Big AI News : OpenAI In Major Trouble, Googles New Robots, Chinas New AI Models, And More...

so in AI news this week the AI race might actually be over now I did a video about this probably yesterday but they are talking about the AI race being over now tldr basically being too long didn't read is that this is where open ey is saying that look if we aren't allowed to train on copyrighted works then the AI race might be over because other companies and countries are going to do that and they're going to fast a passess now in that video I actually spoke about the fact that this would have probably been true Maybe 2 years ago but the fact is now is that we are looking at an AI space where things have changed and what I spoke about was the fact that we are in a situation where yes there is a lot of reasons for them to allow these companies to train on copyrighted works but I'm actually thinking about the fact that data is no longer the big thing that these companies are using to advance their model there are actually a

series of many new Innovations such as test time compute that these companies are probably going to be using so when they're stating that the AI race being over if training on copyrighted Works isn't fair use I think they're just talking about the lawsuits that are probably going to put them in a mountain of debt and legal troubles rather than them not being able to actually access that data for future model either way these companies are in a really strange spot because on one hand technically yes the models don't regurgitate exactly what they've seen in the training data but on the other hand there are instances where really specific use cases or really specific cases of the model taking something from the training data is really clear here I don't have examples for you now but trust me guys there are clear-cut examples where the AI is taking something exactly from the training data so I genuinely do not know

what kind of conclusion they will come to because it is a real gray area where maybe opening eye Majors have to pay back some compensation to those users who are most effective now another thing that I saw that was probably the most underrated story in AI was the fact that the tide is rising everywhere so what I

mean by this is reference to a tweet by noan Brown and noan Brown is the lead at reasoning an open Ai and he actually tweeted something rather interesting he tweeted about how there is the case where creative writing outputs have been a real feel the AGI moment for some folks open Ai and the pessimist line has lately been only stuff like code and math will keep getting better the fuzzy subjective bits will sto nope the tide is rising everywhere now essentially what he's talking about here is of course he's firstly referencing Sam alman's tweet and he's stating that we trained a new model that is good at creative writing we are not sure yet how it's going to get released but this is the first time that I've been struck by something that was written by AI now I really do think that is of course probably true considering the fact that when we look at GPT 4. 5 that is the model that is really smart when it comes to EQ which is its ability to be emotionally intelligent and considering they've trained a model that was rather similar to that it wouldn't surprise me if they now decided to train another model that is a little bit better and steer it towards creative writing so this is something that doesn't surprise me at all what does surprise me though is the fact that this is something that open AI are kind of working on because I do know that the large use cases for AI are mainly and the reason I say that this is rather surprising is because when I look to opening ey from a strategical kind of view I was thinking that opening ey is a company that is really just going to hone down and focus on you know statistics and math and reasoning and just constantly just crushing those kind of you know quantitative benchmarks where we can really look at the numbers

but it seems like opening ey is starting to realize that there is just something about having a model that just truly understands you that is just truly valuable and that's I think why a lot of people have just used anthropic for a large amount of time and somehow people are now starting to switch back to GPT 4. 5 considering that that's the model that is somewhat creative and it is a lot larger and the model is truly interesting so um I think this is a really interesting situation because they're stating that the tide is rising everywhere so certain areas even if they seem to you know spearhead the kind of benchmarked area everywhere is still going to increase and I think that's the case not just for llms but in terms of IM generation Robotics and a lot of other areas that you probably wouldn't expect and that's something that I've

come to understand is that every area of AI is constantly outperforming what I thought it would do even as someone who monitors the space and watches space and it's honestly rather fascinating to see how quick things are going so as well we did have figures bot X and Bot Q is figures new Factory designed to produce humanoid robots initially making up to 12,000 robots per year with plans to significantly increase production to maintain high quality figure manages the entire manufacturing process inhouse building Essential Software tools like Mees Erp and WMS figure redesigned their robots from the Prototype figure 02 to the production ready figure 0 3 to simplify construction replacing slower expensive methods with faster and cheaper techniques like injection molding and metal stamping they also established new teams focused on reliability and safety testing due to limited existing Supply chains for humanoid robots figure designs many robot components internally and carefully partners with specialized suppliers a dedicated manufacturing team ensures assembly efficiency strategically automating tasks like battery testing and gear lubrication to improve quality and speed a key Innovation is using figur own robots powered by their internal AI Helix to assemble other robots and handle materials on the production line This approach boosts efficiency reduces repetitive human labor and positions figure at the Forefront of scalable robot manufacturing now one of the key

AI updates that have happened this week that I truly don't believe has been priced in yet in the sense that we haven't seen the full scale ramifications of this because the model is so good that every day with no exaggeration I've seen a different use case what I'm referring to is Google's Gemini images and what I mean by that is that this model is essentially able to generate images in a way that you truly haven't thought of before because it is like Photoshop in the sense that you can make a character do something like you can change the background to be black or white but you can also take a character and have them in 3D space and just imagine what they would do and have it with a ridiculous degree of accuracy now essentially what I mean by that was this example right here which you can see this person says make me this character then they said put this character in the game with a gameplay screenshot and then you can see that the gameplay screenshot was pretty accurate and whilst that in and of itself would be pretty crazy the craziest thing about this was the fact that like they were controlling the character not in terms of like being able to play the game with the controls and stuff but they were able to say okay can you make the character move over here can you make it climb this wall and everything in that background looked super so I think it just goes to show that this is super insane in the sense that the level of control you can have with these AI systems and this is kind of interesting because this just tells us that these AI systems are a little bit more advanced than we thought because this isn't just some kind of you know AI manipulation engine what they actually talk about in the paper is that this Google AI model is actually really smart because it has a really good internal World model that's why it's able to move things around with such remarkable accuracy for example here you can see it you know asking the character to walk over and then you can see it you know having the character climb that wall having it closer this is just some truly interesting stuff and I can't imagine people not using this for a million different use cases I know I'm certainly going to be teaching a few things in my AI Academy because this opens up a variety of Incredible use cases that I don't think anyone is going to be talking about but that's for another time and an example of this I just grabbed an image of Deadpool you know in fact I said make an image of Deadpool then I said make him fold his hands which it did it had the same image then I said make him stand on one leg and I was like okay that's then make him wear a suit and you can see right there that he's wearing a suit so this was something that was super interesting I don't know I just found it to be super intriguing on you know how quickly we were able to you know just change different things about this image I mean it's definitely going to save you a lot of time if you're working with images that you need to manipulate so I

definitely add this to your AI workflow especially if you work with images on a day-to-day basis and you need things to be a little bit more granular then we had Gemma 3 which is the open source mod news for the week I know I said that about a lot of stories but this one was underrated because if this was released by a different company like a Chinese company the world would have you know spun on its head because they would have been like oh my gosh China has done it again but Google has done it and they haven't really received any credit I got to be honest I don't know what it is about Google ability or credit for the things that they do as much as other companies do and it's honestly quite frustrating because Google literally came up with the architecture that chat gbt is powered by they have you know some of the best reasoning models on the planet and so I don't know if Google has a PR problem or whatever it is but you can see that Google made an open-source model that is 27 billion parameters it cost a fraction of everything else to train and on the chatboard arena elos score it performed better than o03 mini llama 3 mystro large and deep seek V3 only coming behind deep seek R1 and guys that is a 27 billion parameter model basically being on par with models that are 671 billion parameters which is I don't even like I honestly I haven't used a model yet so I can't speak from firsthand but having a model that small formed that well on the chatbot Arena elos score which is a qualitative Benchmark when users feel that the model is better than others is truly surprising because in terms of the size cost and efficiency that thing has to be absolutely incredible and like I said already if China did this people would be like well it's over for the USA so I think this clearly shows us that you know AI models are going to get cheaper more efficient and faster which is kind of something that we already knew but honestly a7b model performing at this level maybe I'll just have to you know run a couple of these as some AI agents because it's definitely super interesting now I will say the model does hallucinate a little bit but you can see right here that model performance compared to the size Gemma 327b is just simply in a league of its own so this is why if you're into the open source area you definitely want to check this out now also there was also this robotic update which is where we've got cougle's new robots this is

something that most people we're bringing Gemini 2. 0's intelligence to general purpose robotic agents in the physical world to be helpful robots need to be interactive responding live to your actions and your voice they need to be dextrous to complete your most complex tasks and they need to be gener to understand things in your 3D world and all of these capabilities need to work across different physical forms we're bringing this together in Gemini robotics our most advanced Vision language action model Gemini robotics is interactive can you put the bananas in the clear container notice how we move the objects and the model reacts and replans on the Fly can you put the grapes in that clear container our model's low latency means it can respond live to rapidly changing conditions and instructions the same model can generalize to all kinds of applications where you can collaborate with the robot live Gemini robotics is dextrous high dexterity tasks are some of the biggest challenges in robotics I can fold the orange square into an origami fox that sounds fun why don't we try that did you know that the word origami comes from the Japanese words Ori meaning to fold and Kami meaning paper these capabilities are enabled by Gemini 2. 0's spatial understanding of detailed aspects of things in your world can point to where the eye should be drawn on the most importantly Gemini robotics is General it uses Gemini 2. 0's World understanding to generalize across a vast range of real world tasks can you flip the red D so that it matches the number on the green D many robots can execute predefined actions but these movements are not predefined the robot is reasoning both about what it sees and how to move it figures out how to make the red Dy match just like we asked and this generalization goes even further this same model can generalize to tasks like this one that it's never been trained to do pick up the basketball and slam dunk it keep in mind these are objects the robot has never seen before but by leveraging Gemini 2. 0's understanding of Concepts like basketball and slam dunks the robot figures out the task we're now inviting more Partners to join our trusted testers program where we're working together to build the next generation of robotic AI agents learn more about how we're bringing Gemini to the physical world at Deep mind. gooogle robotics now there was also this which is detecting misbehavior in Frontier reasoning models and this was an eye opener because it actually talks about something that the AI industry has been talking about for a very long time and essentially what they

were talking about is that the frontier models they exploit loopholes when given the chance and they kind of said okay so since this AI is exploiting this Loop P let's actually see if we can monitor these Loop PS and try to you know give them a sort of punish pment if they try and exploit these loopholes but the problem is that when we punish them when they make these little sneaky things happen they only start to hide these sneaky things and they don't actually stop doing means it becomes even harder for us to realize what's going on so it's basically really hard to control these models if we don't know what they're doing and this is basically one of the things so I'm guessing that you know companies are trying to figure this out because if you have a model that you're trying to you know make sure it's aligned make sure in the right way and make sure that it's incentivized to do the right things of course you're going to want to understand what's going on underneath the hood and this is of course probably not underneath interpretability research but definitely in alignment research because the problem is that when these models reward hack they essentially just break the game in a way that isn't what you want and I made a video on this but I basically spoke about the fact that the researchers in a previous video like from seven years ago they were you know exploring this very thing and they basically asked an AI to get a high score in Mario and essentially if you're asking an AI to get a high score in a video game you would presume that the AI would just play the video game and get a high score but what the AI did was it just hacked the game edited the code and gave itself the high score and of course that's an example of reward hacking and in this example the AI did something similar and basically the AI just got sneakier and sneakier the more open AI checked for it so it was something that is of course in AI safety but the same thing is that you know it just doesn't stop the model so of course they're going to have to develop a new way to sort of figure out how to stop the model doing that so this was something that was super intriguing to me and I thought that it would be worth the cover now of course as well China did reluce a new model and so we have Ernie 4. 5 which is essentially like GPT 4. 5 except it's from China now one of the cool things about this that I didn't really see anyone talking about was the fact that this can actually analyze videos so that's actually a huge

thing I don't think the fact that it's similar to GPT 40 is there I think the fact that this is a huge multimod model that can do a variety of different things is probably the main selling point of this model because not many Frontier models other than Google can actually analyze videos and if we're going to start to move to agentic you know workspaces and experience this is of course something that we definitely need to take advantage of because analyzing videos is something that basically is akin to human Vision so I'm not sure why this you know capability isn't being worked on more I'm guessing it is probably really expensive to do inference for but this is something that you know companies are going to explore certainly in the future and so with this model they essentially have Ernie 4. 5 which is of course a standard model and then they have Ernie X1 which is a reasoning model which is super cheap so over time the cost of intelligence is really starting to go down but I do have to posee this question are you guys actually going to be using this model because despite the models being cheap I don't really see many people genuinely using models like deep seek in the AI workflows I still see people on many different company levels in their business is actually using the opening eye apis rather than using deep seek so I do want to know if it is just a trend to say that you're going to use this model or if it's just purely air hype because I do think it is probably 7030 probably 70% people saying they're going to use these and the other 30% actually using them and deploying them now of course China didn't stop there we actually did have Manis which was absolutely insane this was the you know probably most groundbreaking story in Ai and I think this one was crazy because it's essentially an AI agent that is really good at browsing the internet and doing a lot of different and the crazy thing about this is that it is a lot better than opi's operator agent by a clear margin and deep research by a decent margin as well

now the thing about this was that I think the reason that this was hyped up so much is because there were only you know certain invite code that you know you could only access if you were I don't even know if you were special but they haven't really given out you know worldwide access to this yet they're still in limited beta or Alpha but the crazi thing about this is that okay this agent came out from China and it was crazy it was crashing benchmarks everyone was like okay opening ey are done someone just released something that's pretty crazy but the craziest thing after this was the fact that like it was later revealed that this was actually clawed with an agentic framework just wrapped on top of an open- Source thing called browser use which i' covered a few weeks back now I'm not hating on Manas for this at all but we definitely have to realize what's going on this should just show you guys that a lot of the AI capabilities that we do want are probably wrapped in some kind of framework that we haven't yet explored Manis has been able to explore that by giving us you know clae 3. 5 son it and showing us clearly what it's capable of when that agent powered with some other open source tools is able to browse the internet and the use cases here are completely incredible but it's definitely something that you guys should use because I think the use cases on the website are going to completely change how you use the internet especially if you're doing any kind of knowledge work now China did not stop there we also had Hunan Turbo S which is a cutting edge AI model that was launched by 10 cent on February the 27th

so yes I'm a little bit late to this one but this one was designed to provide Ultra fast responses and enhanced reasoning capabilities that was another competitor in the AI Market many people have just argued that this is something that is just once again taking a huge Digger opening eyes market share now of course I'm not going to do anything crazy here you can of course see all the benchmarks a lot of times we have seen that competition is getting even more Fierce as time goes on I do super interesting to see how all these models manage to compete because one thing that we do know in business is that competing on Price is not really a strategy that you want to do everyone really loses and the only person that wins is the consumer and I'm not saying that this is not going to be good for you guys and me of course cuz we are the consumers of the models but I'm just thinking that at some point these you know companies are going to start to figure out that wait a minute you know we can't really just compete on price here sure we can do that initially but over the long term we're actually going to have to ensure that we can maintain a profit and have a really good product to as well now um another company that managed to make something really useful was of course mistal so mistol made an OCR so mistro

OCR is basically a powerful new software platform that can read and understand documents that is pretty much better than anything available today it's excellent at recognizing test text math formulas tables and multiple languages and can process thousands of pages quickly and it also allows sensitive organizations to keep data secure by hosting it themselves and people are basically using this to organize business data digitize historic documents to preserve culture speed up scientific research by quickly scanning papers and turning complicated documents into easily searchable and usable information you can try it right now with mistral's API it's definitely usable if you've had some PDFs that you've wanted to scan and analyze in a completely different way now in more AI news I was actually supposed to include this earlier on in the video but remember how I spoke about all of these Chinese models coming out opening ey have basically tried to well not tried

they are trying to kind of I think you could say stomp this out because they basically called a deep seek State controlled and they wanted to ban these Chinese models that's actually what they said that these models are you know basically owned controlled and maintained by China so you know having these models available in the US is somewhat of a security risk now of course on one hand I do completely understand where they're coming from I did a video on this and there is evidence to suggest that some of those claims may be true but we also have to understand the opening eyes clear incentive is for them to not have any other Chinese models in the space because they clearly do take some of open ai's market share and we did see this for the first time when we saw that deep seek was number one on the App Store above chat GPT which is just nothing that you could have predicted so this is something that we really have to you know look at because I do think that overall you know these models probably won't be available in the long term like I just think that is something that's going to happen I think National Security is always something that people take zero chances on so these models are probably going to be banned on government devices but considering many of these models are open source I'm not sure how they would enforce that maybe serving the models but if it's open source and you host it yourself you can pretty much you know provide inference just host up an Amazon server so um it's going to be really hard to do that but you know right now literally a week ago we did see that the Chinese government actually did get involved in deep seek and they're actually not letting certain individuals fly out um and those who work at the company so who knows maybe

opening eye is on but for those of you guys who are wondering is China behind take a look at what Dario amade said at a recent interview but I think we're heading for a world where we open AI Google are building billions maybe in tens of millions of chips costing tens of billions of dollars or more it's very hard for that to be smuggled if we put in place export controls we actually may be able to stop that from happening in China um whereas if we if if if we don't I think they may be at parody with us uh and so you know I was a big supporter of the diffusion rule I've been export controls for several years even before deep seat came out because we saw this Dynamic coming and so I think it's actually one of the most essential things ACR not just AI across all Fields um for the United States national security for us to prevent China from getting millions of these very powerful chips and they weren't the only ones discussing this we can actually see the sales for CEO and ralio actually discussing whether or not they think China is behind so this was a super interesting conversation and they talk about whether or not the US can maintain their Competitive Edge and they actually talk about the fact that China is not far behind which is kind of worrying does the US have a Leading Edge on AI against China Mark I think that both have different approaches right now because you know I wouldn't say that the United States has a significant Leading Edge you just saw one of the lead uh models which was deep emerg but also now as I me Baba's model as well and there's other models different approaches with chips so while there are us chips that are very good and very competitive as we saw in the training of the there's different approaches with chips and that there is nobody has a monopoly on training chips or on where do you think this is going Ray we just took up tariffs 20% how do you think they're going to deal with that and also the coming technology B the United States is uniquely competitive in inventing the best chip it's not uniquely competitive it's uniquely uncompetitive China is behind but not by a lot in the best chips and they try to get the best chip but they're ahead in making more chip and making those chips producing those chips and having those chips work together in application and they're integrating chips with robots robotic so and way ahead on that so the application and the use I mean there's going to be a big competition but that's what it really looks like today and this was something that I saw in a recent interview that I saw literally nobody talking about and it's the fact that AI

they are discussing whether or not it should have a pause button and I think it's interesting because anthropic are the only company that seemed to talk about AI if it's you know a real person that could potentially have right someday and they were going back and forth with the audience and one thing that anthropic did say that they were considering is that they would allow AI to have a button that essentially would you know mean the model is out of distress and would allow the button and say look I don't want to engage with that and it was just interesting because like I said before anthropic are the only company that have actually hired a welfare researcher which means there's someone who works at anthropic that looks into whether or not these models are suffering to any degree and tries to stop it which is super interesting so take a look cuz I didn't see this covered anywhere this isn't just a philosophical question I was surprised to learn there are surprisingly practical things you can do um so you know something we're thinking about starting to deploy is you know when we deploy our mod models in their deployment environments um uh just giving the model a button that says I quit this job that the model can press right it's just some kind of very basic you know preference framework where you say if if hypothesi in the model did have experience and that it hated the job enough giving it the ability to press the button I quit this job um if you find the models pressing this button a lot for things that are really unpleasant you know maybe you should pay some doesn't mean you're convinced but attention it sounds crazy I know it's probably the craziest thing I've said so far hi Trooper Sanders