AI News: OpenAI Finally Released Their New Model!

I just spent the last week at Disneyland. And of course the week that I'm gone turns out to be an insane week with tons of big announcements. I'm a day later than normal on getting this AI news video out, so I'm not going to waste your time. Let's just jump right in. There was really two major, major things that happened this week and then a whole bunch of little things. The two major things the new release from OpenAI and the Apple iPhone event that happened. So let's start by talking about this new OpenAI01 preview mod that's been made available. We've been getting lots of teases about it over the last several months from OpenAI. Originally it was called Q Star, then it was called Strawberry, and now they're calling it OpenAI01. It kind of sounds to me like all future models are going to follow this naming scheme. We're probably not going to be getting, you know, GPT5, GPT6, GPT7, et cetera. This is 01 and in their blog post about it, they said we are resetting the counter back to one and naming this series OpenAI01. So I'm guessing their next models that they show off are going to be OpenAI 02, OpenAI 03, etc. And maybe, you know, some decimal points in there as well. Now, in order to use these new models, you do have to be on a pro plan or an enterprise plan. You basically have to be paying for chat GPT, but when you log into your account, every paid member should now have these new options. It still defaults to ChatGPT 4.0, but if we click on this little dropdown, you can see we now have 01 preview and 01 mini with their older models falling under this more models additional dropdown. Now, for the most part, OpenAI still recommends you use GPT4O for most things. If you're going to use the 01 models, it's more for advanced reasoning, mathematics, logic and things that need to be really thought out. More complicated tasks. Now, if you're wondering what makes this model different or better than previous models, well, it essentially thinks through its response before responding. So when you ask it a question, it's going to be a lot slower to respond, but it's going to really think through its response. This is called chain of thought prompting. So If I select 01 Preview here, you can see some of the suggested prompts and one of them is how many Rs are in Strawberry, which you know, the previous model of GPT4 kept on telling you there was two Rs in Strawberry. Let's go ahead and just use one of the demo prompts. Let's tell it to create a puzzle to solve for me. And here you can see it says generate a six by nanogram puzzle for me to solve where the solved grid looks like the letter Q. And you can see it's actually thinking. And there's a little dropdown here. And if I open this dropdown you can actually see its thought process. Creating the puzzle. I'm working on a 6x6 nanogram that forms the letter Q. Sketching the grid layout, formulating an O shape, filling the grid, examining cell patterns, assessing cell patterns, creating the puzzle. You can see the thought process of ChatGPT thinking through all of the steps that it needs to complete this. And then we can see down here when it's finally done, it says here's a 6 by 6 nano gram puzzle for you to solve. If I scroll up to the top of this prompt here, we can see that it thought about it for 30 seconds. So it spent 30 seconds going through that sort of chain of thought logic, thinking through everything it needed to do. One of the other example prompts, solve an advanced math problem. You can see it wrote up this complicated math problem that I don't even know how to read. Let's $sign/math CAL B $sign B. The set of rectangular boxes with surface area. I don't even know what this means. Basically what it's going to do is it's going to try to respond and then sort of double check itself as it goes, finally outputting what it thinks is the best possible response we can see here. It's still thinking it through and it's still working out the math problem. As I'm talking, it came to an answer of 721, is that right? I have no idea, but it took 33 seconds to think it all through. The press release over on the OpenAI blog has all so of examples of how they've been using it. OpenAI01 and Economics in Cognition. That's the company that makes Devin quantum physics, genetics, etc. Now, I did mention that there was the O1 preview and the 01 Mini. Well, it looks like the 01 Mini will become available to ChatGPT free users eventually, but right now both models are only for paid members. Now, they've done a lot of testing on IT and apparently 01 ranks in the 89th percentile on competitive programming questions, places among the top 500 students in the US in qualifier for the USA Math Olympiad and exceeds human PhD level accuracy on a benchmark of physics, biology and chemistry problems. So using that chain of thought reasoning, it's definitely gotten a lot better at solving technical problems. Now, in all of these various benchmarks that they show off on their website, it looks like there's another oh1 model that we don't have access to yet. As we can see in these various charts here, we have an O1 model and we also have the O1 preview model. The 01 preview model is what we're seeing in ChatGPT. So in competition math, it's got 83.3% accuracy, but the version that we seem to be using has about a 56.7% accuracy competition code. Their main 01 model is in the 89th percentile for competition code, while the one that we have access to is in the 62nd percentile. And then under PhD level science questions, you can see that 01 scores a 78, where the 01 preview, the one we do have, actually scored a little bit better at 78.3. Compare that to an expert human, which is about 69.7% accurate. We can see some other benchmarks here with GPT4O in pink and 01 improvements in this blue color, the blue being how far it exceeds the previous GPT4O model. So, for example, in math, GPT4O got a 60.3, while 01 got a 94.8. So about a 34% improvement. Their website has a whole bunch of examples of how it was used. There's been a ton of YouTube videos already put out where people are demoing this and walking you through it. But pretty much the Zero1 preview model outperformed GPT4O in cipher coding, math, crossword, English, science, safety and health science. So for each of these up here, you can see the initial prompt as well as the response. And in pretty much every single One of these, O1 outperformed GPT4O. Now, again, they also released OpenAI 01 mini, which is actually 80% cheaper than OpenAI 01 Preview and it's optimized for STEM reasoning. Now, one of the big complaints that people have had about this new OpenAI 01 model has been the pricing. If you want to use it through the API for one of your software products, it's quite a bit more expensive than what's already out there. It's also quite a bit slower because you saw how it thinks through the process. So this OpenAI01 mini model is supposed to offset some of that, bring the cost down. And it's quite a bit faster than the preview model we can see here. The 01 mini model took nine seconds. The 01 preview model, as of me talking right now, is already at 30 seconds here in this comparison. But I really like this tweet from Jim Fan here where he breaks it down a little bit more clearly about what's happening. And I especially like this graphic that he shared right here where most large language models spend most of the time and money on pre training. So they get a wide amount of data from all over the web. They pretty much scrape everything pre train that into their model. Then they do what's called post training, which is a little bit more like fine tuning and putting up guardrails and telling the model how to respond. And then inference is a teeny tiny sliver of what happens when you use these AI models. Inference being when you give it a prompt and it gives you a response where GPT4O is really fast at inference. You give it a question, it responds within three seconds. It appears they're spending less time on the pre training. The scraping every possible thing it can possibly get, spending roughly the same amount of time on post training, sort of dialing it in, fine tuning it's and then inference, where you actually give it the prompt and receive your response back, is where a lot more time is being spent. So theoretically we should be able to get new, more improved models being released even quicker because less time is being spent on pre training and more of that time might be shifting to the inference phase when you actually prompt it. Companies like OpenAI are actively slowing down the inference phase so that they can minimize the most expensive part, which is the pre training phase. Now this is a super oversimplification. I will link up to all of the articles, all of the posts, all of the tweets I've mentioned in this video below so you can dive a little bit deeper. But I do want to address one thing here that my buddy David Shapiro says. He mentioned that Claude Sonnet 3.5 can do strawberry with the right prompting. There's no secret sauce. We can synthesize the data with any model. Basically the point that he's making here is that chain of thought prompting has been around for a while. You can use chain of thought prompting in any large language model to basically tell it to think through step by step and give it additional prompts to get to the right conclusion. All OpenAI is doing is putting all of that stuff in sort of like the system prompt and telling it to think it through step by step and basically look at its own response, evaluate its response, update its response based on its own evaluation, and continue to do that. In the past, you would just do that through additional prompting. Now OpenAI is doing that for you immediately after the prompt. But that's the biggest news of the week. That's the news that has the AI world buzzing the most. The other thing that happened this week was Apple had their Glow up event. This event was pretty much designed to update people on the latest iPhone, the latest Apple Watch, as well as some updates around the latest AirPod innovation and things like that. Most of the AI features that they talked about during the Apple Keynote we actually already got at wwdc. There wasn't a ton of new AI features that we didn't already see at that. We just got new info that it's going to be rolling out into the iPhone 16. Apple themselves even put out a blog post here solely focused on Apple Intelligence and what's going into their new devices related to Apple Intelligence, including features like being able to rewrite and proofread and summarize emails and documents, the ability to clean up photos by removing things in the background, the ability to prioritize your notifications, the new glowing box around the phone for when you're using Siri, their new image playground which lets you generate AI art and even generate directly inside of notes. And again, for the most part, this is all stuff they previewed at wwdc. This wasn't new exciting announcements during this specific Apple event. There was, however, a few interesting things from the Apple event that I want to highlight, so I'm just going to jump to those parts of the keynote here, including the fact that Apple watches are now going to have AI translation built into them and the Translate app comes to Apple Watch using machine learning for speech recognition and rapid translation. The thing that I found interesting about the new AirPods was the ability to nod or shake your head to respond to Siri.

When interacting with Siri, you can simply nod your head yes or shake your head no.

In response to Siri announcements, they talked about their private Cloud Compute, which is essentially cloud compute that you can use, but it's private, so they're not storing or keeping or training on any of your data. It's also going to allow you to use larger models that that won't actually work on a mobile phone by sending them to a cloud GPU to run the processing. They also once again showed off some of the features to rewrite things like emails for you, creating your own images and emojis with text to image generation, summarizing your notifications for you, as well as prioritizing what it thinks are the most important notifications to the top, as well as the new visual intelligence which doesn't sound like we're going to get right away. This is coming, it sounds like March of next year 2025, but this is a feature where you can take a picture of something and what you take a picture of it will actually then give you information around it. So this person took a picture of a restaurant and then it gave a bunch of information around what time it closes, the cost, the reviews, et cetera. It's also getting some updates to the photo editing which we looked at earlier and again, pretty much everything they announced was stuff that they announced during wwdc. They just showed it off in the context of the new iPhone 16. The Verge put out this article afterwards called the iPhone 16 will ship as a work in progress, basically saying that when you buy an iPhone 16, if you bought it this month in September or early October, you're not going to have any Apple Intelligence features. Supposedly the features are going to start to roll out in iOS 18.1 in October sometime, with more and more of these AI features rolling out over the coming months. And we also know that a sort of visual intelligence feature isn't going to be coming until around March of next year. Even if you went out and rushed out to buy it. So you could be the first one to use these AI features. You're not going to have them on day one, unfortunately. Adobe shared some interesting information this week as well with their new text to video generation version of Firefly. Now from the previews I've seen, this looks like it could actually compete with Sora, and they also claim that it's all ethically sourced video, so it's only trained on openly licensed public domain and Adobe CD stock content. They're calling it commercially safe. Here's a thread from Pierrick showing some examples of some of the videos that have come out of this new Adobe Firefly video model. It looks like they generate about five second videos. Here's one of a galaxy that zooms out to reveal an eyeball. We have one of a detailed portrait of a reindeer, slow motion, fiery volcanic landscape, miniature adorable monsters made out of wool and felt footage of a camera on a drone flying over a desert with wind blowing over the dunes creating waves in the sand below. Detailed, extremely macro close up view of a white dandelion viewed through A large red magnifying glass drone shot going between the trees of a snowy forest at sunset golden hour stop motion 2D animation made of felt of an egg cooking in a frying pan. Now this one looks really cool to me. To make this style of animation I could really see that working well for like shorts and like little explainer type videos. Hand drawn simple line art of a young kid looking up into space with a wondrous expression on his face. An adorable kawaii cheeseball on the moon, smiling 3D render, octane etc. Macro detailed shot of water splashing and freezing to spell the word ice so it looks like it actually spells words inside the video as well. Now we don't have access to this yet, but it looks pretty promising. Of course these videos are likely all some of the cherry picked videos. They probably took multiple prompts, picked the best one and shared those. But this is what Adobe has shared with us so far of what the Adobe Firefly video model can do. But a lot of smaller sort of interesting things happened, so let's go ahead and rapid fire them. I still have a lot of tabs up here I want to share, but I'm going to try to get through them quickly and just give you the quick recap of what's going on with a lot of this stuff. Starting with the fact that Mistral released Pixtral 12B. Now Mistral has both open source and closed source large language models available, but this new Pixtral model is their first model that can accept images as an input. So something we've been able to do with most of the models now we can now do with Mistral's 12B model. And the best part is this is an open source model so developers can build on it, iterate off of it, fine tune it, do whatever they want to it to improve it, and make it an even better model for people to use. Now this is really really cool. Now Google has this tool called NotebookLM and it's really actually super helpful. You can upload a whole bunch of documents into it and then have conversations about those documents. So for example, here's a notebook about the invention of the light bulb and you can see a bunch of different articles have been uploaded about the invention of the light bulb and then we can actually chat with all of these sources so it will look across all of these sources with every question we ask and respond based on the information within the sources that are available. That in itself is really cool and really helpful. Kind of similar to what we could do with Claude projects where we upload a bunch of sources there and have conversations about it. However, they just rolled out a brand new feature that basically generates a podcast about your notes. If I was to come down here and click on Notebook Guide, you can see we have this new audio overview button here. If I click load, it'll say this may take a few minutes and it will create a podcast of two people discussing back and forth with each other. The invention of the light bulb. So here's what that sounds like.

Light bulbs, right? They're so normal now. Like we just kind of expect them to be there so we can actually see what we're doing when it gets dark.

It really is funny when you think about it. Like flick a switch, bam. Instant sunshine. No matter what time it is now.

This is one of the examples that Google loaded in for us. I was curious what would happen if I loaded in a complex research paper. So I actually took a document from archive.org about Lin Fusion, a new architecture for text to image generation that uses a linear attention mechanism to address the computational limitations of traditional diffusion models. I was wondering if I just take this complex PDF with all of this technical jargon in it and I throw it into NotebookLM and then create an audio overview, would I actually be able to understand what the hell the paper was talking about when I was done? And the answer is, yeah, it actually did a pretty dang good job.

Ever get that creative itch? You that feeling when you're absolutely bursting with ideas for some awesome, crazy detailed.

AI art, but then you remember what it's actually like waiting forever for images to render your poor computer sounding like it's about to explode?

Yeah, can really kill the vibe, you know.

Now I'm not going to play the whole thing, but it has that podcast style where they're having a conversation back and forth about the contents of the document that I just uploaded. It's a super cool feature. I highly recommend people play around with it. It's available over at notebooklm.google.com while we're on the topic of audio, Amazon is allowing audible narrators to clone themselves with AI. Amazon will begin inviting a small group of audible narrators to train AI generated voice clones of themselves this week with the aim of speeding up audiobook production for the platform. Now it's a US Only beta test and will be extended to rights holders like authors, agents and publishers later this year. Narrators can also use Amazon's production tools to edit the pronunciation and pacing of their AI voice replica. Amazon says that narrators will be compensated via a royalty share model on a title by title basis. But didn't go into more details than that. SUNO rolled out a cool new feature this week called Covers. They put out this X post here. Reimagine the music you love with Covers. Covers can transform anything from a simple voice recording to a fully produced track and into an entirely new style, all while preserving the original melody that's uniquely yours. I initially learned about this from Nick St. Pierre, aka Nick floats over on X where he showed off what it can do, but I wanted to jump in and play around with it myself and so here's what I did. Now don't judge. I can't sing at all, but I recorded this little audio clip here. Subscribe to Matt Wolf on YouTube and don't forget to check out future tools IO. So that was my brilliant singing. Basically. In order to do that, I went into Suno, clicked on Create, clicked on Upload Audio, click the audio button and then it allowed me to record my voice. Once I recorded my voice, it put that little audio clip of my voice right here in line with the rest of the songs that I've generated on suno. If I jump over to these three dots over here on the right side and then come down to create, we can see we now have Cover Song Beta as one of the options. When you click on that, it will create a couple cover songs of using the audio that you just put in. Now it's not going to use your voice, but it will use the same words and try to match the melody. So here's what it made for me. Subscribe to Mat Wolf on YouTube and don't forget to check out FutureTools IO. Subscribe to Mat Wolf on YouTube and don't forget to check out FutureTools IO. There you go. Pretty cool. Sounds a lot better than me trying to sing it. Honestly. Now I should note this is only available to paid members of SUNO right now and there is a limit to the amount of covers you can make. You can see I have 198 free cover songs remaining. There is a finite number per month and it is only available on one of the paid plans. Moving on over to Facebook News, Facebook and Instagram are making AI labels less prominent on AI edited content. A lot of people were all up in arms because when you post an image to Facebook or Instagram it would have a little note saying generated with AI or something like that. And a lot of people were saying my image wasn't generated with AI. Why is it putting that on there. Now they're making it a little less prominent. You actually have to click into a menu to find the AI info. Ideally, less people will be frustrated by the sort of tagging of AI content when it's not really AI content. Facebook also admitted this week that they are scraping pretty much everybody's photos and posts to train their AI, and there is no opt out option. This was from some sort of Facebook hearing in Australia with Meta's Global Privacy Director Melinda Claybaugh.

The truth of the matter is that unless you consciously had set those posts to private since 2007, Meta has just decided that you will scrape all of the photos and all of the text from every public post on Instagram or Facebook that Australians have shared since 2007 unless there was a conscious decision to set them on private. But that's actually the reality, isn't it?

Correct we use data that's thanks. Thanks for the answer. Now I'm sure Facebook and Meta aren't going to get in too much trouble for this, because I'm sure it's buried in the terms and conditions somewhere that you're giving them the right to use and train on your data when you upload it to Facebook. Unless you explicitly set it to private. If you signed up for one of these platforms and you're using it, you probably unknowingly already agreed to letting them do this. I'm just guessing, though I haven't read the policies myself. There's some really cool AI generative game type stuff coming up. Like for example, Roblox announced late last week that you're going to be able to use AI to create sort of 3D worlds inside of Roblox. Roblox announced that it's working on a 3D foundational model that will power generative creation on its platform. The model will be open source and multimodal and allow creators to generate 3D content using text, video and prompts. A creator can say I want to create a world that is in the Scottish Highlands with castles and a stormy day with a Dragon background, and I want all of this in a Steampunk Steam style. And then the output will be the full scene creation. They did say Roblox isn't trying to replace the creative process, but instead focuses on enabling more people to develop and create games. Here's a little screenshot. Not a whole heck of a lot to go off of, but you can see a little before where they're in like green grass in a road and then after there's a little bit more texture And a little bit more scenery going on on the road. That's all we're working with right now. This one looks even cooler. I think it's called Cybever. They just unveiled their 3D world creation platform. Now, this isn't something that we have access to yet, but here's what this looks like. You can generate a map through text. It creates a real basic map, but then you can adjust it by drawing. You can see they're drawing like a little river here into the map. They can adjust the terrain, adjust the world style. You can see they have some like, templates you can use, like a water village, industry zone, grand bazaar. And it'll create some generated town layouts. And it'll create a 3D preview in less than a minute with some additional assets to kind of give you an idea of what that world's going to look like. And then this was the output that it shows. Now, this is one that, to me, feels like I'll believe it when I see it. This looks like too good to be real, but you can see this is the 3D environment they created here. Load marketplaces or your own assets. There's a deer, there's the ocean, mountains in the background. And it looks like you can pay to add assets into the game as well. To me, it looks so good that I have a hard time actually seeing this work as well as it says it does in the real world until I actually put my hands on it myself. While we're on the topic of game assets, this company DAZ3D, showed off this new plugin which allows you to generate character mesh shapes from text prompts. So if we take a look at their video here, young female African warrior, and it generates the character there, muscular dwarf with a big beer belly and a big nose, large head, etc. Etc. And it made that character a pregnant woman. It made that character male, tall, lean, pale, vampire, tall, lanky, alien male. So we can see you can plug in whatever you want your character to look like and it will create various models that you can use, and it gives you a nice start for your game assets. I'm unclear if it actually does the texturing for you because it goes on to show this image where she's got tattoos and pants and a shirt and a gun and everything's colored in. I don't know if it does that part for you. I'm guessing it does because they're showing it in the preview, but I'm not super clear on it yet. But that's from a company called DAZ3D in collaboration with Yellow3D. Another tool that can help with game assets is a tool called Meshi. Well, they just announced Meshi version 4 where you can enter any text prompt and it will generate these 3D objects from the text prompt. This you can actually use for free right now. You get a certain amount of credits to use it for free. If you head on over to Meshi AI, you've got text to 3D, image to 3D, AI texturing and text to Voxel. I tested this a little bit. I did an image to 3D. I uploaded an image of my head here and I did it with quad and triangle topology to see which one looks better. Here's the quad topology. This is what it made me look like. I mean, it's got the beard and hair color right? I guess. Here's what it did with the triangle topology. And once again, I mean, I've got a beard. Got that part right. So far, not super realistic when you upload real face images. If we head over to text to 3D, I've played with this a little bit more. I just generated this one today with their new Meshy 4 where I gave it the prompt a wolf howling at the moon. And this is what it generated. And it's actually pretty impressive. When you zoom in and look pretty close, you can see that the wolf doesn't have eyes and it's got a really long snout, which seems a little off. But if you ignore the fact that it has no eyes and a really long snout, from this angle, these angles, it all looks really actually pretty dang solid. I'm really impressed with the, like automated texturing that it did. It just could use some work on getting the face. Now. It did give me some other options to choose between, and the one that I selected was the best you can see. This one looks more like some sort of weird beast than a wolf. This one looks like a wolf or something with a giant tumor on its head. I don't know what's going on with some of these. This one came out pretty decent looking. I used the texture feature and that's what I got. I think it looks pretty good. Slight AI news with the new PS5. There's a new PS5 Pro coming out. It's going to use AI to upscale the video quality and make it look better. I personally think this new PS5 Pro is a total joke. They're releasing this at what, 700 bucks or something like that? And it doesn't even come with a disc reader. So if you have a PS5 or a PS4 or already own some PlayStation discs that you'd want to play in it. Well, you're not going to be able to play it without buying an external disk drive because it doesn't come with it with the PS5. That sucks. But it is going to use AI to upscale the quality and try to get even better, better video quality out of the games that you're playing. And finally, some new information out of DeepMind's robotics lab. They have a robot now that can actually tie a shoe. So we can see in this video the robot grabs both the strings and is able to accurately tie the shoe. Robots have not been able to do that before. Here's another one where it picks up a shirt off a table and actually puts it on a coat hanger and manages to to hang it up. And another one where it's sort of repairing and putting some pieces on another robot. Pretty cool that we're seeing this kind of dexterity with a two handed robot where it's going to get better and better at being able to do everyday tasks for us. That's ultimately where we want these robots to get to anyway, right? We want them to be able to do our dishes and do our laundry and I don't know if I need a robot to tie my shoes for me, but we want it to do day to day tasks around our house for us. And it needs this additional dexterity like being able to tie a shoe to be able to accurately do that kind of stuff. And there you have it. There's a breakdown of all the AI news that I caught this week. Again, I was at Disneyland pretty much the entire week, got home, did a super cram session to try to catch myself up. That's why this video is coming out a day later than normal. I normally put these videos out on Fridays, this one's going out on Saturday. But I needed that extra time to catch up on all the news, make my notes, figure out what I felt was worth showing you and put this video together. So I apologize for the day late in publishing this video. But if you do want to keep up with the news on a pretty daily basis, check out Feature Tools IO. There's an AI news section where I keep that up to date regularly. All the news that didn't make this video is on that page. And all the cool AI tools that I come across I share here on Future Tools homepage. There's a free newsletter. Check it out. FutureTools IO. You're gonna like it. It's really cool. I'm not biased at all. It's just the best website on the entire Internet. If you like videos like this, you wanna stay looped in on the latest AI news, the latest AI tools, the latest AI research, and get tutorials on how to actually use some of this stuff in a meaningful, helpful way in your daily life. Make sure you like this video and subscribe to this channel, and I'll make sure more videos like this show up in your YouTube feed. Thank you so much for tuning in and nerding out with me today. I really appreciate you. I'll see you in the next video.

Bye.

Bye.