Huge AI Breakthroughs That Flew Under The Radar (And other AI News)

Well, this week feels like a week where there's been some legitimately big breakthroughs in the world of AI and I don't want to waste your time, so let's dive right in. Now the first thing I want to share doesn't actually feel like one of those legitimately huge breakthroughs, but it is one of the things that I feel like the most people are talking about. So I'll start right here with ChatGPT. Now having a tasks feature inside of ChatGPT, you can now ask the AI assistant to schedule reminders on recurring requests. Now this is only available in the various paid plans and if we jump into Chat GPT here, this is how you'd find it up on the left where you select which model you're using. If you're on one of the paid plans, you'll have the option for GPT4O with scheduled tasks. You can select that and it will open up this version and you can give it a prompt like every morning at 9am Pacific, share the most interesting news in the world of AI with me. I'll go ahead and submit that and you can see it confirmed. Got it. I'll keep you posted every morning. Now if I click on this little edit button and it brings up a window where I can actually edit the task share AI news, here's the instructions. Tell me the most interesting news in the world of AI and then we can mess with the schedule and the time that it comes across every day. And you can see you can set things to daily, weekly, monthly, annually or give it a custom option. And if you turn on your desktop notifications here, it will notify you in a pop up notification. Now you might be wondering, well, what are some good use cases for this feature? Well I came across this really good thread from Min Choi here, a really good follow on X if you're not following him already, where he shared a list of all sorts of ways that people are using it. For example, daily stock market analysis every weekday morning at 9:30 before the market opens, send executive summary of current stock market analysis focusing on my portfolio of stocks Nvidia, Tesla, Apple. Here's one where they asked Every morning at 9:52am create a new very difficult coding problem in Python to solve in Canvas. Now this person could get into a daily routine of solving Python problems. Here's one where somebody wants to practice Italian on a weekly basis every Thursday. Thursday evening. Here's another one Every Friday morning suggest fun weekend activities near Me in Washington D.C. metro area based on local weather and upcoming events. I'm interested in family activities. Every Sunday at 10am Create a meal plan for the week with a grocery list that fits my keto diet. Each morning at 7am suggest a quick 20 minute workout routine based on my fitness level and the equipment I have at home. I'm active and I have dumbbell weights from 5 pounds to 90 pounds. This thread has some more examples and of course I'll share it in the description below so you can find this THR and read through more of these examples. But I think you get the idea. Anything you want to do on a daily or weekly basis that an AI prompt could also help you with is kind of what you'd use this resource for. In other OpenAI news they got a new board member this week in Adebayo Ogunlesi. I don't know if I pronounced that right, but this person is a Founding partner, Chairman and CEO of Global Infrastructure Partners and a Senior managing director at BlackRock. Now if you're not familiar with BlackRock, they are the world's largest asset management company overseeing a approximately $9.42 trillion in assets. So this is a massive company that has large political influence and insanely deep pockets. OpenAI also partnered with Axios this week to work even deeper with the news industry. We've been seeing this trend of OpenAI partnering with all of these news publications so that they can use their news as sort of real time updates to Feed ChatGPT. And this week OpenAI wasn't the only company to go that route. We also got the news that Google inked deal with the Associated Press to bring more real time info to Gemini. This article on TechCrunch here even goes on to compare them to OpenAI reminding us that they've partnered with companies like the Financial Times, Axel Springer and News Corp. And now of course Axios like I just mentioned. And the French AI company Mistral signed a deal with AFP to offer up to date answers in their Chat GPT competitor lechat. Just a day after Google inked a deal with the Associated Press, Mistral has announced a content deal with Newswire to agents France Press to improve the accuracy of answers in Le Chat. At the moment it seems like Anthropics Claude is really the only large language model that hasn't partnered with any of these news sources to actually offer real time updates and news within their large language model. All right, now let's get into the stuff that I feel like is the really truly groundbreaking stuff that came out this week, starting with Titans and well to be honest, this one didn't actually come out this week. It came out at the end of December, so this one's a little over two weeks old now. However, it seems seems to be getting renewed attention this week, especially after my buddy Matthew Berman did an excellent breakdown of this paper in a recent YouTube video. This research comes out of Google and it argues for a new architecture called Titans as opposed to the traditional transformer model. That is what underlines pretty much all of anthropics and chat GPTs and all of the normal existing large language models we're used to. It says here we present a new neural long term memory module that learns to memorize historical context and helps an attention to attend to the current context while utilizing long past information. It's basically saying that this new model has a long term memory and a short term memory. It's much better at focusing on what it is currently reading like a short term memory, while also maintaining and remembering information that it learned long before what it's reading right now. Something else that's really interesting about this paper is that on page three here it says we designed this memory module so an event that violates the expectations, AKA being really surprised, is more memorable. Just like the way the human brain works, if something out of the ordinary or really unexpected or really surprising happens, those things tend to get burned into your memory a little bit better than the sort of average everyday mundane things. And this model seems to pick up on those big surprising things more so than it picks up on the mundane little things. Think of this research this way. Existing large language models that use Transformers are kind of like reading an entire book all at once and trying to remember everything that's in that entire book and then answer questions based on all of that information. It gets overwhelming and computationally expensive. These new Titans, however, work more like the human brain. As it reads through information, it keeps the most important information, the most surprising information and things like that top of mind, while actually being able to sort of ignore and disregard the stuff that's less important to the overall context of the information that it was fed. So instead of like reading the whole book and committing the entire book to memory, is just committing the key portions of the book that are going to be most important to the memory and then responding based on that information, making it a lot less computationally intense and also allow a much larger context window. In fact, this paper says it can effectively scale to larger than 2 million context window size. There is another new model that did get announced this week from the company SAKANA AI called Transformer squared. And this is another one that feels like a really, really groundbre model looking at the paper itself from the abstract. This is the interesting part here. During inference, basically what happens after you give it a prompt? Transformer squared employs a two pass mechanism. First, a dispatch system identifies the task properties and then task specific expert vectors trained using reinforcement learning are dynamically mixed to obtain targeted behavior for the incoming prompt. Basically, what it's saying here is that when you give it a prompt, it's going to decide what specific expertise is needed to respond to that prompt, and that's the first pass. And then it draws from the expert knowledge within its training data set for that second pass to best respond to the original prompt. So imagine a musician, if they see a set of sheet music, consider that like the prompt, they're first going to determine whether the music is, you know, jazz or rock or ska or whatever style of music. Once it figures out what style of music that's the first pass, it will then employ different techniques on the second pass. Based on its understanding from the first pass, all of this expertise is essentially all in its training data already. But because it uses this two pass system, it first figures out what expertise to use from its training set and then on the second pass uses that expertise to then give a response. We can see here that the subheading is self adaptive LLMs. So essentially it is adapting to the information that you're giving it by picking the right set of expert information within its data set to then respond. They can also scale the amount of compute that they use based on the complexity of the task. We can see here on their webpage. These systems will scale their compute dynamically at test time after you give it the prompt to adapt to the complexity of tasks they encounter. Now, probably most interestingly about this research is that it actually seems to combine expertise when you ask different questions. So during that first pass, when it decides which expertise to pull from, it's actually pulling from multiple expertise within the database. So right here it says a particularly intriguing finding comes from analyzing how few shot learning combines different z vectors to tackle tasks when solving math problems. Contrary to expectations, the model does not rely exclusively on its math problem math specialization. This suggests that complex mathematical reasoning benefits from combining mathematical programmatic and logical reasoning capabilities. They observe similar unexpected combinations across other tasks and models, highlighting the framework's ability to synthesize diverse types of expertise for optimal performance. So essentially it does. Its first pass figures out which expertise to use and not always the expertise that you assume it's going to use is solely what it uses. It might use some logical expertise, some creative expertise, some programming expertise to solve the problem that you gave it, and it's not always the exact sort of expertise you'd think it would use. It's very good at finding correlations that maybe humans aren't as good at finding. So between the Titans research paper and this new Transformer Squared paper from Sakana AI, we're actually seeing some huge leaps and some potentially new architectures that could really level up a lot of these language models. AI tools like ChatGPT, Claude and Gemini are amazing, but right now, to get the best results from them, you still kind of need to use pretty good prompts. And that's why for this video I partnered with HubSpot because they recently put out a free guide with over a thousand prompts for productivity and marketing. This thing is literally a prompt playbook with so many prompts. Just check out this table of contents over here. It's got prompts for marketing strategy, brand strategy, copywriting, social media, brand identity, brand analysis, SEO, paid search, influencer marketing, pricing strategies, visual creation, brand sales strategies, CRM strategies, productivity, and so much more. This is like the end all be all guide to prompting. In fact, let's zoom in and pick one at random here. Create a content brief for an infographic that showcases how businesses unique selling proposition addresses a Persona's challenges, including key talking points and data to include. It's going to map out a whole infographic that you can use for your business. Or how about up? What is the optimal price point for our business's product or services that will appeal to our Persona while maintaining profitability? This playbook has every prompt you need to leverage tools like Chat GPT to blow up your business and brand in 2025. And once again, it's totally free and you can find it in the link in the description below. And thank you so much to HubSpot for sponsoring this video. Now the other research that came out this week that feels absolutely wildly groundbreaking to me is this new Matter gen from microso. This is a generative AI that can design new materials now in the same way that tools like Midjourney and Flux and Dall E can create brand new images from prompts. This Matter gen can create brand new materials from prompts. And not only that, but they actually released and open source this research. It's available under the mit license on GitHub right now. So check this out. The discovery of lithium cobalt oxide in the 1980s laid the groundwork for today's lithium ion battery technology. Those are the batteries that now power modern mobile phones and electric cars, and probably most of the devices you have that run on batteries. Well, now we can imagine AI actually helping us find new materials that could make more improved battery technology, more efficient solar cells, new materials for capturing carbon dioxide from the atmosphere, stronger and lighter building construction materials and things like more efficient fuel cells. This opens up so many opportunities. Much like an image diffusion generates pictures from a text prompt by modifying the color of pixels from a noisy image. Mattergen generates proposed structures by adjusting the positions, elements and periodic lattice from a random structure. The base model of Mattergen achieves state of the art performance in generating novel, stable, diverse materials. It's trained on over 608,000 stable materials from the Materials Project and Alexandria databases. Under the Looking Ahead section, it says this will have profound impact on how we design materials in broad domains including batteries, magnets and fuel cells. I feel like we're truly entering that era right now where AI is literally helping solve the world's problems. I really feel like we're living in the best timeline right now. And that's not the only AI news from this week either. The company Minimax has their new Minimax 01 open source large language model. When comparing it on benchmark tests, it holds up pretty well against models like GPT4O, Claude 3.5 Sonnet, Gemini 2.0, Flash, and many other models, which always seems impressive to me from an open source model that we're at that GPT4.O level or Claude 3.5 sonnet level. But with open source models, all right, shifting away from large language models and generative AI for a second, there's been a little bit of news to come out of the White House this week around AI, including this new AI diffusion rule. So according to this statement from the White House, the US Government wants to limit the availability of various processing chips and hardware to run AI to specific countries. If we look through the new rules here, no restrictions apply to chip sales to 18 key allies and partners. Chip orders with collective computation power up to roughly 1700. Advanced GPUs do not require a license and do not count against national chip counts. Basically saying that universities, medical institutions and research organizations should fall under that threshold. Entities that meet high security and trust standards and are headquartered in close allies and partners can obtain highly trusted universal verified end user status. So Basically, if the US government trusts you, you can order more GPUs. And I don't really want to go into all of the Details here, but I will link this up below. Basically the US Government is saying we're only going to allow companies to sell high volumes of GPUs and processing chips to countries that we actually trust. Most countries could still buy them in lower volume. And it feels like this is kind of the US Government saying we don't want this technology to be given to China. Another country that doesn't appear to be on the allies in trusted countries list is Israel, which is interesting because Nvidia just actually acquired a company out of Israel recently and they were not happy with this decision from President Biden. Putting out an article basically saying that this feels like it's going to slow down tech progress, it's a government overreach, it's likely not going to improve national security, and that it will weaken American competitiveness. That's essentially Nvidia's arguments here in this article. They also claim that it is cloaked in the guise of anti China measure. And all said and done, Nvidia is very, very unhappy about this decision. Biden also signed another executive order this week to speed up AI data center construction, basically giving the Department of Defense and the Department of Energy to lease federal sites to private companies that are building gigawatt scale AI data centers and clean power facilities. It also tells federal agencies to prioritize and speed up permitting of AI infrastructure and also seeming allows these companies that are building AI data centers to sort of bypass some of the red tape from the various environmental protection agencies. All right, moving on to some topics that are a little bit more fun to discuss. Luma Labs just rolled out a new video generation platform called Ray2. A lot of people are actually kind of comparing it to Google's VO2 and the videos I've seen so far look pretty solid. We can see some examples right now on the screen. We can scroll through here. We can see that it does natural motion. We can see that supposedly it's really good at following instructions from your prompts. It has a decent understanding of physics and simulation. It's photorealistic, it can produce cinematic scenes and there's all sorts of examples on their website here. But I just want to go ahead and test it. So I jumped into my Luma Labs account and tossed in some prompts here. They're my normal prompts that you typically see me test. I've got my, my wolf howling at a moon and it generated two videos for me here. This one, it looks like the wolf's neck kind of breaks and the moon is sort of moving around weirdly but the first one is pretty decent, I suppose. I wouldn't really call it photorealistic, but it meets my requirements. Next up, I did my typical monkey on roller skates, which surprisingly, aren't that bad. The monkey kind of is going really, really fast in this one, and it looks like the legs sort of merge a little bit. The one below I'm a little less impressed with. You can see the legs sort of like SW places with each other at points. But again, considering where we were with this stuff, like 18 months ago, this is still insanely impressive with how far it's all come. And then finally I did my prompt of a cat eating a taco. And not too bad. The first one looks like a cat sort of devouring a taco. I actually think that one did a pretty good job. The second one, yeah, I don't know what's going on with that. It looks like more like a cat sort of exploding a taco and the cat's got a jacked up face. But yeah. This is now available inside of Luma Dream Machine. There's no wait list like you get with VO2, and anybody can use it over at the Luma Labs website. While we're on the topic of video, the company Freepik released a feature where you can actually generate AI videos that actually include sound. It feels a little off to me. It's. It's pretty cool. Pretty impressive that we've made it this far, but still not perfect. Here's like a singing bird. Here's like a can in the snow that sort of explodes. A Lego parrot. I haven't actually played with this one myself yet. Somebody actually pointed it out. Every week I do one of these tweets where I share. Here's the news I'm looking at, and this is one that somebody pointed out that maybe I missed. And since we're on the topic of Free Pick, they also released a new feature where you can remove the background. We have a lot of tools that can do this now, but apparently if you're a Free Pick user, you can do it directly inside of that app as well. And we're getting a new image generator out of Runway. They have a new model called Frames, and it's available now to the early beta testers. I'm in their creative partner program, so I do have early access to this one. But Runway, in the past, when they roll out tools to the creator partner program, it usually is only about a week or so later before it rolls out widely. So I imagine this will be rolling out Widely, very, very soon. But let's go ahead and try a couple inside of it. Let's do our typical wolf howling at the moon to start. It was actually fairly quick. It took less than a minute. And all of these images are pretty impressive. I actually really like the color palettes of these bottom two especially. Let's try prompt adherence real quick and do something like a three headed dragon watching TV while eating nachos and wearing cowboy boots. I've tested this prompt before. This is my go to when I want to test how many things can I squeeze into a single prompt. And here's what it came up up with. We've got three heads, we've got a tv, we've got chips. I can't tell if they're nachos or not. And they're wearing one cowboy boot and there's some cowboy boots on the side. It got all the elements, they're just not all composed properly. This one looks more like a two headed dragon with no cowboy boots. Well, I guess the boots are on the side here. This one, I only see two heads. This one, we've got three heads and some boots and some nachos and a tv. I mean it got all the elements in there. It just didn't quite, quite pull them all together. Very great, but still a really impressive model. I'm going to try one last prompt. I want it to be super realistic. So let's go ahead and try one of these styles down here Also like high contrast, warm. And let's just do a photo of a woman standing on a sidewalk with a city skyline in the background. And this is what it comes up with. They look semi realistic. I wouldn't say it's the most realistic image I've ever seen. That one, the woman looks real, but the way it's sort of composed, it doesn't look super realistic to me. This one just kind of looks like AI to me. And this one, I mean they're not bad. I wouldn't say they look super realistic. They almost look more like a realistic woman standing in front of like a green screen image or something. But not bad at all. We'll probably play around with this some more on the Monday livestream 11am Pacific on the YouTube channel and really see what we can get it to do. Now another tool that I came across this week that I really, really am excited to play with, I haven't even tried this yet is from Korea, AI. They have a new 3D tool. So check this out. So they have an image here with a car in it. And they're able to select the. And then they're able to drag and drop and rotate the car around. It turned the car into a 3D object. They then drop like a private jet into the image and then they can rotate the private jet around as well and place it in the image. And it also understands what the rest of that image should look like in 3D. This to me seems pretty wild. And I'm gonna go ahead and jump in and try to test it real quick. Let's start by generating an image of a wolf. Wolf howling at the moon. Let's go ahead and download this one right here real quick. And then we'll jump over to the real time tool here. This is apparently where the 3D image tool is pulling our image of a wolf here. Now if I right click on the image, there's an option that says convert to 3D. So let's go ahead and convert that to 3D. Merge the wolf with the moon here. So I've got a wolf with a moon for a head, but I can sort of rotate it around and we can see our real time image changes as I rotate my wolf moon head thing. Oh, that's hilarious. I went ahead and pulled in an image that I created previously of this like orc here. Let's go ahead and convert that to 3D, see if we get a little bit better of a result. Here's the 3D version of the orc that it generated for me. I can actually select here and rotate it and theoretically the orc on the right should follow how I'm rotating it. But it doesn't seem to be syncing up very well with the way I'm rotating it. So maybe I'm doing something wrong here if I am somebody in the comments, let me know if you've played with this new 3D feature, why isn't it composing my or to look like what I'm putting over here on the left. You would think that it would be turning sideways if I turn it sideways on this side. I don't know. Let me know if I'm doing something wrong. I would like to figure that out. Maybe we'll play with it more on the live stream. All right, we've got a bit more smaller news that came out this week, so I'm going to kind of move into rapid fire mode and just quickly talk about some of the other things that happened this week. Adobe came out with a new AI tool that can edit 10,000 and images with one click. Firefly Bulk Create can automatically remove, replace or extend image backgrounds and huge batches. I don't believe this is available just yet, but it seems like it can be very helpful for maybe somebody that owns an e commerce site and needs to edit a ton of images on their site all at once. Apparently companies like OpenAI and Google are paying thousands for content to train AI models. According to this Bloomberg article, YouTubers are selling their unused video footage to AI companies. So far, nobody has approached me or asked me to buy any of my unused videos. But if somebody wants to give me money for a video that I've never actually published anywhere, I'm happy to do it. The company Wyze that makes personal security cameras now uses AI to summarize your motion events and some of the examples that they share on this website I found to be pretty funny. Like this one Babies escaping from crib nursery detected. Three babies are helping each other climb out of their cribs. The babies are laughing at 8:29. Here's another example. Dad and sons playing baseball. Driveway cam detected. A dad and three boys are playing baseball in the driveway. The dad is hit by the baseball in the crotch at 10:20am so if you have a WYZE camera, you should be able to see what happened on the camera now without actually looking at the video footage. And in similar news, if you remember, Apple had that feature where it was summarizing text messages. But some of the examples of the summaries were mom tried to commit suicide when the mom's original text said that hike almost killed me. Yeah, some of those summaries weren't working out very well and Apple is officially pausing the notification summaries. At least for now until they can make them a little bit better. This story is absolutely wild to me. A woman in France was duped by an AI version of Brad Pitt. The scammer managed to get $850,000 out of this woman by claiming that he had cancer and needed money for care and that his divorce with Angelina Jolie was tying up all of his fucking funds. And this lady in France bought it into the scam and sent him a bunch of money. If you're curious about the pictures that this scammer used, well, they're available on Twitter. These are the various AI generated photos of Brad Pitt getting cancer treatment and this woman Ann, actually getting personal messages from him. I like pointing out stuff like this because I want to remind people to be safe out there. It is getting much easier and easier to get scammed by this kind of thing. I think many of us More tech savvy people would have seen some, some red flags pretty quickly, especially this one down on the bottom right where Brad Pitt's little head is popping up during surgery. Like, who took that photo for him? But you know what? People in your family, your loved ones, maybe more elderly members, are probably not going to be as observant as some of us more tech savvy people. So I think it's a good reminder to, you know, stay safe out there and really sort of second guess almost everything you see. This is super interesting. Sony has a new patent that includes an AI system to predict a players behavior so that it can reduce online lag. By using a camera to study and record player actions, Sony's exploring how accurately AI can predict player behavior in a big win for AI this week, AI triumphs over venom. A revolutionary snakebite antidote is unveiled. Scientists have developed innovative proteins capable of neutralizing lethal snake venom toxins, offering a promising, safer and more effective alternative to traditional antivenoms. And finally, finally, let's end with a robot video. This is the Unitree G1 robot and we can see that the walking of this robot has gotten a lot better. But in a second here, we're going to actually see this robot run. And that to me is actually starting to get more and more terrifying because if there ever is a robot uprising, do we really want the robots to be able to run at us? I don't know, but this is still really, really impressive. These unitree robots, I believe they're only costing about $16,000. So in it's really, really impressive how far robotics has come just in the last year alone. I think we're not far off from being able to have robots in our houses that do a lot of our chores and workforce around the house. I mean, some of these new demos we're getting are pretty dang impressive and I actually saw these roaming around at CES a couple weeks ago, so pretty crazy stuff to see. And that's what I got for you this week. Hopefully you learned something new and you feel more looped in on the world of AI news this week. If you enjoyed this video, give it a thumbs up and consider subscrib to this channel. I make videos about AI news every single week, along with tutorials and breakdowns and live streams on Mondays. And also if you want to stay looped in on the latest AI tools and the latest AI news, check out the FutureTools IE website. I share all the cool AI tools I come across. This news page here is updated every single day with way more news than I can even share in these videos. And I've got a free newsletter. If you join the free newsletter, I'll send you two emails a week with just the quick, coolest tools and just the most important news in the AI world. And as a thank you for subscribing, I'll also hook you up with the AI Income database, which is a cool list of ways to make money using these various AI tools that are available. It's all free. It's all available over at FutureTools IO. Thank you so much for tuning in. Thank you to HubSpot for sponsoring this one. I really, really appreciate you hanging out with me, nerding out with me today and hopefully I'll see you on Monday's live stream. If not, maybe I'll see you in one of my future videos videos. Really, really appreciate you. See you in the next one. Bye Bye.