AI NEWS: Self-learning AI, Crazy AI Ads, and More ChatGPT Updates

0:00 It's been another really interesting week in the world of AI. We got new models that train themselves. AI infiltrating every part of advertising, including what AI will look like in the apocalypse, plus some new chat GPT features and probably the best dance moves you'll ever see coming from a humanoid robot. So, don't want to waste your time. Let's get right into it. Starting with Alpha Evolve out of Google DeepMind. This is a self-improving AI that doesn't just write code, it actually evolves the code. So, Alpha Evolve actually uses two of Google's models. It uses Gemini Flash and Gemini Pro. So, it uses Gemini Flash to maximize the breadth of ideas explored while using Gemini Pro to provide critical depth with insightful suggestions. So, imagine it like this. When you're trying to think of like a novel solution for a problem you're

0:46 trying to overcome, you might brainstorm a whole bunch of ideas and just sort of dump them all out. Some of the ideas will be good, some of them will be really bad, and most of them will probably be somewhere in the middle. Well, that's what Gemini Flash is doing. It's trying to just brain dump a whole bunch of potential ideas to solve this problem. And then Gemini Pro steps in to better evaluate the ideas and figure out which ones are actually potential solutions to the problem they're trying to solve. Alpha Evolve verifies, runs, and scores the proposed programs using automated evaluation metrics. These metrics provide an objective, quantifiable assessment of each solution's accuracy and quality. Google DeepMind has actually been using Alpha Evolve internally to help optimize things like their data center, their TPU circuit design, and even to train the

1:33 Gemini models. Alpha Evolve is also able to propose new approaches to complex mathematical problems. It was actually able to discover multiple new algorithms for matrix multiplication. It found an algorithm to multiply 4x4 complex valued matrices. And the previous best known algorithm for this specific mathematical problem was created back in 1969. So it found a better way to solve a math problem where no better way has been found in 56 years. So what makes this alpha evolve so particularly special is that is actually inventing new methods of solving problems as opposed to just using methods that were already known and sort of scraped from the internet. This to me feels like a pretty big breakthrough. And they

2:19 actually announced this the week before Google IO. And if something this big was announced before Google IO, it's really got my wheels turning on what they're actually going to announce at Google IO next week. Along similar lines, a new research paper came out this week from Singa University, Beijing Institute for General Artificial Intelligence, and Penn State called Absolute Zero: Reinforce Self-Play Reasoning with Zero Data. This is actually a new method of training AI models without needing to give it data. If we read the abstract here, in a hypothetical future where AI surpasses human intelligence, tasks provided by humans may offer limited learning potential for super intelligent systems. So they're basically saying once the AI is super intelligent, humans

3:05 are going to be too dumb to further train it. And what do we do then? To address these concerns, we propose a new RLVR paradigm called absolute zero, in which a single model learns to propose tasks that maximize its own learning progress and improves reasoning by solving them without relying on any external data. Under this paradigm, we introduced the absolute zero reasoner, a system that self-evolves its training curriculum and reasoning ability by using a code executor to both validate proposed code reasoning tasks and verify answers, serving as a unified source of verifiable reward to guide open-ended yet grounded learning. So basically what it's doing is it's creating its own problems to solve like coding and math problems. It then attempts to solve

3:50 those problems and then using a code executor double-ch checkcks to see whether or not the solution that it came up with was correct or not. And then it continues to loop on this process. So despite being trained entirely without external data, AZR achieves overall state-of-the-art performance on coding and mathematical reasoning tasks outperforming existing zero setting models that rely on tens of thousands of in-domain human curated examples. One thing to keep in mind, this is really going to only be good at math and code. So, anybody out there going, "Oh, this is how we get to AGI." Well, not really. For something to be an AGI, it would have to have sort of world knowledge and be able to solve any types of problems, not just math and coding. For example, a model like this, you're not going to be able to go and ask it for business

4:36 advice and have it explain to you like chat GPT would on the best next steps to take. you're not going to be able to ask it about world history or current events or things like that. This specializes in creating math and coding problems, solving the problems that it creates, and then using a code executor to doublech check to see if the way it solved the problem actually worked. Pretty much anything outside of that domain, it's not going to work for. But what's interesting about this is that there is a lot of controversy out there about AI writing code, but then people being frustrated because the only reason that AI knows how to write code is because it sort of learned on everybody else's code that was trained into the original model. Well, with research like this, this is no longer going to be the

5:23 case. This is able to figure out and write new code that wasn't trained in by humans. It's actually creating the new code and then testing what it created. All right, let's talk about AI in advertising because AI is about to infiltrate every aspect of advertising. Companies that are trying to get in front of you are going to use AI to figure out the best ways to get in front of you. And services that rely on ads like Netflix and YouTube are going to leverage AI to try to put the best ads in front of you at the exact right time. Now, earlier this week, the president of Stripe interviewed Mark Zuckerberg, and he broke down his vision for how advertising will work in the future for the advertisers who use platforms like Facebook and Instagram. You know, it

6:08 used to basically be if you wanted to run ads, you had to like come up with your own creative and you had to come up with who you wanted to reach. You had to kind of have a sense of who your customers were and do a lot of your own measurement. And over time, we've just like automated more and more of this stuff. So the the basic end goal here is any business can come to us, say what their objective is. I want to get new customers to do this thing. I want to sell these things. Tell us how much they're willing to pay to achieve those results, connect their bank account, and then we just deliver as many results as we can. In a way, I mean, it's kind of like the ultimate business results machine. But it will be possible in the future that if you're running a small business or maybe even a larger one, you won't even need to have to start off with a creative. Maybe you can. Maybe you'll get better results. you could just come with a goal and uh how much that's worth to you and we'll just kind of deliver results across our platform. Now, I did cut out a few things from

6:54 that clip just to kind of speed up the pace of it, but I promise I didn't take away any of the context of the clip. But basically, the idea here is if you have a business that's trying to generate sales, you'll be able to go to Facebook and say, "This is what my business is. Go get me more sales." It will likely go and create the ad for you, figure out the targeting for you, and just continually optimize your ad for you to get you that result. Maybe you're trying to grow an email list. You can tell it, "I need more emails." It will go and do all of the process of creating the ads for you. You just give it the result you're going for, give it some money, and it will go out and exchange that money for that result using AI to continually optimize for you. As somebody who's constantly trying to grow a newsletter, Future Tools newsletter, that's actually something I'm fairly

7:40 excited to try. Now, when it comes to the platforms that we consume content on, we're about to see sort of an overhaul in what that advertising looks like to us as well, thanks to AI. According to this article on The Verge, apparently Netflix is trying to use AI to make ad breaks look less like ad breaks. Now, I'm not 100% certain what this is actually going to look like. The details are a little bit fuzzy still, but you can see the sub headline here. The new format will let advertisers make spots that blend in with its shows and movies. Netflix had an event this week called Upfront, and they demonstrated an example that placed the image of a product over a background inspired by one of its shows, like Stranger Things. I guess you're going to be able to have like various Netflix IP in the

8:25 background of your ads to make it look a little bit more native to the show that you're watching. If you're on Netflix's adup supported plan, that means you might see an ad that blends in with the show you're watching, whether it's Bridgetgerton or Wednesday, Netflix says advertisers can insert this new ad format in the middle of what you're watching or put them on the screen when you hit pause. When we check out the actual press release on the Netflix website here, we can see this section called creative formats. Netflix debuted a new modular framework for ad formats that leverages generative AI to instantly marry advertisers ads with the worlds of our shows. I'm guessing this is one of the examples here of what it could look like. We can see it's for the show Wednesday, but we also see Wendy's booking.com and Cheetos on the actual

9:10 ad. So, if you're on one of the Netflix plans with ads, you've got that to look forward to. YouTube is also changing things up using AI in their advertising. YouTube's trying to optimize the placement of the various ads using AI. They had an event this week called Brandcast where they made a few announcements and one of the announcements was Peak Points, a new product built with Gemini that identifies the most meaningful or peak moments within YouTube's popular content to place your brand where audiences are the most engaged. So essentially what this is telling me is that right when your YouTube video is getting really good, right at that moment where people are sort of leaning in excited to hear what you're going to say next, they know you're not going to click away at that point cuz you're into it. You're invested. And right at that point, boom, that's where the ad's going to show up.

9:56 I don't know if that's exactly what's going to happen, but that kind of seems like what they're going to do is they're going to wait till you're very, very invested in the show and then put those ads in those moments once they know, all right, they're not leaving because they need to see what happens next. And since we're talking about ads, I need to share this. Pika this week put out an ad promoting their platform. And well, just check this out. I'm not going to play with audio cuz I'm sure the music's copyrighted, but you'll still get the idea. So the ad shows this person basically peacifying everything. They're putting a claw for a hand. They're playing with their cat and then glitterifying their cat and then blowing their cat up and making it float away. And it seems like a pretty fun video of them just having fun playing with Pika turning their dirty laundry into butterflies and their plants into

10:42 balloons. And then it takes a bit of a turn. We see from outside the window. It zooms out and it turns out they're in a postapocalyptic wasteland and everything's on fire and falling apart. And then it jumps back inside and she's having fun turning her cat poop into a little mini tree, dancing around the house, and pretty much peekifying everything. While again, we're reminded that the world outside is burning down around them. A crazy person comes to the door and offers some cooked rats. And of course, she peekifies that person and turns the rats into cotton candy. She then proceeds to peekify everything in her life while eating the cotton candy. And then at the end, we are reminded that she's actually sitting there eating

11:28 rats. And the cat we've been seeing the whole time is a dead taxiderermy cat, I guess, with the tagline at the end, everything is terrible. No, it's not. So, I'm really curious. What do you think of Pika's new ad? Let me know in the comments. I think this one is uh really interesting. And I think some people are going to have some hot takes on this one. We got some fun new toys to play with this week. 11 Labs released their SB1 Infinite Soundboard. It's a soundboard, drum machine, and endless ambient noise generator allin-one. Describe the sound effects you want to hear. Then SB1 will generate them using our text to sound effects model. And then it turns it into a little pad that you can play with. So if you head over to 11labs.io/ound io/ound effects/ soundboard and there's a whole bunch of preset sound effects in here like

12:15 thunder cricket waves. There's a drum machine in here with like various kick drums and snares. You can even map them to your keyboard. movie sound effects, animal sounds, but you can create your own soundboards as well. So, if I come down to one of these empty slots down here and click on it, it opens up this sidebar here where I can generate anything like a burp. It gives me four

13:00 options and if I like one, I click add and now it's part of my soundboard. There you go. So, just a fun little AI enabled toy to play around with. And it's seemingly free to use, too. Stability AI and ARM collaborated to release stable audio small. This stable audio open small is an audio generator that can do, you know, sound effects and, you know, short song sort of sound bites. But it's so small that you can actually run it on a mobile phone. And they also open sourced it. Put all the weights up on Hugging Face. put all the code up on GitHub so anybody can use it, play with it, build off it, iterate, do whatever they want with it. And I actually did find a hugging face space that shows off what it's capable of. I don't know who named it and I don't know

13:47 who picked these examples, but I tested the prompt sad trombone. And here's what that sounds like. [Music] I don't know what happened at the end there. It only creates 12 seconds of audio, so I'm not sure exactly what the use case is for this on a mobile phone, but it could be run on a mobile phone. Also, if you ever wanted to know what a ghost peeing in a server room sounds like, here you [Music] go. I'm thinking maybe that end was a zipper. I don't know. We got a couple of

14:34 updates with chat GPT this week. The GPT 4.1 model, which previously was only available in the API, so you're able to build software and leverage it with your software, and you're able to use it in the various coding apps like Cursor and Windsurf and things like that. But now that model is available directly inside of ChatGpt if you're on one of their paid plans. Now 4.1 excels at code. So, if you want to have it write code directly inside of chat GPT for you, that's really going to be the main selling point of 4.1. Now, if we head over to Chat GPT, in order to access this new 4.1, again, on the paid plans, you'd come up to the top left and under more models, you can actually see GPT 4.1, great for quick coding and analysis, and 4.1 Mini, faster for

15:20 everyday tasks. My main go-tos are still 40 and 03, but if I wanted to mess around and write some code directly inside of chat GPT, I would use 4.1. Although, if I am writing code, I'm usually using something like windsurf or cursor. OpenAI also rolled out a feature inside of chat GPT where you can export documents as PDFs after doing deep research and the formatting is all dialed in and it's really nice. So, for example, here's some deep research that I did on a timeline of major AI advancements. If I click on this little share button at the top of my response here, you can see it pops up this window where I can either share the link or download a PDF. And of course, it gives us a nice well formatted PDF directly from chat GPT. Now, if you're a fan of

16:05 using Windsurf, for the most part, you would typically use an existing model from one of the major providers, Cursor 3.7, Gemini 2.5 Pro, ChatGPT 4.1, something like that. Well, now Windinsurf is actually rolling out their own model called SWE1. It's their first family of models optimized for the entire software engineering process. SWE1 is short for software engineer 1 and it's got three models. SUI 1, which is approximately Claude 3.5 sonnet level, SUI 1 light, and SU1 mini. Now, I'm imagining at the moment something like Claude 3.7 or Gemini 2.5 Pro are still going to be a little bit better than this model. But right now, if you do use Windsurf, this

16:51 model is available to all paid users, and it uses zero credits per prompt. So, there's really no limit to how much you can use this model right now. if this model's good enough for what you want to use it for. I also came across this little diagram here over on X from ADIT called this is literally how Microsoft is winning the AI race and I just really liked this image cuz as we know Microsoft has a 49% profit share in OpenAI and Microsoft also owns VS Code. Well, both Windsurf and Cursor the two biggest AI coding platforms out there were both forked from VS Code. OpenAI now owns Windsurf. They recently acquired them for $3 billion. And OpenAI was an early investor in Cursor. So they fully own Windsurf. They own a

17:38 percentage of Cursor. And Microsoft essentially makes money when people use Windsurf or Cursor due to this sort of pass through from OpenAI here. Now VS Code is free and open source. So, I don't know if Microsoft really makes money off of VS Code other than maybe some of the like paid plugins for VS Code, if there even are paid plugins, but I don't really think Microsoft makes much off of VS Code, but it sounds like they're going to make something off of these various forks through their relationship with OpenAI. I just found that fascinating. All right, a few last rapidfire things here. If you're a fan of the Android ecosystem, pretty much everything in that ecosystem is about to get AI. There was an Android event this week where they made a bunch of different announcements about Android

18:23 ahead of Google IO next week and they announced that Gemini is coming to the where OS smartwatches later this year. So, you'll be able to have AI conversations directly with your smartwatch. It's coming to Android Auto, allowing for hands-free conversational assistance in the car. It'll be able to do things like summarize and translate your text messages for you. And if you're just bored on the drive, you can just chat about current events or get a personalized news digest or get an answer to any burning question on your mind while you're driving. It's also coming to TVs. Gemini will also be available on your Google TV later this year. It'll be able to help with TV recommendations, such as finding the best content to watch, and I assume you'll probably be able to just have conversations with your TV like you will with your car and your watch and your phone. And pretty soon we'll just be

19:10 talking to every inanimate object in our house. I for one am excited about when Gemini comes to my toaster oven cuz I really need someone to talk to when I'm making my toast. There's a new AI model for text to Lego. You give it a text prompt and it will actually figure out how to make that thing with Legos. Carnegie Melon University unveiled Lego GPT this week. And some of the examples they showed are like a sofa, a bed, a mug, a chair, a camera. Now, this was actually trained on a pretty small data set. So there's definitely a lot of things it won't create. And most of what they show in their examples are pretty basic things, but they were also able to feed the information to a robot. And the robots were actually able to assemble the things for them that came out of Lego GPT. So if we sort of fast forward through this video here, we could see at

19:57 the end the two robots essentially made a very basic Lego guitar. I'm assuming from the text prompt guitar. Now this is available on hugging face and I gave it the prompt wolf howling at the moon. Imagine that. And well this is what it generated right here. So but we can see here the model is restricted to creating structures made of one unit tall cuboid bricks on a 20x 20x 20 grid. It was trained on a data set of 21 object categories. Basket, bed, bench, birdhouse, bookshelf, bottle, bowl, bus, camera, car, chair, guitar, jar, mug, piano, pot, sofa, table, tower, train, vessel. So no like animals or really any sort of living objects. And we can see here performance on prompts from outside of these categories may be limited. Now

20:44 if you do want to use this yourself again, you can find it by searching out Lego GPT on hugging face. But I will warn you now, it is extremely slow. I gave it this wolf howling at the moon prompt and it probably took a good 10 minutes before it gave me back this thing. But hey, it's the worst it's ever going to be. I mean, it's pretty bad right now, but it's the worst it's ever going to be. And finally, I promised you robots. We got robots this week. Elon on Twitter posted this video of a dancing robot. And we can see it's actually a pretty good dancer. I don't know if that is like teaoperated and somebody else is doing the dance motions and it's following along or what, but we can see in this video it's actually tethered to like this rail up here. And I came across this other video on Reddit here of the same Optimus robot, but this time

21:29 no tether and the video's a little bit longer. And we can see that this is actually 1x speed. So, this isn't like a sped up video. This is actually how fast this robot moves in real time. And it's able to do all of these like really cool dance moves. And the motion that these robots can do is really really quite impressive. And it doesn't actually need that tether that we were seeing in a previous screen. So, pretty cool. I mean, I don't really know what the use case of dancing robots are, but it's really cool to see that robots are able to do that kind of motion now. They're just getting more and more human like every single day it seems like. But that's what I got for you this week. I think this week is sort of a calm before the storm week. Next week I think is going to be absolutely insane. We have Microsoft Build up in Seattle followed

22:14 immediately by Google IO. They overlap those events in the same exact week. So we're going to be getting announcements out of Google next week. We're going to be getting announcements out of Microsoft next week. OpenAI has a tendency to try to overshadow the announcements of some of the bigger companies. So, don't be shocked if you see some OpenAI announcements next week. We're also hearing rumors that Enthropic is about to release like Opus 4. There's also rumors about Deep Seek R2 coming out soon. Elon Musk promised that Grock 2.5 would be out like a week and a half ago, and we still haven't seen it yet. But all of this is kind of bubbling up, which means there's probably going to be a really big week next week. And well, really the rest of this month is probably going to be a ton of announcements. I'm going to be up in

23:00 Seattle for Microsoft Build and then down in Mountain View for Google IO. And if you want to stay looped in on all of the latest AI news that's coming out and really keep your finger on the pulse, make sure you like this video and subscribe to this channel and I'll make sure that you stay looped in and more videos like this show up in your YouTube feed. Thank you once again for tuning in. I really, really appreciate you. If you want to find some really cool AI tools, check out futuretools.io. This is where I share all the cool AI tools I come across. Also, there is so much AI news every single week. I've started to really filter it down and only share the stuff that I think is really impactful or really quite interesting. But for all of the rest of the AI news, maybe the smaller little tidbits, you can check out the AI news page where I share it all and it's updated every single day,

23:47 so you'll still continue to be looped in on all the latest AI news. Again, that's futuretools.io. And of course, there's a free newsletter. So maybe you don't want to check in to the website every day. Just twice a week, I'll send you an email with just the coolest tools I come across and the most important AI news that you need to know. It's totally free. You can get on board over at futuretools.io and click this join the free newsletter button. And thanks again for hanging out with me and nerding out with me today. It's been really, really fun sharing all this cool stuff. I really, really appreciate you and hopefully I will see you in the next one. for bun.