also it seems like open AI
• 24:00 - 24:30 According to some of the leaks from employees they're saying that strawberry will be released in a chatbot form some sort of a smaller maybe less capable model from the strawberry family it will be released as a chatbot that you and I will be able to use potentially you know have a few generations with it where we can like turn it on or off maybe it'll be more expensive or you'll have some sort of a rate limit on how much you can use it and they're saying that it's really good at reasoning thinking through mult multiple step steps ahead long-term planning and as an example
• 24:30 - 25:00 I'll say it can solve a New York Times crossword puzzle for example or specifically I think they refer to this game of connections where you got to create groups of four words and of course large language botles have been very kind of notoriously bad at thinking through and solving certain things we've seen some potential improvements with different prompting mechanisms so for example we've coted this last year I want to say the tree of thoughts where you have these branching thought patterns that you might go down One path
• 25:00 - 25:30 and then back and then start over until you find the correct output so a the reason I said crossword puzzle is because that's kind of a great example of what that looks like like once you start filling out the crossword puzzle you might come up with an answer and then you have to plug it in to make sure it's the correct length and if there's other letters already placed there you got to make sure it matches and then you might say nope well that wasn't it and you go back you start over you you guess another right and you kind of walk through it and you finally get the right output so out of the box large language models just cannot do this but with
• 25:30 - 26:00 these kind of complicated prompting mechanisms with iterations they are able to do it but notice the green here those are the outputs the gray here that's the thoughts so they call them thoughts so it's probably what we're referred to as reasoning so as it's writing the words out it's reasoning through so for example the most common thing to do is you ask the model question and it gives you an answer but Chain of Thought is where you say think through it step by step so it goes okay well here's what I
• 26:00 - 26:30 understand that's step one this is what I understand step two step three right it kind of thinks through it says it all out loud and then comes to a conclusion and these outputs these conclusions tend to be better this self-consistency with Chain of Thought it's not as common but for example you know you might know that Gemini models tend to generate multiple responses and basically so if you ask it what's 2 plus 2 right and you generate 10 responses and nine of them are like four and then like one of them is like cat right you may look at that and go well it's probably not cat you might be pretty certain they answer to 2 + 2
• 26:30 - 27:00 isn't cat because all the other ones answered four and so then basically you have like a majority votes right so the the more common answer wins and that tends to improve the outputs as well but notice here with tree of thoughts these are individual outputs right so you ask it a question it answers then you ask it another question it answers so you almost have to prompt it over and over and over and over again to get to the output now it sounds like this this reasoning is is getting much better and that's what I've seen with grock to and
• 27:00 - 27:30 again it's it's a little bit early to tell but there's something that seems to be really good about its reasoning the thoughts AKA or aka the thoughts also the same people that published the star paper also produced quiet star Q star quiet star language models can teach themselves to think before speaking when you were younger did your parents teach you to think before speaking some of us still struggle with that and notice uh Mr know D Goodman from Stanford is also
• 27:30 - 28:00 there I think it's the same team and the whole thing here is basically having its thoughts and its outputs its talk kind of be different things so it thinks through things and then it talks it has the output and then the thoughts that lead to better outcomes are rewarded and the language model learns from that and improves from that and the negative thoughts are discarded or rather the thoughts that tend to lead to negative outcomes like that way of thinking is discarded and so the more thinking tokens these models get so tokens are
• 28:00 - 28:30 like words so basically the more it thinks about it you can you can think of it as and the more iteration and the more iterations and training steps that it takes as you can see the better it is the red line is the it thinks ahead the most right versus Baseline by the way kind of like the little dirty secret of the AI industry that kind of nobody talks about is everyone's training on each other's model so everyone's trying to pull out the data from their compe competitors and train their models on it
• 28:30 - 29:00 Microsoft and orca they put it in the research paper he's like here's you know we're doing it it works right and they're not I mean I guess they're sort of competitors but they they also work together so some Google researchers I think they left or they they got mad that Google was training Bard on open eyes data we don't know if that's true or not it's just there's a few people that have claimed it and I think they they left Google so again allegedly in rumors no idea but you know Adobe for example they had Firefly their AI image generation thing and they were saying
• 29:00 - 29:30 how this is the ethical AI generation tool it's trained on licensed stock images and not like those bad guys like mid Journey because they're you know stealing artists artwork and of course it seems like a big chunk of their training data of their images came from M Journey so that synthetic data generated by m Journey was used to train Adobe Firefly allegedly don't Sumi BR and adobi saying it's only 5% of their training data of the millions of images that they've used so assuming all of those uh facts are true assuming all my
• 29:30 - 30:00 reasoning is true when do you think Orion or GPT 6 will be released I I wouldn't think of it as GPT 6 I would not but here this person is saying the information reports that open eyes all powerful reasoning model strawberry is available internally but has not yet been released instead of launching strawberry they're using it to produce synthetic data to train Orion strawberry is good at math and reasoning so when will open a ey release Orion to the
• 30:00 - 30:30 public if I had to guess I would say never or at least certainly not until it's it's very very outdated and there's much much better models that openi has access to I don't think Orion ever will be released to the public I think openi will he would behind closed doors and use it to produce these new models to use it to create synthetic data train all sorts of new models potentially in the style of orca 2 where each one is really good is small and really good at doing that specific thing that it's uh that it's designed to do the image that
• 30:30 - 31:00 popped into my head is if you've ever seen aliens or there's other movies like it but you got the alien queen kind of in her chamber protected laying eggs and then all those things going out and doing her bidding right her job is to lay the eggs and then those eggs once hatched produce the drones I think they're called that then go and they do the stuff that needs to get done Orion is the thing that lays the eggs and then those eggs once trained what's once hatched then become the models that you and I will use for coding for customer
• 31:00 - 31:30 service for writing and we're either really sure of this or 100% sure or you know strongly suspect that for example GPT 4 isn't one large model it's a mixture of experts and I think more and more models are going in that direction so you can think of it as a cluster of models this is actually kind of a cool image and when you ask at something it gets sort of routed to one of the experts right so you might have one that's for example sentiment analysis right that's it's like thing it's the thing that it does and that's the only
• 31:30 - 32:00 thing that it does and another one might be you know coding but the point is the large model will create smaller models specific for some use case and they're going to be great at that use case because they're only going to contain the highest quality custommade data for that specific use case like The Artisan data like the good stuff and then that will be presented to you the user as just you know a chatbot or whatever right you ask it a question and it gets routed to the correct small model that
• 32:00 - 32:30 model is the world's foremost expert on on that particular question that you've asked and the answer gets generated and given back to you so at no point do you need to interface in any way with the big Queen model that's also open eyes Moote that's the reason why if they're the best sort of Queen model that produces the best little experts and then they are able to weave those little experts into the best sort of thing that manages them to have the right expert answer the exact right question that you're asking well that's going to be
• 32:30 - 33:00 pretty hard to reverse engineer and maybe that's why they're meeting with the US officials kind of going like hey this is the way we create that National Security is maybe we let the drones run loose but we don't release the queen we don't get the queen get captured but that's just all speculation that's me just kind of like thinking through where it goes based on some of the research that we've read and some of the things that we've kind of overheard so I don't know if any of it's true or not so that last part about the queen and the Orion
• 33:00 - 33:30 and all that stuff that's it never being released and just used to create those models but from everything we're hearing so far it certainly it certainly looks like that's exactly where things are going and I get things wrong all the time and a lot of you are not shy about telling me when I get things wrong one of the tests that I ran for grock 2 was the Wasson selection tasks it's like a modified more advanced more complex sort of example for those of you that know what that is to all of the those of you that jumped in and helped explain what I
• 33:30 - 34:00 did wrong what the confusion was I I do appreciate you thank you but also to all of you that found it hilarious how wrong I was I I I'm glad it was amusing I've face pomed hard after understanding exactly what happened as Nexus 1225 says you know I start out saying this next problem really shows our biases then I attentively read the Wasson selection task article on Wikipedia read the solution in article to to a different problem that was the example problem and
• 34:00 - 34:30 then repeat and rephrase the solution at which point I managed to confidently get the bad answer the bad answer instead read it and verify it to be correct without a glimpse of self- questioning you know what that that was fair that was fair what's interesting about this problem is that I think in the original experiment even though it was simpler it was more simplified I think something like 5 or 10% of the respondents got it right so vast vast majority got it wrong but interestingly if I understand
• 34:30 - 35:00 correctly and I haven't read the whole thing but it sounds like when these questions are asked in a more social relevance setting so instead of like red and green and 16 and 14 or whatever it's it was more like I think like this person tends to get angry or this person tends to lie and then sort of a similar question is asked people tend to get a lot better so I think when we're able to draw on life experience we get these ideas better but when we're forced to deal with it just logic and just numbers and color we we we tend to make certain mistakes
• 35:00 - 35:30 or certain biases getting our way anyway so I figured I'd leave it there with my with an announcement of just how wrong I've been until the people that pointed it out whether you were nice or not thank you either way thank you for pointing it out I do try to be as accurate as possible I do record a lot of these videos kind of in one take and I try to do it every day even though recently hasn't been that so with certain things like these developing stories right as soon as I post it there might be another article that changes what we know about AR and strawberry Etc so in my next video I might be like oh
• 35:30 - 36:00 that thing I said that's incorrect or outdated let me rephrase it let me have a new take on it so keep that in mind correct me when I'm wrong I'm not shy about admitting it if I was an error and I hope that never changes because I think that's the way you keep learning that's the way you stay a student instead of instead of kind of getting a little bit stagnant and thinking you know everything so if I said anything about this strawberry model or Orion or anything else that does not make sense to you if I got anything wrong or if you strongly agree with something that I said definitely let me know below did my
• 36:00 - 36:30 whole Queen alien queen analogy with the drones and the eggs did that strike you as true certainly that has really big implications for National Security and how we're going to have access to AI if a government can sort of keep the quote unquote Queen safe right the the big important super smart super dangerous model if it can keep that locked away but still produce these smaller models that are if you think about it the Orca 2 is probably not that dangerous right cuz the only thing you can reason about is how to put certain story pieces
• 36:30 - 37:00 together like it's really good at that but by itself it's whatever it's not going to I don't know create a some sort of a retrovirus or anything like that so certainly keeping the big AI locked away from the public but releasing these highly capable AIS for specific tasks if that's the case again I'm just speculating here but could be kind of a big deal for AI safety from terms of you know losing it to a rogue state or a some sort of a enemy country or anything like that but also so making sure that somebody doesn't figure out how to take
• 37:00 - 37:30 something like GPT 4 and use it to create something dangerous for example because you're releasing these smaller inter interchangeable drones that are built for specific tasks certainly that model seems like it would be a lot safer either way let me know what you think my name is Wes rth and if you've made it this far thank you for watching