The Verge DRS some big news open AI plans to release its next big AI model by December the Orion model is the Big Kahuna keep in mind that the next few releases of these models specifically by open AI will probably to a great degree determine how hot this Market stays how much hype remains because if the progress cools down especially by you know one of the market leaders opening eye some people might think you know things are slowing down however if we keep seeing Improvement in coding in autonomous research if it gets the gold medal
at the International Mathematic Olympiad for example which by way Google's model got very close they were one point away from gold but that was one of their Alpha models Alpha geometry plus Alpha proof I believe if a large language model can land gold I think that would be kind of a earth shattering breakthrough I would say I think it would be fair to say so The Verge continues the startup's next Flagship model code named Orion is slated to arrive around the 2-year anniversary of Chad GB this day really stands out in my memory so December
3rd 2022 now of course Chad BT was released November 30th 2022 a few days prior but I remember Elon Musk saying Chad GPT is scary good we are not far from dangerously strong Ai and I remember thinking to myself what the heck is chat GPT at the time if you recall they didn't know it was going to make quite this big splash Chad GPT was kind of like this tool this prototype the GPT sort of architecture they were around in fact before that I was using a tool that ran on GPT GPT 3 probably or
something like that I remember being pretty impressed by how well it could write I think it was called Jeeves or Jasper or something like that and I didn't know at the time was running on opening ice technology it was very good but I think for a lot of people this was kind of like the firing of the starters pistol letting them know that the race is on this was what woke a lot of people up to what was coming potentially so if this article is true and the Orion model the big one is arriving around
the 2-year anniversary of Chad GPT that means it's days away now a lot of the breaking news in the a field you probably noticed we pay attention to a publication called the information.com they've been pretty uniring correct in a lot of their leaks and breaking news about open AI about AI in general I actually would not be shocked if there's somebody out there in the Bay Area that's getting two paychecks one from open Ai and one from the information.com just a hunch we've also seen some great Scoops by Publications like Bloomberg and Business Insider on
for example the qar leak back in the days now this is coming from The Verge but take this with a grain of salt I'll show you why in just a second let's read what they're announcing but again there maybe skull dugery a foot if you know what I mean open a plans to launch Orion its next Frontier Model by December the Verge has learned unlike the release of opening eyes last two models GPT 40 and 01 Orion won't initially be released widely through Chad gbt instead openi is planning to Grant access first to companies it
works closely with in order for them to build their own products and features according to a source familiar with the plan now this is interesting because when we first started learning about Orion and again keep in mind some of these are leaks speculations Etc it seemed like there was one model that was meant to be kind of the model that trains other models it was the big smart model that would produce the data the synthetic data the AI generated data that then would go into these smaller models kind of customized for their own use cases
in my first video about the oran model and again there were multiple kind of code names that they've used so we weren't sure which one's which but basically the idea was one of the models was the one that makes the other models and I kind of Illustrated this by talking about if you've seen aliens there's this idea of the alien queen that just sits in its chamber and Hatches the alien kind of drones that then go out and do stuff right so that model wouldn't be the one that's powering you know Chad GPT it's not
the one that's answering your questions it's not the one that's coding it's not the one that's that's doing anything that a regular user would do instead it rapidly creates the data to train the smaller models we've went over a number of studies that show how that's done but kind of based on that idea so this really makes sense to me so they're saying unlike the release of you know all the other models that are power basically powering Chad GPT and you can select Which models you want to use in that scroll down the little drop
down menu in Chad GPT here they're not doing that this model it will instead be granted to companies for them to build their own products and features which again that could mean their own models yet another source is saying that so Microsoft right they're preparing to host Orion on their Azure service so the cloud service for a lot of these AI models as early as November so the Orion model is seen as sort of the Next Generation model after GPT 4 right so it's probably what was referred to initially as GPT 5 but in the
eye they have their own kind of naming convention and they always kind of like flip it around on us so I mean it's probably not going to be called that externally when it rolls out they'll probably have some other name for it this story also came out recently Microsoft prepares for opening eyes next model as their relationship strains if the stars align we should see open ai's next Orion AI model by the end of the year if the stars align Bravo very well said you you get it stars Al line Orion it's fishlyn man called
the story fake news Sam Alman has been a little bit extra chatty on Twitter SLX recently saying it's not that the future is going to happen so fast it's that the past happen so slow my thoughts on the matter are in one image this I had that image lying around and uh lo and behold I got to use it but Sam Alman did address the claim so this is Kylie so this was posted by one of the authors Kylie Robinson Kylie has a pet peeve that she wants you to know about it's Robinson not Robinson
did you read it as Robinson I think every single person that looks at this reads Robinson anyway she posted the article and Community notes immediately strikes and shortly after Sam the great Sama himself says fake news out of control the comments continue bro you just ruined all our Knights Sam Alman replies don't worry plenty of great stuff coming your way just offends me how media is willing to print random fantasy Jimmy apples replies to Sam back to the patient cave until clarifications I suppose Sam Alman replies you and Pikachu Jimmy replies but it well actually
that's where the thread ended by the way I'm curious and this is a genuine question please answer in the comments if you like I'm curious watching AI development unfold watching interactions like this you know I mean what you have here is you know you got a journalist somebody that's uh writing for the Fortune Magazine business inside of the Verge kind of like the old school Legacy Media with with massive reach you know shaping our politics shaping thoughts and opinions and just like wielding this this massive influence for for so long up until you know fairly
recently where you might say that that the balance has shifted a little bit so you know Kylie Robinson posts an article right that that that that post immediately gets you know Community noted for those of you that don't use Twitter and are not familiar with Community notes you can think of it as basically if you've ever been to some social Gathering and you say something wrong you get some minor detail wrong and somebody immediately goes um actually and then they proceed to explain to you very pedantically why you're wrong with like clear like references and
examples of why and and just like very clear proof have have you been there have you um have you felt that or or are you more the um actually person but that's what community notes are basically then you have you know Mega billionaire unkind Tech genius that by the way a lot of people are very divided we did a post you know in this community like a lot of you simply do not trust Sam Alman but you have this conversation just breaking down in it between I mean basically the post author Legacy Media the billionaire
Tech founder investor that is running this company and you could say has a big effect on this technology coming out into the world and then a random Anonymous troll leaker who is I mean probably not Sam alman's alt account that he uses to leak information that no one has access to at all in the industry so those are two separate entities you know what I'm going to come back to that idea I'm going down this crazy rabbit hle way too early in this video but they do note Sam's you know fake news comment here and
basically what he said in that kind of exchange not the cave and the Pikachu bit the bit about the uh a lot of other great technology being released now Orion has been teased by an openingi executive as potentially up to 100 times more powerful than GPT 4 now this this might be a little bit misleading because I think what they're referring to is if you recall that presentation by Microsoft there was the kind of the idea that you know GPT 5 is going to be massive much bigger than GPT 4 four and the reason for
that importantly to find to to point out is it wasn't that the model was 100 times bigger that number was sort of like relative to the previous one if you take into account algorithmic improvements so it's like if you have something powered by one Nvidia card and then you add another Nvidia card then you can say that it's 2x but if you improve the model efficiency you say you double the model efficiency then you can say well now it's 4X so that's what that 100x is referring to and they're saying the company's goal is to
combine it LM over time to create an even more capable model that could eventually be called artificial general intelligence or AGI not sure what they mean by the goal is to combine its llms over time certainly we've heard the idea of distributed training there were rumors that you know open eye can like sort of train uh different pieces of it and then combine them a lot of people were saying that basically the amount of sort of power it would take to run to to train some of these models like you you can't somebody was saying
like you can take it down a States power grid if you were able to just basically put all the um data centers in in one place so combining its llms might be referring to to that idea that it's uh some sort of distributed training then they continue was previously reported that opening ey was using 01 code named Strawberry right so that was the big deal that everybody was including myself freaking out about because that was qar so the idea was that the strawberry models to provide synthetic data to train Orion and they referenced this article
by the information now again information I tend to trust they uh they again they they know people it seems like and again this is why you need to be careful where you kind of get your information from because you don't want to get this uh sort of the broken telephone effect is that what it's called where basically like you get the truth and it's passed down passed down passed down until it just doesn't resemble the truth anymore so this is the information so this is what I would consider a trusted Source when we're talking about
reporting on on AI and more specifically on on open AI what's happening in there so everything that they write here as far as we can tell this this all checks out out and this all lines up with everything that we've been hearing going back to the qar leak so strawberry right the model that was previously called qar so we're calling that the the strawberry model so keep that in mind qar strawberry same thing and so how I'm interpreting this so that what they're saying is like it's not clear whether Chad bot version of strawberry that
can boost the performance of GPT 4 and Chad GPT will be good enough to launch this year this was August 27th 2024 and it was that's the 01 model and they're saying the chatbot version is a smaller simplified version of the original strawberry model known as a distillation so they're saying it seeks to maintain the same level of performance as a bigger model while being easier and less costly to operate and this bigger version of strawberry is going to generate data the synthetic data for Orion okay so here's kind of like the problem here so
this is from The Verge so the released of opening eyes last two models GPT 40 and 01 and then they're saying it was previously reported that openai was using 01 code named Strawberry here's the problem I can't even blame them for this because all right here's a chart from opening eyes I think this was from the um one system card so shout out to TP hang who uh posted this I was too lazy to find the uh the original I have it somewhere but so this is showing the math performance of these models versus inference
cost so again remember the big breakthrough with these models or at least the new thing that's so cool about them is that up until now how you improve the model is you would improve it during training you do training fine tuning Etc so here's for example G G PT 40 GPT 40 mini the previous kind of generation and we would ask them a question and we're like what's the answer and they got to answer right away they can't take their time and think about a little bit so that's what inference means inference means like prediction
this is what these models do you ask them a question they predict what the answer would be and then there's uh training time compute so while we're training that's how much resources we give it how much compute we give it versus inference time compute that's how much resources and time we give it to answer the question to think about answering the question so it's like think about it is like if let's say you're trying hire a lawyer and you're like let me give you a lot of money so you work on this really really hard
for a long time to make sure that the best possible thing happens right that's the cost of letting these models work and think Etc when you ask them the question and so the previous models they just have a set sort of like they just answer and then these 01 models what makes them unique is we're allowed to give them more time to think or you know the cost you you we pay more resources to have them think and as you can see here with the 01 mini right so it improves like this dot line shows
you how well you know as the inference cost increases so do the results where it gets kind of confusing is the thing that we all been playing with is not the 01 gasp right you don't have to take my word for it you can go into chbt and see what models you have available to you you'll notice that the 01 model is nowhere to be seen we have the 01 preview and the 01 mini the thing that everybody refers to as 01 is 01 preview that's this purple thing here and the 01 that's the unreleased
model the actual unreleased model the actual strawberry model right that gets the highest score on this uh aim which is like a math Olympiad for the best and brightest I believe high schoolers in in in America basically like high math performance you know the o1 were basically able to really increase its ability to solve those kind of problems by giving it by by paying more in terms of resources in terms of compute and allowing you to think and process the stuff more after we ask the question it's like use the question think about it tell
me what the answer is in a little bit so again this is the uh information so this is from the same article that's from August so this was before this was released they're saying the chatbot version is a smaller simplified version of the original strawberry model known as a distillation so it's a strawberry light so to speak AKA 01 are you confused yet the reason I mentioned is because of how many people get this little detail wrong interestingly even an exop um former member of the technical staff at openi William Saunders you know testifying in
front of the US Senate committee so about you know some of the potential issues like talking about the the good and the bad threats potentially of AGI Etc talking about open AI it's funny because in that sort of testimony right he's like open AI announced a new AI system called gp01 so again I'm not sure how significant that is that everybody gets it wrong I just it's a little bit hilarious and also why doesn't open the eye just call things in a way that um you know makes sense now in interestingly they do post this
tweet by Sam Alman from a while ago from September 13th 2024 saying I love being home in the midwest the night sky is so beautiful excited for the winter constellations to rise soon they are so great this was taken by most to mean that Orion is coming in the wintertime so certainly it would line up with this article but hopefully this isn't the only sort of thing that they're going on they're saying they do have a source that knows about this but again you know Sam Alman saying fake news out of control to this specific
article and more specifically saying just offends me how media is willing to print random fantasy so this is Jimmy apples back in September talking about the strawberry release uh saying you know view strawberry in context of better models to come and then the sort of the the big boy GPT 5 potentially as early as December but just you know for your sanity so we don't uh die from the hype you know You' probably should think of it as q1 2 of 2025 so let me know what you think about this do you think that Kylie
Robinson is spreading fake news or is it Sam Alman that's wielding the Cloak and Dagger are we going to see the Orion model within Chad gbt or is that meant for other uses like for creating other models are we going to see it in December the reason I mentioned some of the stuff feels like we're living in a similation a little bit number one is because it does seem like we're beginning to create intelligence it's not far-fetched I think to say that we could potentially create our own simulation at some point with little chat gpts
running around convinced that they are the real deal two is we realize now what the simulations are for cuz back in the days we kind of didn't know we like oh it's for you know science experiment or something something something now we're seeing places like Nvidia and many many others building simulations because it provides incredible data data like for example for training robots or for simulating certain social interactions many other things so yes it's science but it's specifically we're able to extract data and data is really important we're realizing that simulations can be almost seen
as like kind of a like what oil wells used to be in the Heyday of oil now it's all about data and three as this kind of like AI AGI this incredibly powerful and interesting technology is emerging into the world we also happen to be at kind of the peak of social media and transparency where people like Sam Alman right because a lot of us follow him for news and events and stuff like that when a publishing Corporation prints something he's able to respond he's able to argue to go go back and forth other people
can jump in and kind of add their own opinions the point is it it didn't have to be like this it didn't have to be this entertaining it didn't have to be this transparent just like attention is all you need for AI for Transformers for the AI architecture to work really well same thing here attention is needed or at least very much desired for us watching the stuff as it unfolds making decisions by hitting the little uh hard icons and the hint hint the like button down below please much appreciated when I set up this
green screen so I can appear on camera in front of various news articles and Twitter posts it wasn't working out too well it had these little grainy I'm sure you can see on the camera now kind of it has this like a grainy effect it's got the green outline around me it was a poor quality job done by somebody that doesn't really know how to do it right that's me but a lot of you really liked the cool Matrix SL simulation effect that this provides some of you have even asked how to recreate that in
your own videos so now moving forward I'm going to pretend that this is exactly the effect that I was going for it was all planned and meticulously crafted with that said if you made it this far thank you so much for watching I truly appreciate you and I hope to see you in the next my name is West rth and talk soon
