OpenAI's New Models: Orion and Strawberry Developments

OpenAI's new model, Orion, utilizes the strawberry model for generating synthetic data, raising concerns about potential model collapse.

### Major Developments in AI: OpenAI's New Models, Government Collaborations, and Industry Innovations
* 🚀  OpenAI is developing a new model code-named **Orion**, utilizing the **strawberry model**, which aims to enhance **logic, reasoning, and complex task performance**.
* 🤝  Sam Altman showcased this strawberry model to the federal government, raising concerns about **potential regulatory capture and influence over AI laws**.
* 🧬  The strawberry model is expected to **generate synthetic data for training Orion**, potentially creating a **feedback loop that updates the model continuously with new data**.
* ⚠️  However, there are worries about "**Model Collapse**," where models trained on synthetic data may **degrade in performance over time**, akin to inbreeding.
* 💰  OpenAI is currently seeking to raise funds at a **$1 billion valuation**, with major companies like **Apple, Microsoft, and Nvidia** expressing interest in participating.
* 📈  Anthropic has released details about their system prompts for Claude and made their artifact feature widely available, **enhancing user experience across different plans**.
* 🌐  Google has launched experimental language models and introduced **AI summarization features** in Google Meet, while also rolling out new capabilities in their **Custom Gems** and **image generation tools**.
* 📅  Upcoming events include a panel featuring **Oprah Winfrey, Bill Gates, and Sam Altman** discussing the **future of AI**, along with Apple's iPhone 16 launch expected to highlight **AI advancements**.

there's been an absolute ton of things that happened in the AI world this week, and I don't want to waste your time, so let's just get right into it, starting with the fact that OpenAI showed strawberry to the feds, and they're using it to develop their next model, which is code-named Orion. So, let me quickly break down what this actually means. OpenAI has been working on a new model that was once code-named QAR, and then the code name became strawberry. But this model that they were working on was a model that was supposed to be a lot better at logic and reasoning. Supposedly, it's also a lot better at complex tasks such as math problems that conversational AIs have traditionally struggled with. We've gotten some little teases from Sam Altman. He posted on Twitter, I love summers in the garden, with pictures of strawberries and things like that. Well, now more information is coming to light, and apparently, he showed off this strawberry model to the federal government, so they kind of have a clue

of what's coming. There's also theories that Sam has this of regulatory capture, and if he's working really tightly with the federal government, then a lot of the laws and regulations may potentially go the way that Sam Altman wants them to go. What is this Orion? Apparently, Orion is OpenAI's next flagship large language model that's in development right now. From my understanding of this, what they're doing is they're using this strawberry model to essentially create what's called synthetic data, or data that the AI created, and then using that data to then train this new Orion model. Theoretically, this means instead of going and scraping data off the internet, like from websites and from copyrighted works and things like that, they can create new data with this strawberry model, use that new data, that new text that the strawberry model creates, and then turn around and train it into this new Orion model. Now, it sounds like that this strawberry model is not the fastest model in the world; it takes some time to actually get an output when you give it

an input. However, they're working on a process called distillation, which essentially simplifies and shrinks down the model so that it could give you responses faster, like what you would typically expect from something like Chat GPT. But putting all the pieces together, what it kind of sounds like is going to happen is that maybe this strawberry model isn't necessarily publicly released, but the strawberry model that is really good at complex reasoning and logic and math is kept sort of behind the scenes at OpenAI, but it's used to create new synthetic data to then train the next Foundation model that OpenAI and Chat GPT are going to use. Theoretically, this can create a new sort of loop where the latest model, let's say for instance this Orion model they're talking about, is constantly being updated because the strawberry model is constantly giving it new data to train on. However, when synthetic data is used, there's this theoretical concept called Model Collapse that also has researchers worried. Model Collapse is essentially when the data that these

large language models are being trained on is coming from data that other artificial models created, creating this sort of loop similar to like animal and human inbreeding, where the model could get essentially dumber and dumber over time, creating models that essentially just don't respond as well as they used to. Now, we don't know if the model collapse is going to actually happen or if the models will actually continue to get smarter by using the synthetic data. That's something that's still yet to be seen. But this week, we got a little bit more of a glimpse into what OpenAI has been working on. This strawberry model with better logic, reasoning, and math capabilities is being used to create synthetic data. That synthetic data is then being used to train the new model called Orion, which is likely going to be the model that we might see next in our user interfaces for things like ChatGPT or whatever they call the next version of GPT. Now, this is all speculation based on very little information that has come out, but from the sounds of things, that's

kind of what's happening right now with the latest OpenAI models. And like I mentioned, they are sharing this information and giving early access to peek at what these can do to the government. We can see Sam Altman here tweeted out, We are happy to have reached an agreement with the US AI Safety Institute for pre-release testing of our future models. For many reasons, we think it's important that this happens at the national level. US needs to continue to lead. Hopefully, the people they're showing these models to early are, you know, smart engineer-type people that actually understand what's happening and not just, you know, US senators that are basing their decisions on what they see in the news. OpenAI is not the only company that's agreeing to share their models early with the US government. Anthropic also said that they will share their models with the US government ahead of time so that they can, I guess, get approval before they're launched to make sure that the government is okay with what these models are capable of before they get in the hands of normal

people and business. I've got a few other things I want to talk about regarding the government and regulations around AI, but we'll circle back around to that a little bit later in the video here. But I found this really interesting while we're on the topic of open AI, that open AI is currently trying to raise at a $1 billion valuation right now. Now, to me, that number is not what's crazy. To me, what's sort of really interesting about this is who is involved in the raise. Apple, Microsoft, and Nvidia have all discussed joining in on this raise. Nvidia, the world's biggest chipmaker, has discussed joining the funding round. Apple and Microsoft have also been in talks about participating in the financing. And to me, that's just fascinating because obviously Apple and Microsoft have been essentially the biggest competitors. I mean, they've been the two top largest companies in the world for a while now, and they're constantly switching places and going head-to-head on what they do. So for them both to be investing in open AI seems

just fascinating to me. All right, moving back to anthropic news. There's been some small updates from anthropic this week, including the fact that they published their system prompts so that you can actually see what sort of instructions anthropic is giving to Claude behind the scenes. If you go to this URL here, docs.anthropic.com, etc., etc. I'll link it up in the description below. You can actually see the system prompts for Cloud 3.5, Sonic Cloud 3, Opus, and Cloud 3 Hau. So, for instance, we can take a peek at the Cloud 3.5 Sonet system prompt. You can see it's pretty detailed. There's a lot of information that it's feeding through to Claude when you give it a prompt. For instance, Claude always responds as if it's completely face blind. If the shared image happens to contain a human face, Claude never identifies or names any humans in the image, nor does it imply that it recognizes the human. There's also this line, If asked about controversial topics.

It tries to provide careful thoughts and clear information. It presents the requested information without explicitly saying that the topic is sensitive, without claiming to be presenting objective facts. Now, there's a lot more to this prompt, but again, I will link it up below. So, if you want to do a deeper dive into the system prompts that Claude is using when you ask it questions, you can kind of get a better idea of how it's going to respond and why it responds that way. In another update to Claude this week, the artifacts are now generally available. Now, if you don't remember what artifacts are, artifacts are this right sidebar here or this right section of CLA when you ask it a prompt, and it writes code or writes a script for you or things like that. It will often open it in this right artifact window over here, and it's also capable of running some code like JavaScript, HTML, things like that. You can actually run that code right here inside of the artifact window. Well, this feature is now widely available to everybody. No

matter what plan you're on, it's now available to users across free, Pro, and team plans. In my opinion, artifacts and projects are the two greatest things that anthropic ever added to Claude that make it so that I use that way more than chat GPT now. Personally, I think everyone should have a website, whether you're a business or you're somebody just trying to have a presence online. It makes sense to have a website where you own everything you put on the website. That's why for this video, I partnered with Hostinger. Not only is Hostinger one of the least expensive options for building a website, but they also have a ton of AI functionality built in. If you head on over to hostinger.com/slmtWolf, you can claim the special deal. And if you choose their Business website builder here, you get access to all sorts of cool AI tools. When you hit the checkout page here, you have a bunch of options to prepay for months in advance, but what's even cooler is if we come on down here to the

bottom click on this. Have a coupon code button and enter the coupon code Matt wolf here. Click apply. We'll watch this total number drop by an extra 10%. Here once you're logged in. They have the option to let AI create your website for you in minutes. Just click Start creating. Pick your website type whether it's a Blog portfolio landing page online store. Their AI is set up for all of them. Let's build up a portfolio for brand name. I'm just going to go Matt's photos. And then for the description. I can literally click this voice button and talk out what my website is going to be about. This is a website where Matt wolf will share all of his photos from his travels and the tech that he likes to purchase. Sure that's good enough. And then I'll click create website. We'll let Hostinger AI do its work here. And in less than a minute I have a full website here where I can start customizing my images. The text on the website and dial it in to look however I want. And it's

already pre-designed to look like a travel website with a portfolio of images ready to go. It looks great on desktop and mobile. I can easily edit my website with their drag-and-drop Builder, and I can design it down to the exact look that I want to go for the site. And it's even got some more cool AI tools like an image generator, a writer page generator, section generator, blog generator, heat Maps, AI SEO assistant, and an AI logo maker, all within the editor. It's pretty dang cool. You can learn more by going to hostinger.com/Wolf and entering the coupon code Matt wolf at checkout. Thank you so much to Hostinger for sponsoring this video. Google had a flurry of announcements this week as well, including the fact that they rolled out three experimental new language models. They rolled out a smaller variant of Gemini 1.5 flash 8B, a stronger Gemini 1.5 pro model, which is better on coding and complex prompts, and a significantly improved Gemini 1.5 flash model. You can try them all over at aist studio.

google.com here and switch between the models by clicking in the top right up here, and you can see that we've got a handful of models that are in preview mode, including the new models that Logan just mentioned over here. These models are intended to gather feedback and to get their latest updates into the hands of developers. So, if there are some errors or mistakes, just know these weren't necessarily meant for wide adoption like ChatGPT was yet. These were meant to sort of test and for developers to play around with. If you use Google Meet for your calls, which is what I tend to use for most of my calls, mostly because it's built into Google Calendar, but it's what I tend to use for most of my calls these days, they now have an AI summarization feature that takes notes on your meetings for you using their Gemini AI. If you're in a Google Meet call, there should be a button up in the top right. If you click on it, you can see that Gemini is taking notes. It shows a summary so far along

with the latest updates at the moment. It's currently limited to the English language and only works on computers or laptops, so no mobile version yet. And when you're done with a meeting, your notes will appear in Google Drive, and the document can be automatically sent to attendees or added to the calendar event after the call. Seems pretty handy to me. Google also rolled out a couple of new features in their Custom Gems and an improved image generation with imagin 3. I still, to this day, don't know if this is supposed to be imagin 3 or imagine 3. I think it's imagin like image, just with the letter N, but I'm not sure, so I'm going to keep calling it imagin. Now, the custom gems are kind of a similar concept, I guess, to Claude's projects and ChatGPT's custom GPTs, where you could give it its own individual system prompt so that it acts in a similar way every single time you chat with it. Some examples they gave: a learning coach, which helps you break down complex topics; a brainstormer gives you easy

inspiration, a career guide, a writing editor, a coding partner, and it says here they are rolling out on desktop and mobile devices to Gemini advanced, Gemini business, and Gemini Enterprise users in more than 150 countries. So, it looks like in order to use these Custom Gems, you do need to be on one of their paid plans. They've also improved image in 3, like I mentioned. Here's some of the example images that they shared from image in 3. Here's a comparison between image in 2 and image in 3. And the other thing that's not noted in this article is that it will now let you create AI-generated people. Again, if you remember last time they let you generate people, they were having issues where it was generating things like black Nazis and Asian Vikings and things like that because it was trying to be ethnically diverse in all of its Generations, but it went a little too far in the wrong direction with that. Well, they claim that problem has been fixed, and now you can actually generate people's faces with imaging.

Again, here's another little cool thing that Google released. I actually don't know when this released, but I just discovered it this week through this article. If you go up to your Google search bar up here and you type at Gemini, you can see a little chat with Gemini dialogue box pops up. And I can say, What are the best taco shops in San Diego? right here in my URL bar, and it will jump me right over to the Gemini web page with some answers to my question. I thought that was pretty handy. It may have existed for a while, but I just learned about it. If you're a fan of Grok from X, it just got a little bit of a boost this week. We can see from this X post here, Grok 2 Mini is now two times faster than it was yesterday. This was posted on August 23rd. One of those things where I record these on Thursdays, they come out on Friday, so Friday news gets missed. This came out last Friday, but it's now 2x faster because some of

The developers rewrote the inference stack from scratch. They also claimed that both models didn't just get faster but also slightly more accurate. So, if you're one of the premium members on X and you want to use grock here, it should be a little bit faster this week than it was last week. We've talked about this in previous videos where Elon is building a brand new data center. I guess the data center is called Cortex. They're building it over at Tesla Headquarters in Austin, and he gave a little 20-second tour of the facility here. So, let's take a peek here. This entire thing is designed for training new larger AI models, and I'm assuming for inference of actually using these models as well. So, pretty insane how big of a system they're setting up here. All right, I mentioned earlier in the video that there was some more AI regulation news earlier this week. Elon came out in support of the SB 1047 bill. That's the bill that I sort of broke down in last week's video. We also did a next wave.

podcast with Anan over at a16z about the entire bill as well. Basically, quick recap the bill basically states, and this is what most of the AI companies are worried about, that if people use an AI model to create something that turns out to be disastrous, the AI companies that made the model are also at fault in addition to the person who used the model for the bad stuff that they used it for. That's what the AI companies are sort of freaking out about. It is a California bill, but most of the AI companies are based in California, so it's going to have an impact on pretty much the AI industry as a whole here in the United States. Elon came out in support of it at the end of last week, and Ric came out and said the benefits of it probably outweigh the cons. They didn't necessarily say they endorsed it or supported it. The wording of this bill has gotten lighter and lighter and lighter with each sort of iteration of the bill, and now it looks like the bill is likely about to pass. It just made it

through the California legislature, we can see here the bill also gives the State Attorney General the power to sue if developers are not compliant, particularly in the event of an ongoing threat such as the AI taking over government systems like the power grid. Over time, the language of the bill has gotten sort of weaker and weaker each time it's sort of gotten another pass through Congress. I think now the latest version of it actually removes even more liability from the companies that made the model. Now personally, I don't think the bill is a super smart bill to be passing right now. I think it's going to slow down the progress of innovation a little bit, but at the same time, I also don't think it's going to be as catastrophic to these AI companies as some people are making it out to be. I think part of the problem with the bill is that the language of the bill, there's a little bit of ambiguity that leaves some things open for interpretation. Things like the companies must implement reasonable administrative, technical, and physical cybersecurity protection to

prevent unauthorized access to misuse of or unsafe post-training modifications of etc. etc. The term reasonable I feel like is always sort of up for interpretation and there's a few things like that where the wording makes these companies uncomfortable but we can see here how many times this bill has been amended. We're still in August and it's been amended twice in August alone where they keep on making the terms lighter and lighter. They also Define critical harm here which is what this bill is trying to prevent and what they're saying that the model Creator could be liable for but if we look at how they Define critical harm the creation or use of chemical biological radiological or nuclear weapon that results in mass casualty or at least $500 million of damage resulting from cyber attacks on critical infrastructure by a model conducting or providing precise instructions or conducting a Cyber attack or series of cyber attacks on critical infrastructure or $500 million of damage resulting from an artificial intelligence model engaging in contact that does the following acts with limited human oversight intervention or

supervision results in death, great bodily injury, property damage, or property loss, and would, if committed by a human, constitute a crime CRI specified in the penal code, etc., etc., and other harms to Public Safety and Security that are of comparable severity. So, I feel like if an AI is able to create chemical, biological, radiological, or nuclear weapons, for instance, maybe we shouldn't be as concerned about the company who initially created the model and more of like the larger implications on the world from that. But at the end of the day, this is one of those things where I can see both sides of the token here. I can understand why the model creators are worried. Companies that are making open-source models and making them publicly available, they don't know how these models are going to be used, 'cause people, since it's open-source, can go and iterate off of them and change them and tweak them and build off of them. And once the open-source model is out into the world, it's sort of out of the hands.

of the initial creator of it. Now, there's probably some flaws in my logic, and I'm sure you'll let me know in the comments if there are, but that's sort of how I feel about this in the grand scheme of things. I don't really think this is going to have as big of an impact as a lot of the AI companies are worried about, but I can also see why the companies are worried. There is another California bill that it seems like most of the AI companies are on board with, and that's that content created with AI should require a watermark, which personally I've got no problem with. It's probably like a watermark in the metadata. I don't really think it's going to solve a lot of problems. I think whatever way they use to watermark, people will figure out ways around the watermarking, but I see no problem with making this a bill and getting the AI companies to get on board with this bill. I just don't really think it's going to prevent the bad actors from doing what

They do another big talk of the week was the Nvidia quarterly earnings call that happened this week, which was pretty interesting. Goldman Sachs actually called Nvidia's stock the most important on planet Earth. And when they announced their earnings, they announced 122% growth in second-quarter revenue earnings. However, after these announcements, the stock still fell by 7%. So, despite the fact that they had huge growth, still, people were still unsatisfied because I guess they wanted even bigger earnings. Now, their profit margins dropped a little bit from 78.4% down to 75.1%, and there's been some issues with delays on shipping some of their latest chips. But it seems to me like Nvidia is still in a really strong position. Not financial advice at all, I don't know what I'm talking about, but they still seem to be growing. However, Nvidia does seem to have some new competition popping up in the form of Cerebrus. They apparently just launched the world's fastest AI inference previously Grok.

held that record. We can see on this chart here that Grock was the fastest inference before, and now this new Cerebrus is even faster. I'm not going to get too into the weeds with this stuff, but it looks like these guys are probably going to have a very similar model to Grock, where they're not actually out there selling the hardware, but they're allowing people to use their hardware to train and run inference. And this appears to now be the fastest way, and apparently the least expensive way, and there might even be another competitor in hardware. I don't actually know what this means yet, but I'm sharing it with you because it's out there in the world, and it's fascinating. Mid Journey claims they're getting into hardware. What kind of hardware? Who knows? We could see here Mid Journey tweeted, We're officially getting into hardware. If you're interested in joining the new team in San Francisco, please email blah blah blah. And then somebody said, Care to share more? And then Mid Journey said, We aren't announcing anything.

specific yet but we have multiple efforts in flight. So apparently, mid-journey is also getting into hardware. The kind of hardware they're getting into, I guess we got to wait and see. Suale from Playground AI just announced a new model of Playground that is really good at designing graphics like a pro. He says you can make t-shirts, logos, social media posts, and even invent new kinds of memes. It's powered by their next image foundation model, Playground V3, which is in beta right now. And from his examples here, they're looking pretty good. We can see some other images that he generated here, like this Radiohead ramen, Ben there, the sun like a tarot card kind of thing. It looks like it's generating some pretty good images. So if you're looking to check out even more AI image generators, maybe Playground's another one you can check out. Leonardo got a little update this week as well, Leonardo Phoenix, which is one of my favorite models, not because I'm biased because I'm an adviser there, but because I genuinely like their models. But that

first part's true. Also, they just rolled out their Phoenix model into the API, so now you can actually build apps yourself that leverage the Leonardo Phoenix model for the image generation. Moving on to open-source AI news, Meta's Llama large language model is approaching 350 million downloads to date. They claim it was downloaded more than 20 million times in the last month alone, making it the largest open-source model, which is not really that surprising. There's also a new model called Quinn 2 VL, which is also an open-source model. But here's what's interesting about this model: it's got state-of-the-art understanding of images of various resolution and ratio, which a lot of the models can do at this point, but it also has understanding of videos of 20 minutes and more. So, it can understand videos over 20 minutes and answer questions based on the dialogue and the content of the video. That, to me, seems pretty cool. Some examples on the page of what this model is capable of output: the color and number of each box.

we can see top row single block blue with the number nine, and it nails it. It solves algorithmic problems, but of course, the thing that impresses me the most is the video understanding. They can upload a video, and then it gives a description of what the video is about right here. And then the user asks, What is the color of the clothes the two astronauts are wearing? The two astronauts are wearing blue and black clothes, zooming to where it shows the two astronauts. I'd say they pretty much got that right. Here's something else that's pretty exciting in the world of large language models. This company, Magic, has figured out a way to have a 100 million token context window. That's 75 million words between the amount of text you input and the amount of text you get back. You could input an entire library of books and ask questions about any of the books in that library, and theoretically, it should be able to find the answer. Like, this is insane. It's gotten to the point where context

window is just not going to matter. We're going to get to a point where unlimited context is pretty much a thing, and that's pretty exciting. You can feed it the entire content of the Harry Potter series, The Lord of the Ring series, The Game of Thrones series, pretty much every book series you could think of. You just feed them all into there and ask questions about all of them, and it should theoretically be able to answer them because the context window is so large that it just doesn't even matter anymore. This topic has been going kind of viral this week as well. There's a new AI model that can hallucinate a game of 1993's Doom in real time. I actually made a YouTube short about this, but we can see here that this is actually Doom being generated in real time. So, as the character is moving around, each frame is being generated. So, every single new frame is just being generated in real time. There's no game engine behind this. Pretty insane, like the ammo, the health, all of this stuff is just generated.

pretty mind-blowing, in fact. My buddy, theoretically media, did a cool breakdown of this thing. I did a short on it that's only a minute long. Check out Tim's video as well because he did a pretty deep dive into what this is and what it actually means. All right, so I know I've covered a lot already, and I have quite a few more things I want to share, so I'm just going to kind of Rapid Fire the rest of the news that I found interesting that I thought you should know. Like this article here about how AI spots cancer and viral infections with nanoscale Precision. Basically, they have an AI model now that can find microscopic signs of cancer and detect the cells really, really early. It can look at cells that are 5,000 times smaller than a human hair and detect if they're normal cells or cancer cells for super early detection. Some of the healthcare advancements that AI is enabling is really kind of mind-blowing right now. Another major issue that AI

could potentially help solve is the issue of climate change. This week, researchers have figured out how to produce light-harvesting molecules that are four times more stable than the starting point. Basically, what this means is that they figured out how to make solar panels that are more efficient and last a lot longer. If you're a video editor that likes to edit in Da Vinci Resolve like me, a new version of Da Vinci Resolve just came out this weekend, Da Vinci Resolve 19. Most of the new features in them are AI features. Now, there's over a hundred new features in this new version of Da Vinci Resolve, but some of them include Intel Track AI, which is better at tracking what's on the screen using AI. Fairlight AI audio panning, so if something is moving across your video, it can actually notice that and pan the audio from left to right or right to left. So, if somebody's listening in headphones, they can actually hear the audio pan using an AI feature. And a ton of other cool AI features. Again, over a hundred new features.

were added. I'll link up to the article below so you can learn more about this. It was recently reported that Meta was working on a new headset that was designed to compete with the Apple Vision Pro, a much higher-end, more expensive, more technologically advanced version of The Meta Quest. Well, this week they announced that they're scrapping that product, but they also started talking about how they're working on mixed reality glasses. I think what's going on here is that the Meta Ray Bands have been so successful. I mean, they really are awesome. I love my pair, but they've been so successful that they're realizing that's the form factor people want. I don't think people really want bigger, bulkier, cooler VR headsets. They want things that look like normal sunglasses that they can wear every day that give them some of those features that they might have used VR for, things like heads-up displays. The Meta Ray Bands, they've already got a large language model in them, they've already got cameras in them, they can already

take pictures and tell you about your environment. They've got good speakers. They're already really cool. But imagine that with like a heads-up display that can show map directions or help you remember people's names and put their name right above them as you see them walk in front of you. Really cool things like that that these glasses can enable. Now, it doesn't look like we're going to see them until 2027. There's still some hurdles to overcome. The worry now is that it's going to require a battery pack. Nobody's going to really want to wear Ray-Ban sunglasses with like a cord going to their pocket with a battery pack. That might be something that they need to solve as well. But this, to me, is actually even more exciting than the announcement that they might make a more higher-end version of a Meta Quest. I would much rather see the Meta Ray-Bans get more and more advanced over time. In October, Amazon is going to roll out a new revamped Alexa with generative AI. If you've used Alexa now, it's not the

smartest AI in the world. Future models are going to use more generative AI, like what we get out of things like Chat GPT. Probably going to use one of Amazon's Titan models, if I had to guess. But it should make your Alexa a lot smarter and enable more conversation, similar to what you get out of something like Chat GPT or like what OpenAI demoed with their GPT-4 advanced voice. Maybe now we'll get that right in our Alexa. We'll find out in October. The camera company Wyze is implementing AI inside of your camera that can now search for things that it saw on camera. They have this demo video here that sort of shows off what it can do. Like, you can search cute animal, and it will actually search for video footage that your camera recorded of animals that came onto your video screen. It says you can get even more specific. I typed in woman under umbrella, and you can see it actually found video footage of a woman under an umbrella from their camera. They said you can also be as

vague as you want, so they type in truck and it finds clips of trucks that popped onto the camera. That seems like it'll be pretty handy. Oprah Winfrey is about to do a panel with Bill Gates and Sam Altman all about the future of AI. It's going to be on ABC, and it's going to air September 12th, so mark your calendar. And finally, Apple has announced when the iPhone 16 launch is going to happen. It's going to be on September 9th, and a lot of the speculation is there's going to be a lot more AI talk. The tagline for the event is it's glow time, and you can see their picture of an apple glowing. If you remember when they showed off the newest version of Siri, whenever you talk to Siri, it shows a glowing box around your phone to show that Siri is enabled and listening to you. So, a lot of people are speculating that it's glow time means it's AI time or it's time to show off what Siri is really capable of, but

we'll find out on September 9th, and I'll likely do a breakdown video of any sort of AI that came out of that event. So, make sure you're subscribed to this channel and you like my videos to make sure that that shows up in your feed. And that's what I got you this week. I know it's been a ton of stuff this week. I broke a lot down, but a lot is happening. There's a lot of things going on behind the scenes. Some of the OpenAI news is really, really big. The strawberry training, the new Orion model is huge. That probably deserves its own breakdown video. Again, I mentioned I'm doing a lot of traveling right now, so my video production has slowed down a little bit. But once all my travels are over, you can bet I'm going to make a lot more videos breaking down a lot more details into the intricacies of these various things that I've talked about, as well as make more tutorials on how to do a lot of the cool stuff with AI. I'm also

experimenting with some shorts. So, if you see one of my shorts pop up in your YouTube feed, give it a little heart. I'd really appreciate it. I'm trying to crack that shorts code, and any sort of content that I don't feel deserves an entire video on, and I can sort of break it down in a minute or less, I'll probably make some shorts on that. So, we'll see how that goes as well. Anyway, I'm just rambling at this point. I really, really appreciate you tuning into these videos, nerding out with me around AI. I record these once again on Thursday, so if any big news comes out on Thursday night or on Friday, it typically makes the following week's video. And as always, check out Future Tools. This is the website where I curate cool AI tools and the latest AI news. I've got a really killer newsletter all around AI and the latest advancements, so check that out. Totally free. You don't like it, there's an easy unsubscribe button. So, what do you got to lose? All right, once again, thank you so

much for tuning into this video. Thank you to Hostinger for sponsoring this. I really appreciate everyone who tunes into my channels, even the ones who leave negative comments. I love the fact that you watch the video anyway, and I appreciate you as well. So, thanks again. Bye-bye. See you in the next one.

AI News: OpenAI's Next Model Revealed!

AI, No-Code, Tech, Futurism - I'm a tech nerd and talk about tech nerd stuff

Matt Wolfe

OpenAI's New Model Development: From Strawberry to Orion

OpenAI is using its new model, code-named Strawberry, to create synthetic data for its next flagship model, Orion.

The Strawberry model aims to improve logic, reasoning, and complex task handling, while also potentially influencing regulatory decisions through its collaboration with the federal government.

There are concerns about Model Collapse from using synthetic data, but the process could create a continuous loop of updating and training the Orion model.

OpenAI's $1 Billion Valuation and Anthropic Updates

OpenAI is seeking to raise $1 billion, with investment interest from Apple, Microsoft, and Nvidia.

Anthropic has published its system prompts for Claude, providing insight into the AI's instructions and response strategies.

Additionally, Claude's artifact feature, which allows users to run code and scripts, is now available to all users across free, Pro, and team plans.

Google has rolled out three new experimental language models: Gemini 1.

Google Meet now offers an AI summarization feature for meetings, taking notes and providing summaries, with documents sent to Google Drive afterwards.

Additionally, Google has introduced Custom Gems for personalized interactions and improved image generation with imagin 3, including the ability to create AI-generated people's faces.

Elon Musk's AI Model Updates and Regulation News

The inference stack for AI models has been rewritten, resulting in faster and slightly more accurate performance for premium X members.

Elon Musk is building a new data center, Cortex, at Tesla Headquarters in Austin, to train larger AI models.

The SB 1047 bill in California, supported by Musk, holds AI companies liable for disastrous outcomes from their models, causing concern in the industry.

AI Companies and Their Recent Developments

AI companies are concerned about a California bill requiring watermarks on AI-generated content, but the effectiveness of this measure is questionable.

Nvidia's quarterly earnings showed impressive growth, but the stock still fell due to unmet expectations and shipping delays.

Meanwhile, Cerebrus has launched the world's fastest AI inference, potentially disrupting Nvidia's dominance, and Mid Journey is also entering the hardware space.

Additionally, Playground AI is releasing a new model for graphic design, while Leonardo Phoenix is updating its offerings.

The Leonardo Phoenix model is now available through the API for image generation.

Meta's Llama model has reached 350 million downloads, making it the largest open-source model, while the new Quinn 2 VL model offers state-of-the-art image and video understanding.

Magic's new model features a 100 million token context window, potentially allowing for unlimited context in text-based queries.

Additionally, an AI model is now able to generate a real-time game of Doom from 1993, and another model is using nanoscale precision to detect cancer cells for early diagnosis.

Researchers have developed light-harvesting molecules that are four times more stable, leading to more efficient and longer-lasting solar panels.

Da Vinci Resolve 19 has been released with over a hundred new AI features, including improved screen tracking and audio panning.

Meta is shifting its focus from a high-end headset to mixed reality glasses, inspired by the success of the Meta Ray Bands, while Amazon is set to revamp Alexa with generative AI in October.

A new AI tool will help users find truck clips by typing in "truck.

" Oprah Winfrey will host a panel with Bill Gates and Sam Altman on the future of AI, airing on ABC on September 12th.

Apple's iPhone 16 launch is set for September 9th, with speculation on a focus on AI and Siri's capabilities.