# AI NEWS : OpenAI's INSANE New Breakthrough, No GPT5?, New Robots, AI Becomes Alive and LLAMA 4 ## Метаданные - **Канал:** TheAIGRID - **YouTube:** https://www.youtube.com/watch?v=izgJF9uJqys - **Дата:** 06.11.2024 - **Длительность:** 29:02 - **Просмотры:** 39,314 ## Описание Prepare for AGI with me - https://www.skool.com/postagiprepardness 🐤 Follow Me on Twitter https://twitter.com/TheAiGrid 🌐 Checkout My website - https://theaigrid.com/ 04:35 AI Plateu 06:42 OpenAI Breakthrough 07:52 AGI Election 09:33 China Pulls Ahead 11:12 LLAMA 4 12:06 AI Is ALIVE! 16:19 Googles Learn LLM 17:20 Robots Climb Ladders 20:30 No GPT-5 22:14 o1 Leaked 22:49 New AI religion 24:56 1 Billion Agents 27:30 500X Cheaper GPT4 Links From Todays Video: https://www.physicalintelligence.company/blog/pi0 https://www.physicalintelligence.company/ https://www.cnbc.com/2024/11/04/jeff-bezos-and-openai-invest-in-robot-startup-physical-intelligence.html https://www.physicalintelligence.company/blog/pi0 https://x.com/GuangyuRobert/status/1852397398112514445 https://arxiv.org/pdf/2411.00114 https://x.com/tsarnick/status/1853898866464358795/video/1 https://x.com/tsarnick/status/1853543272909775038 https://www.transformernews.ai/p/anthropic-ai-welfare-researcher https://www.reddit.com/r/singularity/comments/1gfspw4/thomas_friedman_endorses_kamala_because_he_says/#lightbox https://x.com/tsarnick/status/1852451760885145918 https://x.com/Ahmad_Al_Dahle/status/1851822285377933809 https://x.com/tsarnick/status/1852097723627151646 https://x.com/tsarnick/status/1851740698481188895 https://x.com/tsarnick/status/1852097723627151646/video/1 https://www.reddit.com/r/singularity/comments/1ggpjkn/google_has_quietly_released_learn_about_a_new_ai/ https://www.youtube.com/watch?v=_N2mmcxINmo&t=140s https://www.reddit.com/r/singularity/comments/1gjm1wa/anthropic_tries_to_fight_the_recent_rapid_fall_in/ https://futurism.com/the-byte/openai-leak-o1-model https://www.reddit.com/media?url=https%3A%2F%2Fpreview.redd.it%2Fno-gpt-5-this-year-something-else-v0-w86ys0p0p4yd1.png%3Fauto%3Dwebp%26s%3Daec528682d1fddd162ee4c0edf2ce589ae916298 https://mobile-aloha.github.io/ Welcome to my channel where i bring you the latest breakthroughs in AI. From deep learning to robotics, i cover it all. My videos offer valuable insights and perspectives that will expand your knowledge and understanding of this rapidly evolving field. Be sure to subscribe and stay updated on my latest videos. Was there anything i missed? (For Business Enquiries) contact@theaigrid.com #LLM #Largelanguagemodel #chatgpt #AI #ArtificialIntelligence #MachineLearning #DeepLearning #NeuralNetworks #Robotics #DataScience ## Содержание ### [4:35](https://www.youtube.com/watch?v=izgJF9uJqys&t=275s) AI Plateu could really be applied to in the future now I don't know about you guys but there is a lot of different things swelling around the AI Community now Mark Anderson on a podcast basically said recently that AI models are hitting a ceiling SL wall in terms of capabilities remember the phrase deep learning is sign War maybe that's what's happening right now or at least that's what this billionaire believes and of course you know I did hear inclinations that Google Deep Mind is having trouble with their second generation of Gemini like the results aren't as promising as they want them to be but I wonder what you guys think me personally I'm kind of 50/50 cuz there's a video coming out later today on my channel I'm not sure when I'm going to post it but it doesn't pose the best kind of implications for the future of AI research because it kind of shows that AI models might not be as smart as we do think there is this very interesting ASM totic kind of thing that's happening right now um where you know two years ago there was one you know llm that was like way out ahead of everybody else's which was Open the eyes and sitting here today there's like six um that are like on par with that and interestingly at least for right now they're all sort of ASM toting at the at sort of the same point they're kind of hitting the same ceiling um on capabilities now they you know well there's lots of smart people in the industry working to break through those ceilings but you know sitting here today if you just looked at the data charts of performance over time what you would say is there's at least a local topping out of capabilities that's happening right yeah if and if you look at like the improvement from GPT uh 2. 0 to GPT 3 to 3. 5 and then compare that from like 3. 5 to 4 you know we've really slowed down in terms of the amount of improvement and the thing to note on that is the GPU increase was comparable so we're increasing gpus at the same like rate um but we're not getting the intelligence improvements at all I would have to say I disagree on this just because of the fact that there's 01 but then again like I said before I'm 50/50 based on some new research coming out one of them is Apple's recent paper and a second thing is some research coming out of China that is has it has some really Stu implications for the future of AI but um I think be solved so I guess we will have to see because a lot of things are just speculation so it will be interesting to see what new paradigms do pop up now what's funny about this is that like these guys are saying okay it's topping out of AI capabilities but take a look at this recently in this 30 second clip samman ### [6:42](https://www.youtube.com/watch?v=izgJF9uJqys&t=402s) OpenAI Breakthrough actually talks about an AI result that he says I can't talk about it but it is breathtakingly good so it seems that open AI may have made another significant breakthrough which could be quite similar along the lines of strawberry I don't know what breakthrough it is these air Labs I generally think that they have breakthroughs going on all the time but uh take a look at this cuz this is pretty crazy cuz on one hand you have you know people saying oh the capabilities are you know topping out things are not going as good as they are and other side you got like hey buddy uh things are going pretty crazy so I want you guys to see this because I know some people might think it is just AI hype but then again I guess 01 did live up to expectations and that probably will come later today or tomorrow what in AI does no one focus on that everyone should spend more time what I would love to see there's a lot of different ways to solve this problem but something about an AI that can understand your whole life doesn't have to like literally be infinite context but some way that you can have an AI agent that like knows everything there is to know about you has access to all of your data things like that what was one thing that surprised you in the last month Sam it's a research result I can't talk about but it is breathtakingly good now look I'm not a politics guy I don't pay attention to politics so I apologize if I'm not on ### [7:52](https://www.youtube.com/watch?v=izgJF9uJqys&t=472s) AGI Election the left or the right I'm just a guy that's focused on AI but I saw this on Reddit last week and I was meant to include this in video but I didn't now this was basically said Thomas fredman endorses Kamala because he says that AGI in the next you know four years is likely and we must ensure superintelligent machines remain aligned with human values as they you know use the powers to go off in their own directions basically guys I think this is uh pretty crazy because one of the things that I think about this election uh whether you're on right or the left honestly I I don't really you know pay attention to that stuff but I think okay the statement that was made okay that this election okay coincides with one of the greatest scientific turning points in human history is a statement that is very profound because the birth of AI is very likely to emerge within the next four years so I think Trump just became predident so with that okay I mean what is the USA going to look like when we do have AGI that happens to come in and of course is able to do a variety of different things I do wonder how the USA is going to handle that but yeah I I think it's going to be a pretty crazy election term because think about it like this the next four years are going to be essentially it's now 2024 that's 2028 and I mean by 2028 a lot of people have predicted that we're going to have super intelligent systems or at least systems that are capable as humans who knows you know there could be walls there could be things that could pop up I mean it really just depends on where the a research goes but this is definitely going to be one of the most important times because when you think about it when you have things like AGI and Asi running around the democracy is going to be absolutely insane I mean are we going to see Technologies are going to be influencing how people vote I mean it's just going to be really crazy and this is what we literally have Alex zard Wang of scale AI okay giant AI company okay basically stating that look okay if we actually do want to maintain our lead ### [9:33](https://www.youtube.com/watch?v=izgJF9uJqys&t=573s) China Pulls Ahead we actually need to do a lot more because China are actually racing ahead and I didn't actually know this okay and I actually pay a lot attention to AI but apparently okay China are training uh systems that are three to five times larger they like they are basically uh like if we're at 100 miles hour China are at 500 to 300 milph which is uh not good when you actually think about the future implications of that so we're going to need to uh speed up a bit my true belief is that um if the United States maintains leadership in AI you avoid a lot of the most devastating scenarios of AI safety the total investment from Tech giants into AI programs will exceed the entire us defense budget at some point you know in the next however many year the we so that I think is a bullish sign for America where we're not seeing the investment is actually from the US government itself and um and so I think it's a very real scenario where in the commercial sector we have Exquisite AI capabilities but none of that Tech translates to um military organ National Security advantages that enable us to prevent all these really bad outcomes just as a simple example you know the United States on uh data and AI the United States Department of Defense on data and AI invests you know depending on how you count it uh 3 to 10 times less than the uh Chinese people liberations army so the pla out of China so um we are as a portfolio dramatically underinvested compared to China which is you know the second largest defense budget in the world and um there's a lot of talk about defense modernization and how we like innovate on defense procurement these are very common conversations in the dod but my concern is look we're just not moving fast enough on it and now we also ### [11:12](https://www.youtube.com/watch?v=izgJF9uJqys&t=672s) LLAMA 4 have meta basically stating that look they recently visited one of their data centers where they're training llama models on a cluster bigger than 100,000 h00 which is I don't even know if my brain can even THM that much which is just absolutely insane so we can see that the trade runs that are currently being conducted are absolutely incredible okay so it's clear that Lama 4 is of course in the works and they're training this on an insane model I mean an insane server so I really do wonder what that model is going to look like the kind of results because I remember even last time when we actually look at llama 3 they said that llama 3 was literally just continuing to get smarter so I mean it's going to be pretty crazy to see exactly what happens I have no idea I'm really excited for llama 4 because llama 3 was really good so it will be interesting to see what llama 4 is if it's like multi modal so I mean they did talk about that in the previous paper if you guys did pay attention to it when it came out so here's another story okay and this one I don't even where do I even begin so this story is ### [12:06](https://www.youtube.com/watch?v=izgJF9uJqys&t=726s) AI Is ALIVE! basically saying that anthropic is hi an AI welfare researcher for those of you guys who actually don't think AI are real human beings and that they are just statistical programs that predict yada yada and they're not humans this is the thing that is going to tell you or I guess you could say it's going to give uh an entirely New Perspective so basically anthropic okay has hired its first full-time employee focused on the welfare of AI systems basically saying that look okay these systems might actually be conscious and if they are conscious okay we actually might need to be a little bit considerate about how we talk to these systems how we interact with these systems and essentially we're going to need someone okay who actually is going to Grapple with their welfare okay so he says um it's the clearest sign yet that a companies are beginning to Grapple with the questions about whether future AI systems might deserve moral consideration and whether that means we might have obligations to care about their welfare so for those of you guys who constantly you know are constantly rude okay to Claude anthropic Gemini okay who say do this you know you're calling it dummy thises and that guys maybe these systems has feelings okay we don't know okay it's still very early on in the grand scheme of things okay it says that Cal fish who joined the company's alignment team since uh you know in mid-september told Transformer that he is tasked with investigating modal welfare okay and what companies should do about it okay and the role involves exploring he philosophical and Technical questions including which capabilities are required for something to be worthy of moral consideration and how we might recognize such capabilities in Ai and what practical steps companies might take to protect AI systems interests if they turn out to havy so basically saying that look these things they might have interests okay and we might have to protect those interests which is absolutely insane okay you know what it is I think it's I think it's a little bit different okay because when we take a look okay at how uh you know opening eye have positioned Claude versus chat gbt I mean look just look at this okay so like one of the things that we can see here okay this is the logo of anthropic okay and Claude okay now if we actually look at the trat GPT logo okay I'm not going to get an image of it right now but look at this image of clae okay what does that look like to you okay does it look like an AI system it it has like a face okay like a face it kind of looks like a human okay and I think that is the difference between anthropic and opening eye okay anthropic I do believe okay that they do believe that these systems are a little bit more human but I think open ey just has taken the opinion and the belief that these systems are just completely tall so it's super interesting because they're both Smart Companies they're both well respected in their fields and they both taken an entirely different approach okay opening eyes branding is look this is a all it's chat gbt but Claude you know anthropic they've given it a name Claude is a name so I mean you know it's it's completely different and it's crazy okay it says the report which fish worked on before joining anthropic argues that there is a realistic possibility that in the near future some future AI systems will be conscious or robustly agentic okay and two criteria believed by many experts to be sufficient for something to be worthy of moral consideration okay and according to the reports authors this possibility means the question of AI welfare is no longer an issue for the sci-fi or distant future but instead something researchers and companies need to start thinking about now it's time to take AI welfare seriously I'm going to make a full video on this it's definitely going to come into the works but guys do you think that AI needs well do you think we should be able to talk to AI how we do I mean I'm always respectful and poite I never say do this dummy or y I know some people do and they don't think it's kind of system but I mean for future AI systems are we going to have to start being a lot more personal with the kind of systems I don't know it's going to be kind of interesting to see exactly how things go but this is one of the stories that I didn't think you would have yet but I'm guessing that they've hired someone for this role it means that they're taking this very seriously because they're putting money behind this and I don't want to say that anthropic would be tight on money but I would say that they're mainly focused on the issues that are most pressing to the company because things are very competitive so they wouldn't waste time on this if they didn't think it was genuinely worth their time so this being the case is truly incredible this is a remarkable kind of thing like one once in a one time event because this is going to be like the first time company's actually done this now Google ### [16:19](https://www.youtube.com/watch?v=izgJF9uJqys&t=979s) Googles Learn LLM has also released quietly released a tool called learn about and this is something that I found on Reddit I'll leave a post uh in the description I don't believe this is available in the UK I that is quite unfortunate you know when you live in the UK this is the kind of thing that happens maybe possibly in the EU but if you are in America this is probably something that I would say you know you try it's basically like a version of notebook LM but I guess it's a little different Google has a variety of different products and services that integrate AI that are really good but they always put them in this experimental Zone I guess they just don't want the backlash from you know the general public of you know having access to the stuff and then when they have fail something and then of course Google gets the whole p nightmare that they had previously so I would say um I'll leave a link to this in the description but definitely try this out and let me know what it is like because I would like to try this but like I said before I am unfortunately living in the UK okay I mean it's not that bad but you know what I mean um so yeah this is something that seems pretty cool so for those of you who are looking for new AI tools this one by Google is pretty cool now eth Zurich actually came out with a complete ### [17:20](https://www.youtube.com/watch?v=izgJF9uJqys&t=1040s) Robots Climb Ladders Banger of a video where they basically talk about how they now made robot dogs be able to climb ladders okay so for those of you guys thinking okay you can h abusive chat gbt and clae and then run up a ladder and you'll be fine because the humanoid robot won't be able to climb a ladder guys they are doing that too for those of you guys who think you're going to be fine you're not cuz these things are getting the ability to climb up ladders as well so they can sneak around they can jump over these kind of abilities but humans who've always had the ability to climb up ladders okay now we are getting that ability with these robots too I mean it's a crazy video okay I'm going to include it here but uh this is something that I think this kind of research doesn't get respected enough until you see a robot chasing you up a ladder and then you're starting to wonder okay when did we cross the barrier to where robots can now climb ladders I wasn't around for that video in this work we enable robust ladder climbing with quadrupedal robots quadruped robots are proliferating in industrial environments where they carry out routine autonomous inspection missions however quadrupeds are still unable to reliably negotiate one common feature of industrial infrastructure previous research on robotic ladder climbing has been conducted mainly with humanoid robots has been characterized by slow climbing speeds and has not generalized Beyond specific ladder configurations and structured noise-free environments we achieve robust quadriple ladder climbing via a reinforcement learning based control policy and complimentary hooked end effector first we designed a hooked end effector that allows the robot to apply both compressive and tensile forces in the ladder to stabilize its Center of mass while climbing then we trained a control policy in simulation using a privileged teacher student reinforcement learning approach the teacher has access to privileged observations only available in simulation including noisess position and orientation of the ladder external disturbances and other information which helps to learn robust ladder climbing we then distill a student policy which mimics the teacher actions but only has access to noisy onboard observations and estimate of the latter State we deploy the policy zero shot on the real hardware and show a 90% success rate on real ladders the policy's robust to slips and missteps while climbing that arise from inaccurate State estimation we compared the policy which uses the hook end Defector to a policy trained with the regular ball foot we found that the hook substantially outperforms the ball in terms of speed and robustness particularly at steep inclinations where the ball can't apply forces to stabilize the base the robot is also able to climb ladders when subject to unmodeled disturbances here the robot uses the hook to brace itself against the ladder and successfully recovers after being pulled off the ladder in the future we look to use different sensing modalities such as depth images that allow the robot to reliably Traverse ladders in unstructured environments without a motion capture system overall this work expands the application scope of industrial quadraped robots Beyond nominal terrains to challenging infrastructural features in the environment highlighting the synergies between robot morphology and control policy when performing complex skills now more open eye news for those of you who are wondering what we're going to get later this year one thing did forget to include from my video where I ### [20:30](https://www.youtube.com/watch?v=izgJF9uJqys&t=1230s) No GPT-5 actually did the coverage on the Sam Alman slop AMA one thing I didn't real realize was that they actually said that okay we've got some good releases coming later this year but nothing that we're going to call gbt 5 though so for those of you who are thinking okay the election is now over things have now cooled off we're going to get some more releases apparently we're not going to be getting gbt 5 just yet so that means that gbt 5 is quite likely to be sometime next year and I'm guessing that since we do have gbt 40 and now it's quite likely that we're going to get the 014 release before the end of the year this is I guess rather interesting I'm guessing that they're probably saving compute for something else but I do wonder what these very good releases coming later this year are going to be I mean is it just a full1 is it something with their image generation research that they did they actually recently did some image generation research I covered a video that basically dives into exactly what they did and it's actually pretty insane what it is like it was research that unfortunately got overshadowed just because of the fact that like you can't use it yet but I still think that you know now that they've shown us that I think it's something that's like wo open ey are still ahead of the game in many different areas so it will be interesting to see what those releases are I mean it could be a number of different things but the problem is that opening eye literally has so many different products in the pipeline like they will show us okay guys we've got systems that can do x y z but we're only going to lease point a but of course remember ages ago this was their process of iterative deployment which means that they're only going to deploy things iteratively which means if they have a 10 feature system they're only going to release one feature every 2 months so it slowly and surely gets out to the population and like I said before if you weren't paying attention actually yes the full1 model the latest full 01 model did leak on Friday and I would expect ### [22:14](https://www.youtube.com/watch?v=izgJF9uJqys&t=1334s) o1 Leaked this model to be out soon because basically of course in a statement to futurism an opening ey spokesperson actually confirmed so for those of you guys on my video commenting saying you know it was just an accident y y they've actually confirmed that we were preparing limited external access to the opening old one model and ran into an issue and that issue has now been fixed so of course right now you can't use the full1 model you could at that time with that specific link but now unfortunately we cannot do so it will be interesting to see once that model does come out I don't think it's going to take much longer considering people already had access to it but it will be interesting to see now what was also interesting was Bill Gates also said ### [22:49](https://www.youtube.com/watch?v=izgJF9uJqys&t=1369s) New AI religion something in an interview that someone also said I'm going to play the other clip after but basically he said that AI is going to become so good okay and this is genuinely what I want you guys to respond to okay he says AI is going to become so good at solving problems and creating attractive activities with humans that we might need a new religion or New Philosophy to stay connected to one to another basically St the look we're going to be in this abundant age where people are just going to be off you know converging with AIS and agents and doing all sorts of entertaining things that how on Earth are we all going to be living in the same shared reality with each other I want to ask you guys okay uh leave you in the comment section below if AI does take every job and if it does do absolutely everything what are you guys going to completely spend your time doing now that there's no pressure to do anything to an that's what I would love to know and I think it's a thought-provoking question that most people never had the time to answer because number one we have to focus on money number two there's a variety of different problems that face us on a day-to-day basis the potential positive path is so good that it will force us to rethink how should we use our time you know you can almost call it a new religion or a new philosophy okay how do we stay connected with each other not addicted to these things that will make video games look like uh nothing in terms of the attractiveness of spending time on them so it's fascinating that we will the issues of you know disease and enough food of climate if things go well those will largely become solved problems and you know so the Next Generation does get to say okay given that some things that were massively in shortage are now not how do we take advantage of that you know do we ban AI being used in certain Endeavor so that humans get uh some you know like you don't want robots playing baseball probably uh cuz they they'll be too good uh so we'll keep them off uh the field okay how broadly would you go with that we are so used to this shortage world you know I I hope I get to see how we start to rethink the these deep meaning questions we also had this video okay from Altera AI this is a really ### [24:56](https://www.youtube.com/watch?v=izgJF9uJqys&t=1496s) 1 Billion Agents cool company they're actually focusing on you know AI agents okay and they said what would the world look like with a 100 billion digital human beings today we share our tech report on Project Sid a glimpse at the first AI agent civilization powered by our new piano architecture okay so that's something that they're working on and it's really cool because they're basically uh focusing on how AI can sort of merge together in this kind of piano architecture where they're looking at how agents can you know become together and just form these really interesting relationships you can see the conversations they're having the kind of dialogues that they're doing and the kind of relationships that they do I mean it's just really interesting to see how these agents manage to work together now of course this does take place in the sandbox game called Minecraft I think it's a really good uh place for these agents to learn to evolve and adapt because there's just a lot of uh areas where you can control the variables and you can really manage stuff in a really easy way and this is basically where they talk about their new cognitive architecture called piano and this is basically where they say how can AI agents think and act concurrently over multiple time scales consciously and subconsciously it says imagine the keys of a piano as different brain modules played together they can make beautiful cords or in this case emergent human qualities and agents so it's really cool because you get some emergent capabilities when you have ai systems working together and these are the kind of things that they're trying to find out it says can AI agents form meaningful social relationships at scale and we show that agents accurately inferring other sentiments forming lasting friendships and rivalries having introverts naturally form fewer connections than introverts in societies of many agents which is actually pretty cool I'd love to like you know simulate myself going into some kind of world where I could accurately say that okay these are the friends I have these are relationships kind of map it to the real world and see how certain scenarios play out I mean I guess maybe that might be something that happens in the future you can also see interestingly that they said that okay how do specialized roles emerg in AI societies when we initially put identical agents in the village they self-developed distinct professions like farmers and Engineers removing social awareness led to agents choosing more homogeneous roles that didn't persist across time and they also said okay and this is really interesting can agents develop and modify their own laws and we tested this by implementing a tax system they not only followed tax laws but democratically voted to change rates based on public sentiment and then of course here interestingly it said Can AI agents develop their own culture and we observed organic spread of memes and tracked out agents primed with a fictional religion spread through agent societies and rural areas show different cultural patterns versus talents I mean it's so interesting that we're now getting AI agents being collaborating on different time scales and just in different worlds I mean it's just bringing us back to that question of are we just a bunch of AI agents simulation run by a giant alien computer and here ### [27:30](https://www.youtube.com/watch?v=izgJF9uJqys&t=1650s) 500X Cheaper GPT4 we have Kaiu Le okay this is a Taiwanese businessman computer scientist investor and writer who's currently based in Beijing China and he's worked at an executive at Apple and Microsoft and Google okay and this guy is talking about his new AI model that is better than gbt 4 and wait for it 500 times cheaper and he says that we will see phenomenal change in the next 2 years as AI reaches PhD level intelligence we're looking at five to six year Vision but what's going to change in just the next two years will be phenomenal if we want to see what's going to change in the next two years let's look back at what happened in the last 2 years or a year and a half 2 years ago Chad GPT was launched shortly after that GPT 4 we're now a year and a half after gp4 is launched and my company 01 has just launched a new model called e lightning that is better than GPD 4 and 500 times cheaper so what of course other companies have also built great models and at lower costs we are seeing technology getting better and better we're seeing every year in a half or so with additional computation AI goes from a kindergarten to elementary school to high school college and soon PhD level and that will help us do so many tasks uh in our lives and we can almost predictably see what will do given this the change final output so you guys enjoyed the video I will see you in the next one --- *Источник: https://ekstraktznaniy.ru/video/13811*