all right welcome so obviously a slightly different setup than usual I'm in New York right now IBM invited me here to show off all of their new AI technology and I'm going to have a video on that coming soon but for today let's get right into our first story open AI just released their own a gentic framework open source and this is both a surprise and not a surprise because if you remember just a week ago they released Chad GPT canvas and that was a huge upgrade to their interface and why are these things similar well if you think about open AI as a platform company essentially providing the pipes for intelligence they've talked about only doing that but now it's becoming more and more obvious that they're going upstream and I've talked on this channel a lot about platform risk and now we're seeing it now they're also releasing their own agentic framework they said that there's no promises behind it they're not going to even promise updates to it you get no support so really it's just a research project for now now but it's cool to see open AI get into the agent game all right so here it is just a few days after release nearly 13,000 Stars it is open source as I mentioned it is called swarm experimental educational so all of the caveats are there I haven't tested it out yet but it looks to have a decent amount of functionality already it describes itself as swarm explorers patterns that are lightweight scalable and highly customizable by Design approaches similar to swarm are best suited for situations dealing with a large number of independent capabilities and instructions that are difficult to encode into a single prompt essentially agents the assistance API is a great option for developers looking for fully hosted threads and built-in memory management and retrieval however swarm is an educational resource for developers curious to learn about multi-agent orchestration swarm runs almost entirely on the client and much like chat completions API does not store State between calls now because it's open source you can likely plug in any model that you want it is not specific to open AI models but I have a feeling it works best with those models here's an example and oddly enough it is about New York what's the weather in New York triage assistant transfers to the weather assistant so that's kind of unique where it's a really explicit way of transferring between agents transfers to the weather assistant the weather assistant calls the tool get weather New York City gets the answer and then replies with it's 67° in New York city so very basic agent functionality but cool to see them get into the game next Nvidia has delivered one of the first new chipsets to a client and of course the client open Ai and open AI was the first to get one of their previous generations of chips so open Ai and Nvidia are very close together here bojun says yesterday Nvidia delivered one of the first b200s to open AI the spec for these beasts are insane first let's look at this picture I mean it's gorgeous here it is the entire front of it looks like a chip in itself but here's some of the open AI team and you can just see the massive size of this GPU and for one of these you can get it at the low low cost of $400,000 for one now imagine you need a cluster of 20,000 or 50,000 or even what x. a did in 100,000 of them I mean these things are essentially the cost of houses look at these benchmarks so these are training performance and speed up over the h100s here's the h100s at 1X and you can see the dgx b200 is three times faster and inference speed gets an even bigger bump a 15x improvement over the previous generation of Chip so this monster has eight Nvidia Blackwell gpus that is their newest generation of chips, 1440 GB of total vram which is crazy 72 pedop flops training and 144 pedop flops of inference it uses 14. 3 KW Max it comes with two Intel zeeon Platinum 8570 processors 112 cores total and 4 terb of system memory so they are starting to deliver their Blackwell chips I heard there's a backlog and a que for these chips for like years so if you wanted to get your hands on one put your name on the list today's video is is brought to you by mamut mamut AI brings all of the best models together in one place for one price Claude llama GPT 40 mraw Gemini Pro and even gpt1 and rather than having to pay for each of these AI separately you pay $10 to mammut and they bring it all together in one place plus they have image generation mid Journey flux Pro Dolly and stable diffusion again all for $10 models are frequently updated as soon as they're released so be sure to check out mamut for access to all the best models for one lowprice m.
ai that is m a MM o u t h. a thanks again to mamut and let's continue about Nvidia they dropped a monster of a model but I've been traveling so I really haven't had a chance to try it out yet they dropped a fine-tuned version of llama 3. 1 that apparently beats GPT 4 and Claude 3.
5 Sonet it's called the neatr 70b instruct model and it is really doing well on all of the benchmarks but of course benchmarks are useless I want to test it myself I'll be doing that next week but most importantly it can count the number of RS and strawberry so you know it's at least that good so it's been doing well on all the benchmarks a lot of people have good things to say about it and we'll see next week when I test it next Google is getting into nuclear which is kind of crazy but also seems to be a requirement if you want to run massive supercomputers at scale Sundar the CEO of Google just announced today we signed a pioneering agreement to purchase clean energy in the US from chyos Power a leader in building small modular nuclear reactors it's the latest step in our history of accelerating clean energy sources and will help support AI Investments this is not the first nuclear story even in the last month and I love it nuclear power is clean efficient and we basically we being the US shut down all of our power plants after a couple accidents a few decades ago and really have not Revisited it since then we know China is building a ton of nuclear reactors and so the fact that we're starting to get into it and it's led by private Enterprise it's good news for the US and especially small power reactors seems to be the fad right now and so that means smaller more modular spread throughout the country country possibly less risk I actually don't know a ton about nuclear energy maybe that's a topic I should interview somebody on let me know in the comments next Adobe had a bunch of announcements where a lot of people thought Adobe was on their way out meaning dying because all of these AI features text to video text to image were going to essentially kill Adobe but it seems like they're keeping up just fine they're building AI functionality into their Suite of products they just released a text to video product that is commercially safe as they call it meaning it was trained on only things that they own or have rights for Firefly video supports text of video imaged video and is designed for Immaculate prompt coherence they gave a bunch of examples during their Max event this week which I didn't get a chance to attend but really exciting stuff and I'm glad that adobe is able to keep up and apparently these aren't just features that they're throwing into their product that nobody is using generative Phill in Photoshop is Apparently one of the most popular features in it already and I demoed that I showed it off a while ago but it's basically you highlight a certain section of an image and you can just type in any prompt and it will change whatever you're looking at and it was pretty impressive at the time and I can only imagine it's gotten better since then next World coin the human identification framework has launched a bunch of new stuff and is now just called World drop the coin it's cleaner I think that was a great decision I just interviewed actually just today I'm about to hit publish on an interview with Saturn and Puget who is one of the founders of worldcoin also Sam Alman is another founder and what world now is aiming to do is to identify humanness to basically give everybody an identification it's completely open source they made a lot of news a few months ago with their orb device which scans your iris and that's the way they're able to really verify that you're a human and then they take that information they store it in the blockchain and you can essentially verify that you're a human anywhere on the internet and so why would you do that why is that important well when you have artificial intelligence that can imitate humanness online and do so at such an incredible accuracy and you have thousands millions billions of Agents running around the Internet it's going to be important to validate that yes you are human this is me and I am human and so that is what they aim to do they just aim to validate if somebody is a human or not there's an entire infrastructure that can be built on top of that technology and I encourage you to check out my interview with sat because it was amazing we talked about ASI we talked about Ubi we talked about world coin of course and the future with super intelligence next another text of video model is out there and this time it is out and completely open source and why do I say that well it seems like a lot of different companies are announcing text to video models meta open AI and none of them are actually releasing them this one is called pyramid flow sd3 you can find it on hugging face it is based on stable diffusion so stable diffusion continues to gain functionality but anyways check it out let me know what you think next it seems everybody is getting into agents Brett Taylor a Silicon Valley Legend who was the co-ceo of Salesforce he's the head of the board at open AI he just raised a bunch of money for his agent startup and could value it at over $4 billion people are raising rounds with almost nothing I don't know if that's the case for him but if you're a name who has done something in the past and you're in AI now you're going to raise at a huge valuation so his startup is called Sierra an artificial intelligence startup co-founded by former Salesforce co-ceo Brett Taylor is Raising hundreds of millions of dollars in new funding led by growth stage investor Green Oaks Capital according to two people who have knowledge of the deal according to the information it triples the company's valuation which was only a billion dollars back in January Sierra which is selling an AI agent that can automate certain tasks such as customer service including voice calls was founded just over a year ago so yeah a $4 billion valuation sounds about right next open AI seems to be losing a ton of money and anybody who thinks open AI is dying right now that could not be further from the truth they will just raise more and more money but they're just growing so fast but they are still the hottest ticket in town in terms of VC investment and according to the information they have implied losses tripling to 14 billion in 2026 that is a ton of money to lose but again they just closed the biggest private round in history so they have some cash reserves to burn and all of this is going towards growing their team building new models doing Partnerships and acquiring new data to train those models so there's a lot of uses for that cash but of course they have to be careful next andal Industries founded by the co-founder of oculus which was acquired by meta the founder was really fired from meta and then he went on with a huge chip on his shoulder to found andil a defense tech company that really merges artificial intelligence into defense Tech now just released a new product called bolt bolt is a family of man packable autonomous air vehicles featuring both ISR and munition variants basically the future of Warfare is with these little drones and this drone is super capable powered by AI I find this technology to be really fascinating it's a mixture of drones and artificial intelligence and other Tech and so even though it is for war it is really cool technology that they're building introducing the bolt family of man packable autonomous vehicles featuring both ISR and munition variants bolt M equips Ground Forces with simple lethal and reliable Precision Firepower it uses advanced economy software to enable operators to focus on four simple decisions where to look what to follow how to engage and when to strike basically it's a drone packed with an explosive you give it a Target it flies over it it can do reconnaissance it could do anything and if you needed to it could also explode next SpaceX did something that seemed impossible only a year or two ago they launched a three-story building into the air in the form of a rocket it went up came back down and was caught by two chopsticks midair and it is so cool let me just play the video briefly we can see those chopck [Applause] now now why is this so important well first of all what used to happen with these boosters is it would go up into the air and then would fall into the ocean and basically just disintegrate and they would have to collect the pieces and rebuild them over a long period of time then SpaceX started allowing these boosters to be brought back down to earth and landing in a position where they could be reused and not just explode on impact now they're actually going to be able to catch the booster refill it add the main cabin on top and then launch it back into space all within a few hours so the reusability and the speed at which we're getting these Rockets back into orbit is going to accelerate greatly and that just means that we are that much closer to becoming a multiplanetary species next it seems like the tensions between Microsoft and open AI continue to grow Sebastian bubeck who was one of the main researchers at Microsoft just left Microsoft to go to open AI so this is kind of a win for open AI because lately they've been having major brain drain BBC was a public face of much of Microsoft's development of generative AI models over the past P two years according to the information his team used special access to technology from open AI to purer its research thanks to financial and product development partnership between the companies again Microsoft and open aai are very close Microsoft owns about half of open AI they have had a lot of friction in the past with Satia the CEO of Microsoft saying even if open AI shut down today we have all of their Tech we could recreate what they do whether that's true or not and they also have stated in public reports that they consider open AI to be a competitor so it'll be interesting to see how that story plays out next we have a new non- Transformer model this one is from a company called Zyra Ai and they announced today in collaboration with Nvidia we bring you Zamba 27b a hybrid SSM model that outperforms mistal Gemma nama3 and other leading models in both quality and speed it is the leading model for the less than billion parameter weight class and so here we can see the MML U score on the left on the Y AIS and the time to First token in an 8K input sequence length on the xaxis so not only is it the fastest in terms of time to First token but it is also the highest quality on the MML U Benchmark compared to llama 3. 1 mraw 7B and Gemma 7B now you know how I feel about benchmarks it's great that they're putting this Benchmark out there but I don't trust it until I actually use the model and if you've watched this Channel at all you know that I have not had a ton of luck with non- Transformers models but either way I'm very happy to see another open source model out there and I'm very appreciative to Zyra for putting it out there doing the work and contributing to the open source AI Community do you want me to test this model of course let me know in the comments next apparently everybody's going to get access to search GPT soon and it's not really search GPT like we've kind of had beta access to it's really just search GPT built into chat GPT if you search something that requires real-time information then all of a sudden it's going to be able to search the web which it has already kind of been able to do so it's this weird new hybrid approach where they're using some of the tech from search GPT and then putting it into the chat GPT product now that makes sense because they didn't say search GPT was going to be an official product long term they just called it an experiment but now here's some evidence that it's going to be showing up in chat GPT for all users next mraw has dropped some incredible Edge models and you know by watching this channel I love Edge models I love being able to run small models on edge devices whether that's a computer a laptop a phone whatever it is I strongly believe in a future where all of these models are very vertical very specialized and can run on consumer hard Ware we don't need an 01 level model for 98% of use cases and if we fine-tune these models to do the right use case at the right time they can be really good but also really efficient really cost efficient as well and have very low latency so I really believe in that architecture of having very small models deployed on lots of different devices and now we have some new ones for mraw we are proud to announce two state-of-the-art models for on device Compu Computing and at the edge use cases we call them less minist stra I hope I'm saying that right I know I'm not please tell me in the comments mini St 3B and mini St 8B so these are very small models 8B is more common but 3B is starting to be even more common and a small model so I definitely love it these models set a new frontier and knowledge common sense reasoning function calling and efficiency in the sub 10B category and can be used or tuned to a variety of uses from orchestrating agentic workflows to creating specialist task workers both models support up to 128k context length and minrol AP has a special interleaved sliding window attention pattern for faster and memory efficient inference so here are some of the benchmarks this is Gemma 22b llama 3. 2 3B and mini St 3B as you can see minist St across the board is just dominating and in the 8B or so class we also see mini straw really just dominating except for the human eval with llama 3.
