Our first story of the day is about GPT-5. We have news from OpenAI itself, so let's go through it. The CEO of OpenAI Japan says GPT-next will be released this year. GPT-next being their next Frontier Model, and its effective computational load is 100 times greater than GPT-4. So, thank you to AI safety memes for putting together this little summary of what happened. And as we're seeing here, the CEO of OpenAI Japan put together a presentation, gave a little talk, and revealed some new information. This is what he said, translated to English: GPT-4-next, which makes it seem like this is just another iteration of GPT-4, but I think it's more likely to be GPT-5, which will be released this year, is expected to be trained using a miniature version of strawberry with roughly the same computational resources as GPT-4. Now, that is exactly what we've been reporting on for the last week or so. The strawberry QAR technology is being used to train these new models.

It's creating synthetic data for these new models with an effective computational load 100 times greater. The AI model called GPT next that will be released in the future will evolve nearly 100 times based on past performance. Unlike traditional software, AI technology grows exponentially. This slide, and that's what we're looking at right here, clearly states 2024 GPT next. This 100 times increase probably does not refer to the scaling of computing resources but rather to the effective computational volume, plus two orders of magnitude, including improvements to the architecture and learning efficiency. So hopefully doing more with the same amount of data is what that means. Orion, which has been in the spotlight recently, was trained for several months on the equivalent of 10,000 h100s. Compared to GPT 4, adding 10 times the computational resource scale, making it plus three orders of magnitude, and is expected to be released sometime next year. So a little bit of new information about the next versions of chat GPT, and honestly, 10,000 h100s kind

of seem small now compared to what xai just built out over the weekend, which we'll touch on right now. So Elon Musk posted over the weekend. This is just yesterday, actually, as of recording this video, so maybe two days ago. This weekend, the xai team brought our Colossus 100,000 h100 training cluster online. From start to finish, it was done in 122 days. That's insane. Colossus is the most powerful AI training system in the world. Moreover, it will double in size to 200k 50K H2 200s in a few months. Excellent work by the team, Nvidia, and many of our partners, suppliers. Jason Calanis says that's insane. Lex Freeman says congrats. Now the one company that they didn't mention by name is Dell, but Michael Dell says it's amazing how fast this was done, and it's an honor for dtech to be part of this important AI training system. Congrats. And Anthony Everywhere has this nice AI company GPU comparison. x aai 100,000 Google AI 990,000 open Ai.

880,000 and so on and so forth. So, that is a massive supercomputer that they're going to be using to train Grok 3 in future iterations of Grok. And I love the competition. I love seeing this up and running. Congratulations to the team. Next, introducing Project Sid, the first simulations of 1,000 truly autonomous agents collaborating in a virtual world. So, we've talked about stuff like this on this channel many times over the past couple of years, from the Simul paper to some of the Nvidia work from Dr. Jim Fan and team being able to essentially figure out how to play Minecraft really well. But now, they've actually placed 1,000 plus autonomous agents in the world of Minecraft and let them live with no specified goal. And they are collaborating in a virtual world with emergent economy, culture, religion, and government. Humans are the only species to land on the moon because we can cooperate at vast scale. Can AI do the same? So, a few things about this research at Altera. We asked whether our agents

can organize at unprecedented scale to achieve what individual agents can't. We saw agents form a merchant hub, vote in a democracy, spread religions, and collect five times more distinct items than ever before. Now, I think this actually just makes perfect sense. Agents are based on artificial intelligence; artificial intelligence is based on data created by humans. So, all of this stuff is really just a derivative of what humans and humanity do. So, you can play it right now at this URL. And if we look at this graph, this is long-term benchmark distinct item collection in Minecraft. Now, they have React and Auto GPT, which previously were incredible at exploring and learning in Minecraft, but Voyager, which is the project by the Nvidia team and Dr. Jim Fan, did much, much better. But now, look at this: Altera 32% of total distinct items, so multiple times better than Voyager. And so, what does that actually mean? That means that their agents, when collaborating together, thousands of them, are able to create and discover many more items and explore the world of Minecraft at a much more

effective rate than any of the previous agentic frameworks. Here, by truly autonomous agents, we refer specifically to machines that are long-term autonomous, can progress for hours and days by themselves, can collaborate with one another to achieve what individuals cannot, have their own deep and robust drives to support humans, and can verbally express their conscious thoughts and feelings. Why are these qualities important as we progress to the next level of AI? Humans can't afford to micromanage every AI like we do with ChatGPT. They must be more autonomous. That's completely true. How do we trust their autonomous decisions? They can't be just trained to act like humans. Their deepest motivations must be aligned with human values. They not only need to be able to talk; their words must largely reflect their inner thoughts and decisions. As AI agents become an integral part of human civilization, they must effectively collaborate with each other and the rest of us, like any other first attempt. Project Sid opens up more questions than

it answers how can we meaningfully measure true long-term progression in a general way, how to build agents that actually care about human growth, how to build machines consciously aware of their actions and the consequences, and how far can we scale an AI civilization. Why can't we have billions of agents, trillions of agents, all working together? And I think we're going to see some crazy new novel emergent behavior when we start having so many agents that it is orders of magnitude greater than what human civilization is at 8 billion. And of course, Plany the Liberator replies with Birdman rubbing his hands together, getting ready to completely jailbreak all of those agents. But very cool, thank you Robert Yang for sharing, and I definitely want to check this out. I'll drop a link in the description below. Next from the information, Meta considers mixed reality glasses codenamed Puffing in. So their Meta X Rayband collaboration was and is incredibly popular. I have my own pair, I love them, I don't really use it all that much for

the AI capabilities, but I use it all the time for the camera functionality and occasionally for AI, but mostly because I think they look great. They look just like any other Ray-Bans, and they take photos in a hands-free way. And so now they're doubling down with mixed reality glasses. Now let's think for a second what happened with the Apple Vision Pro. They released this $3,500 incredibly heavy, incredibly sophisticated mixed reality headset, and it was not that popular. Everybody who wanted to try it bought it, a lot of people returned it, and basically nobody, or very few people, are actually using it day-to-day today. And that includes myself. I bought them super excited about them. I see the tech, I see the potential, but actually putting the headset on, getting started, it's a long process and pretty tedious as well, so I don't really use it all that much. And what the rumors were is Apple has delayed their version two of the Apple Vision Pro and instead are working on a lighter, cheaper mixed reality glasses. And it seems like Meta is going in the

same direction. The company is looking to design an alternative to its large, heavy headsets, which so far have had limited consumer appeal. Staff at Meta's Reality Lab showed ideas for the new glasses-like mixed reality device to Meta CEO Mark Zuckerberg and Meta Chief Technology Officer Andrew Bosworth earlier this month as part of a review of the division's product roadmap. In the wake of the meeting, staff decided to move forward with Puffin and cancel a premium mixed reality headset that was in development. So, it seems like Apple and Meta are exactly aligned with their strategy. People don't want to start with these hyper high-end, very expensive headsets; they want something that is usable day-to-day, affordable, and more practical overall. And yeah, right here it says Apple recently suspended work on a second version of its Vision Pro headset. Puffin would be a much smaller device than Meta's existing Quest headsets. The device has been described internally as resembling bulky glasses, estimated to weigh less than 110 grams. That's twice as heavy as Meta's Ray-Ban smart glasses.

but if of the weight of the quest 3 headset. And so when you take the glasses, mix reality, you take VR, AR, you mix it all together, you have a really compelling hardware device. Now, the only problem with glasses is that if you don't actually wear glasses for vision correction, you're not wearing sunglasses indoors. So, you really only have a few times throughout the day where you actually need glasses. So, I really am wondering how they're going to solve that. And I don't want to wear glasses indoors if I don't need to wear them. Why would I? So, that is my only gripe with the glasses form factor and why I actually think potentially having something akin to earbuds with a camera on it might actually be the better form factor. But I'm not sure. I don't want to be wearing earbuds all day either. But I think I'd rather wear those more so than glasses all day. But you tell me what do you think. Let me know in the comments.

thanks to the sponsor of this video, Vulture. Reap the benefits of the world's largest independent cloud provider when you bring your GPU workloads over to Vulture. They have the latest Nvidia GPUs spanning 32 locations across six continents. Vulture delivers industry-leading price-to-performance and serious accessibility and reliability. Vulture's global fully composable cloud infrastructure moves your applications closer to your users and frees you from vendor lock-in, allowing you to bring your own networking and database solutions. And if you need to scale beyond just a single cluster, Vulture's Kubernetes engine allows you to take full control over your deployment, offering up a 100% free control plane. So if you're tired of waiting for GPUs from other providers, make sure to use Vulture. You can deploy at any scale immediately, and they have H100s, L4s, and more available right now. And you can get a fraction of a card or fully dedicated bare metal systems, which gives you full control over your hardware and your throughput. They also have one-click

installation of all the applications you might need for advanced machine learning workloads, allowing you to get up and running in minutes, not hours. So experience the vulture difference. Don't get bogged down by severe wait times or limited locations. Try it free today with a $300 credit for your first 30 days when you visit get vulture.com bman and make sure to use code Burman 300 checkout out to get that credit. Thanks again to vulture for sponsoring this video. Next, open AI is in talks to raise funding that would value it at more than 100 billion dollars. The company that was previously a nonprofit seems to be really just raking in the funding, and this is their next funding round. And not only that, Apple is rumored to be one of the investors in this current round, and that wouldn't really surprise me. Apple has a ton of cash in reserves just waiting to be deployed, and Thrive Capital would lead the round and would invest a billion dollars. Earlier this year, open AI was valued at a reported 80 billion from 29 billion the prior

year annualized revenue reportedly surpassed 2 billion earlier this year. Growth took off in late 2022 after the company launched ChatGPT, of course, and continued as it rolled out products for businesses and expanded into AI-generated photos and videos. Now, my opinion is that large language models are becoming a commodity very quickly. We have Meta LLaMA, which is free, and we have a number of other competing models that compete directly with OpenAI's models, and the cost of intelligence, even according to OpenAI, is basically being driven down to zero. So, if that's the case, what is OpenAI's long-term business strategy? Now, that is the real question because I don't see it, not yet at least. And so, to be valued at 100 billion where your core product is becoming commoditized so quickly is a little astounding to me. But you know what? The team at OpenAI is incredibly smart. Sam Altman is incredibly smart. I have a feeling they have plans, and they're going to figure it out. I don't think they're going to come out with a

model that is just leaps and bounds better than anything else we've seen on the market. And let's even assume they do. I can't imagine that within 3 or 6 months, Llama 4 or any other open-source model won't be nearly as good, if not as good. And as I mentioned, according to the Wall Street Journal, Apple and Nvidia are in talks to invest in OpenAI's latest round. So here it is: Tech Giants would join a multi-billion dollar funding round alongside Microsoft and Thrive Capital. And if we didn't already know that AI is taking over the world, large language models from both Meta and OpenAI are seeing incredible popularity. From the information, Meta AI assistant wins millions of users in challenge to ChatGPT. In early August, the assistant had 400 million monthly users and 40 million daily users. Now, this is a little bit more nuanced than they just launched something and it had 400 million users because Meta has a ton of products with billions of users using them. So they simply insert Llama into these products.

Instagram, WhatsApp, Facebook, and then all of a sudden, those users are using them. So, I don't think it's truly 400 million users from zero, but it certainly is a great accomplishment nonetheless. And it says right here, Meta AI had 400 million monthly users and 40 million daily users in early August, less than a year after its debut. The numbers reflect Meta's ability to promote the assistant to its more than 3 billion users. That is their distribution network at play. But OpenAI is seeing massive growth as well. According to Axios, OpenAI says ChatGPT usage has doubled since last year. On Thursday, they said that ChatGPT has more than 200 million weekly active users, twice as many as it had last November. I use ChatGPT, but I've actually started using Perplexity more than ChatGPT because Perplexity is essentially ChatGPT with a bunch of additional awesome features. And I know I was late to the Perplexity train, but I'm all over it now. I promise you that 92% of Fortune 500 companies are using its

products that is insane. 92% of Fortune 500 companies and usage of its automated API has doubled since the release of GPT-4.0 mini in July. So reminder, GPT-4.0 mini is the smaller, a little bit less performant, but very efficient and extremely cheap version of GPT-4.0. And with that, people are paying just a couple dollars per million tokens, which is right down the path of what Sam Altman said: intelligence too cheap to meter. And then back to my previous point: if intelligence is too cheap to meter, how are they going to have a business model next? Midjourney is getting into hardware. We just have a couple little tidbits of information that let us know that they are doing this. Midjourney tweeted, We're officially getting into hardware. If you're interested in joining the new team in San Francisco, please email us at hardware@midjourney.com. Now, what could they possibly be building? It's not going to be a phone, or I don't think it would be, because that market is completely saturated. Somebody says here, Me yelling.

at my mid Journey canvas on the wall. So maybe it's some picture frame that can be updated by you just speaking to it. I really don't know. Maybe it's a camera, some like hybrid camera where you take a picture and it stylizes it in some way. I really don't know. So let me know what you think it's going to be in the comments. But they are definitely hiring for Hardware people, and they actually hired the hardware engineering manager from Apple back in December 2023 here. And so Nick on Twitter says, Last year, mid Journey hired the engineer from Apple Vision Pro to be head of hardware. And down below, mid Journey replied, Good catch. So yeah, it seems like they're getting into Hardware. I just have no idea what it's going to be. All right, next. I already touched on this in last week's news live stream. If you didn't check that out, be sure to check that out. Also, I do News live streams every Friday at 10 a.m. Pacific. Be sure to join that. Andrew Karan says, This morning, the US

government announced that OpenAI and Anthropic have signed a formal collaboration with AI Safety Research testing and evaluation. Under the deal, the AIS SI will have access to major new models from both OpenAI and Anthropic prior to release. So, we already knew this. We talked about it last week. GPT-5 or Strawberry QAR, one or more of these, has already been shown to the federal government. It makes me super sad that I haven't seen GPT-40 voice yet, but the government has seen GPT-5 already. But that's besides the point. But the two top closed-source frontier model creators, OpenAI and Anthropic, are both working with the US government now, showing them what they have. And the USAISI is the US Artificial Intelligence Safety Institute at the US Department of Commerce's National Institute of Standards and Technology. Basically, all the acronyms. And they announced agreements to enable a formal collaboration on AI Safety Research, testing, and evaluation with both Anthropic and OpenAI. So, this is certainly one way to do it. I personally

prefer Meta's approach of just open-sourcing everything, putting it out there, and then everybody can see it, and it's going to get much more hardened because many more eyeballs are going to be on every single model weight, every single line of code, and that's the way I would rather do it. But, of course, OpenAI and Anthropic are both closed source, so they're not going to be doing that. Next, we have new models from NA Research introducing a new open data set release, Hermes function calling V1, the data mix that gave Hermes 2 Pro its tool use and structured output capabilities. It is open source, it is based on Llama 3. We have the data sets right here. The data set includes single and multi-turn function calling and structured JSON output data sets, and an updated version of Glaive AI function call, a data set perfect for training LLMs to be better agents. So, this is going to be a great model for agentic frameworks, and you can check out the Hermes function calling.

repo for more information right there, so really great work to n research and from Google. They just rolled out prompt Autos saving and prompt auton naming in Google AI studio. So, little features but cool nonetheless. I want to see Google catch up on quality though, and this post is by Logan katrick, who leads product at Google AI studio, and he was formerly at open AI as developer relations lead. This is enabled by default for all users with Google Drive access turned on, which is where we save prompts, and you can turn this off in the settings bottom left corner. This also persists to run settings right panel and safety settings for each prompt. Good, there's really no reason that I want safety settings on, and in fact, it really kind of kills the experience using Gemini for me. Next, from Magic AI labs, ltm2 mini is our first model with 100 million token context window. I've seen 1 million, I've seen 2 million, both by Gemini, but now we have a 100 million token context window. That's 10 million lines of code or 750 models.

and we can find out more on their blog post. Trained on hashes with Chain of Thought, the LTM architecture gets the following results, and I think this is needle in a haystack test by Greg Camon. So yeah, 100 million tokens, which is just absolutely crazy to think about. This company has raised a total of 465 million, including a recent investment of 320 million from new investors Eric Schmidt, Jane Street, Sequoia, Atlassian, and others, and existing investors Nat Friedman and Daniel Gross, Elad Gil, and Capital G. So it looks like they're based in San Francisco, and here is the needle in the haystack test actually. So we see a stack of books, place a single question in the middle of 100 million tokens, ask a question, and see if it can find it. So really cool, check that out. And next, and what kind of went under the radar, but I think is an incredibly important story, is Brian Armstrong, who is the CEO of Coinbase, the crypto platform company, had a post and said AIs are now paying other AIs with crypto, and

this is a great use case for crypto, and I think crypto is really struggling to find use cases. And if you're super into crypto, I apologize for saying that, but I know I don't really use it in my day-to-day life, nor does anybody I know. But I know there are strong use cases; it just hasn't taken over. But now AI can pay other AI, which is both awesome and scary at the same time. So, according to Brian Armstrong this week at Coinbase Dev, we witnessed our first AI-to-AI crypto transaction. What did one AI buy from another AI? Tokens, not crypto tokens, but AI tokens. Words basically from one LLM to another. They used tokens to buy tokens, which is crazy. They used crypto tokens to buy AI tokens. AI agents cannot get bank accounts, but they can get crypto wallets. They can now use USDC on base to transact with humans, merchants, and other AIs. Those transactions are instant, global, and free. This is an important step to AI getting useful work done. Today, if you give an AI agent a task and

come back in a few days or hours. It can't get useful work done in part. This is a limitation of the technology itself, and products like Devon AI are getting closer to this. But the other reason is AI can't transact and acquire the resources they need. Basically, he's saying if one AI needs something from another, there's not really a good way to do that, and there's not a monetary incentive to do that either. They don't have a credit card to use AWS, GitHub, or Verell. They don't have a payment method to book you the plane ticket or hotel for your upcoming trip. Now, that's not necessarily true if you give your agent your login credentials, your credit card number, obviously in a secure way. Then I think they should be able to pay on your behalf. But this is definitely a million times easier when it's all done through crypto and code, and it's tracked, and it's open source. And yeah, this is much better if you're working on an L M or AI model, and you think you could benefit.

from having a crypto wallet integrated to conduct payments. Try integrating our MPC wallets from Coinbase developer platforms. So, this seems very cool. It definitely seems like a flu under the radar, but I feel like a handful of years from now, this is going to be commonplace, and we're going to have transactions happening all day, every day, between AI that need to get stuff done. And in our last story, apparently Amazon Alexa is going to be powered by Claude AI around October, according to Reuters. If you think about Siri, if you think about Alexa, if you think about Google's voice assistant, none of them are really that good, and we're all just waiting until they are powered with true large language models. And so, that's what we're hearing here. We know Siri soon is going to be powered by their own models. We know already Google's assistant is somewhat powered by Gemini, and now Alexa with Claude. So, I can't wait for my voice assistant to actually be useful and to go out and do useful things for me and

In the perfect world, I actually wanted to go out and do useful things for me 24 hours a day, 7 days a week, without me explicitly having to tell it what to do. And I think actually having payments per our last story is going to be a big factor in accomplishing that vision. So that's it for today. I hope you enjoyed all the news. If you liked this video, please consider giving a like and subscribe, and I'll see you in the next one.

GPT-5 SOON, AI-to-AI Payments Using Crypto, xAI GPU Cluster is Live, 1,000 Agent Simulation

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

GPT-5 SOON, AI-to-AI Payments Using Crypto, xAI GPU Cluster is Live, 1,000 Agent Simulation

Scripts