AI News: Mysterious AI Shocked Everyone (Then Vanished)

A lot of interesting things happened in the world of AI this week, and I don't want to waste your time, so let's just get right into it. One of the most interesting things that happened this week that had the whole AI world buzzing was this mysterious GPT2 dash chatbot large language model that popped up on a website called Chatbot Arena. And it was interesting because although it was called GPT2, it seemed to be outperforming even GPT4 and, and Claude Opus by most people's tests. It was weird that it had a different naming convention. Almost everything else is GPT dash and then the number. This one just had GPT two and then dash Chatbot. What's even weirder is nobody was claiming this large language model. No company stepped forward and said, we are the ones who developed this. Ethan Mullock here, a professor at Wharton who studies AI innovation in startups, how had this to say. There's a mysterious new model called GPT2 chatbot, accessible from a major LLM benchmarking site. No one knows who made it or what it is, but I've been playing with it a little and it appears to be in the same rough ability level as GPT4. He then gives some examples of some of the prompts that he tested here and got some good results from. And then we have Andrew Gao here, who claims that the GPT2 chatbot solved an international Math Olympiad problem in. In one shot, meaning they just gave it a single prompt and it solved the problem. Now, I'm not a math genius. I don't even know how to read most of this stuff on the screen, but if you know what you're looking at, this is his findings. And then there's Dr. Alvaro Cintas here, who works at the rundown, and he gave it a test to see if it could code a simple Snake game, and it actually managed to code this Snake game on the very first try. I mean, I don't Even know if GPT4 will code a Snake game on the very first try right now. And then Sam Altman, I guess, having fun with his following, decided to add some more confusion to the mix by tweeting, I do have a soft spot for GPT2, and one thing to note is how he wrote GPT2. He didn't put GPT2, which was the naming convention of OpenAI's GPT. He put GPT2 without the space. In fact, this was even an edited post where, where he originally wrote, I do have a soft spot for GPT Dash 2 and then he edited it to get rid of the dash. At this point, I think Sam Altman is just trolling everybody. The speculation at this point is that OpenAI released a new model, a new version of GPT4. Some people were speculating it was GPT 4.5, other people were speculating that this was actually GPT2, just trained on new information from GPT4. But to this day we we still don't really know what the heck was going on with this. And on Wednesday of this week, Sam Altman, while speaking at Harvard, told the audience that the mystery bot was not GPT 4.5. So as of right now, we still don't really 100% know what the heck this GPT2 chatbot actually was or who made it. But we do know that in the testing it was actually outperforming some of the state of the art models. That really threw the AI community for a loop this week while while we're on the topic of OpenAI, they finally rolled out the memory feature in ChatGPT to all of the plus users. This literally came about two and a half months after they announced that Chat GPT was going to get this memory feature. At the moment it's also not currently available in Europe or Korea, but if you are in one of the countries that has access, you can log into your ChatGPT account, give it some information about yourself and actually tell it to remember these things. And, and you see it will save it to the memory. You'll get a little notification saying memory updated. If at any point you want it to forget stuff that's in your memory, you come down to your name, click on Settings, click on Personalization, and then under memory click on Manage and you can see all that it is remembered about you and delete any of the memories it has saved. If you want a sort of incognito mode of ChatGPT, you can come up to where it says ChatGPT 4 at the top left, click on that and then turn on Temporary Chat. And it won't remember anything, it won't save any of the details, and it won't use any of the information for OpenAI's training data either. There's also a rumor going around right now that OpenAI is getting really, really close to launching its own search engine. Lior here over on LinkedIn says OpenAI is about to go after Google Search. This could be the most serious threat Google has ever faced. OpenAI's SSL certificate logs now show they created Search.

Microsoft, Bing would allegedly power the service. And then we have a screenshot here showing this search.chatgpt.com within their logs. And here's something interesting that Pete Huang from the Neuron pointed out. He said, here's something that Sam Altman said on the Lex Friedman podcast, the intersection of large language models plus search. I don't think anyone has cracked that code yet. I would love to go do that. I think that would be cool. So, kind of hinting that he wants to do something with search. By the way, Pete was recently on the Next Wave podcast, so if you haven't heard that yet, make sure to check that out. Pete also seems to somehow have an idea of when this might Launch. He wrote search.chatgpt.com May 9. Now, where that information is coming from, I don't know, but Pete's a good friend of mine and I trust him as his source. So I would say there's a pretty good chance something's gonna happen on May 9th. Whether it's actually the release of the search engine or it's the announcement of the search engine, I don't know. He's a pretty credible guy. This open AI news actually came out at the end of last week, but it was after I recorded, so I thought I'd mention it here, that Apple and OpenAI have intensified talks about incorporating OpenAI's tech into future smartphones. But I've also reported recently that Apple has been in talks with Google about possibly using Gemini. But from the sounds of it, it just kind of seems like Apple is exploring all the options for who might power the AI in upcoming versions of iOS. A lot of people are speculating that at WWDC this year, Apple is going to announce a new version of Siri with better AI features incorporated into it. And it still sounds like they're kind of shopping around what model they're going to use for that AI. Last week Sam Altman spoke at Stanford, but this week the video was actually put online on YouTube for us to watch. Now, most of the discussion here with Sam, he kind of gives his stock answers. Most of what he says we've heard on a lot of podcasts before. This was the most interesting little clip from this interview here. GPT4 is the dumbest model any of you will ever have to use again, by a lot. But it does kind of suck to ship a product that you're embarrassed about, but it's much better than the alternative. And in this case in particular, where I think we really owe it to society to deploy iteratively One thing we've learned is that AI and surprise don't go well together. People don't want to be surprised. People want a gradual rollout. So it seems to me that Sam is really hyping up GPT5 and what we can expect next. When asked in this same interview what's gonna be different about it, he basically said, it's just gonna be a lot smarter. So the questions that you ask it, it's gonna give a lot better answers and just be smarter. And that was really all he said about it. But the way he sort of doubles down on this is the dumbest model you're ever gonna use. Kind of makes me think that the models that he's seeing and they're testing right now are quite a bit better. This week we got the first music video generated with OpenAI's Sora. And before I get to it, I gotta share a little bit of a pet peeve. Whose idea was it to design a website that looks like this? Like 50% of the screen is taken up by an ad. Another, like, I don't know, eighth of the screen down here is taken up by this subscribe box and the content window is this large. I'm sorry, but the UI UX of a site like this just sucks. Anyway, end of rant. Here's what the actual video looks like. I muted the audio on it because again, copyright. I don't know if I can play play the audio. I'll probably get in trouble for it, but it's actually looking pretty good. I'm not gonna play the whole video. I will link up to it in the description so you can watch the entire video if you want. But it's pretty impressive actually. There's some funkiness with the walking and you can tell it's AI generated, but I thought it was pretty cool looking. But Sora might be getting some competition soon. There was an announcement out of China for a new AI video tool called Vidoo. And here's the demo video that they put out for it. Some of it's in Chinese, I'm not sure why it started in English, but this is the footage that they were getting out of it. And it looks pretty good that especially. But it doesn't quite look like Sora level to me yet. The people are pretty realistic, but the movement and the animation doesn't quite look on Sora level. And you can also tell that they actually did videos that were designed to be side by side with Sora because it's a lot of the same kind of scenes Which I find very interesting. But if you put them side by side, Tim over at Theoretically Media actually did a really great side by side comparison. These still aren't quite at SORA level. Not bad. Way better than what we were generating a year ago. But I don't really feel like this is real competition for Sora yet. This apparently generates 15 second videos. Sora does 60 second videos. Sora looks pretty realistic, this one. There's still a little bit more weirdness in my opinion. Really dang good. Probably better than what we're getting out of things like Pika and Runway right now. And just like Sora, we don't have access to it yet. This was just a sort of teaser video that was put out and we can't use it just to remind you what SORA looked like. Here's some of the videos from the Sora demo. We saw a very similar Jeep in the video demo. This one just. It's another level, let's be honest. Here's Sora's version of that stack of TVs. Definitely a lot more coherence. You can definitely tell what's going on on more of the TVs. Again, not quite there, but still cool to see these advancements. Ideally stuff like Vidoo gets released so the public can use it and that sort of lights a fire to let us also use Sora. Hopefully OpenAI wants to stay competitive. They push out SORA for us to use. But I'm not holding my breath. More companies are suing OpenAI. This news came out this week. US newspaper sue OpenAI for copyright infringement over AI training. A group of newspapers, including the New York Daily News and Chicago Tribune sued Microsoft and OpenAI in New York federal court, accusing them of misusing reporters work to train their generative artificial intelligence systems. This feels very familiar. I feel like there's a lawsuit like every other week with OpenAI. It's almost becoming background noise. Usually when this kind of stuff pops up, the courts kind of shut it down in favor of OpenAI. But as this unfolds more, I'll share whatever news comes out of it. You ever find yourself having to walk people through the same thing over and over again to different people? I know with my growing online business, I definitely do and that's why I partnered with Guide on this video. Guide is an AI powered tool where you just go through and do the steps of the thing you want to show people and it will actually create a step by step tutorial in the process. Simply install the free Chrome extension. Make sure you Got it pinned up in your top bar so you can actually see it. And now let's say I want to show a team member how to add a new piece of news to the Future Tools website. I simply click on the guide button, click Capture, select the type of content we're creating and give it a name. And finally we click Go. Now I simply just go through the steps. If I come over here, click on this button, click on News, click on New News up here, type the name of the news article, enter the URL, the source and the date and then click Save as Draft and then I will come through and review the drafts. I've walked through the step by step process. Now I can click the guide button again to end my recording and we can watch as the AI actually creates a full tutorial for me. We've actually got a video here now and as it walks through the video you can see it shows off each step that needs to happen. Click News, click New News, click Here, fill the name of the news article, click here. And as you can see it created a full on tutorial complete with screenshots and step by step instructions. And that now I can share this document with anybody on my team and they'll now know how to add news to the Future Tools website. I can now share this tutorial wherever I want or embed it wherever I need to to give team members access to this education guide is actually a tool that I was using before they even approached me to sponsor this channel. So I'm absolutely happy to tell you about it and to recommend it to anybody that finds themselves explaining things more than once. And again, you can use the extension completely for free. I'll make sure the link is right at the top of the description below this video. And thank you so much to guide for sponsoring this video. Also this week, Anthropics Claude rolled out two new features. They now have a team plan which allows for collaboration and shared conversations within anthropics Claude. And they also released an iOS app. I actually showed off and played around with the iOS app on a video earlier this week. So if you want to see more of what the Claude app looks like and hear my thoughts on it, check out this video called the Perfect AI Prompt Forever Everything. Because I dig in a little bit about the new Claude iOS app. At the end of last week we learned that the Biden Administration here in the US established the Artificial Intelligence Safety and Security Board with 22 initial members. And among those members are people like Sam Altman, Microsoft CEO Satya Nadella, and Alphabet CEO Sundar Pichai. It's just interesting to me that the people on the safety board that are probably going to be steering the direction of regulations around AI include the CEOs of all of the biggest companies in the world that are building the AI. So hopefully this doesn't lead to the big companies creating regulations that really benefit them and push out the little guys, which tends to happen in scenarios like this. There are some non AI related CEOs involved as well, including the CEOs of Delta Airlines, Northrop Grumman and Occidental Petroleum, also Maryland's Governor Wes Moore, Seattle Mayor Bruce Harrell, and the leaders of two civil rights organizations. Honestly, to me it seems like a really tough decision to figure out who should go on this board, but I do wonder about some of the potential conflicts of interest of having some of the richest billionaire CEOs on the board that's going to steer the direction of where all this AI is headed. There was a big announcement this week out of GitHub. They announced the GitHub Copilot workspace. To me, this looks a lot like what we saw out of that Devin demo that was circulating for a while that was eventually kind of debunked as sort of being faked a little bit. This seems to be somewhat like an AI agent from GitHub where you tell it what you want it to code and it actually pre plans out what it needs to do. We can see it actually planning out JSX files and CSS files in advance with like little checkboxes for it to check off as it does the various things. And then Copilot will generate the code based on the plans that it generated, lets you do a live preview of the code that you just generated. Here's a sort of PONG example that they created in this video here. It also appears to be collaborative, so multiple people can work inside of this workspace and jump in and help generate new code and update the software that you're building. It's not available yet, but they do have a wait list over at this URL that's on the screen. So if you're a coder or you want to play around with a tool like this to let it write code for you, get on the wait list and eventually we'll get access. Now the other really big story that was Circulating all over AI, Twitter or AIX was this review of the rabbit R1. Now in last Friday's news video, I talked about how the rabbit R1 was starting to get into people's hands. But I had only seen reviews from AI nerds like me that were very biased on AI. I wanted to see a review from an actual tech reviewer is not specifically focused on AI. Someone like a Marques Brownlee. There was a lot of drama about his review of the Humane AI pin where he said it was the worst product he ever reviewed. And very similar fallout came from this video, the rabbit R1, which he titled Barely reviewable. Now he did have a rant in this video that I really, really agree with. I'm gonna play a little clip of it here.

It used to just be make the thing and then put it on sale. Now it's like put it on sale and then deliver the half baked thing and then iterate and make it better and hopefully with enough updates then it's ready and it's what we promised way back when we first started selling it. And then this, this whole period in the middle is a mess. And now these AI based products are at like the apex of this horrible trend where the thing that you get at the beginning is like borderline non functional compared to all the promises and all the features and all the things that it's supposed to maybe someday be. But you still pay full price at the beginning, which is what makes it so crazy.

And the point that he's making there, I totally agree with. That's like my point of view too. I'm kind of sick of these companies saying this product is going to be this, so buy it now. And then when they ship it they're like, well it's going to be that, but it's not that yet. It's going to be that, but you still made me pay full price for it that I think is really, really frustrating. We did see it with the Humane Pin, we saw it with the Rabbit. But one slight area of hypocrisy that I want to point out a little bit is that we also saw that with the Apple Vision Pro. And I feel like Marques for the most part has had nothing but good things to say about the Apple Vision Pro. I've got the Apple Vision Pro, it's sitting on my shelf right over there. It's still unfinished. There's still a bunch of things inside of there environments inside of there that say coming soon. As of right now, several months after I received the product, it doesn't have Netflix yet, It doesn't have YouTube yet. There's so much stuff that was sort of promised around the Apple Vision Pro that we don't have yet. So to me it's just interesting that he's making these rants about AI products specifically. But Apple's big recent release, which had the same kind of issue, maybe to a lesser degree, there was a lot of features baked into it. And the rabbit R1 seems like very minimalist right now with not a whole lot in it, but still, it's the same kind of thing. We were promised all these features, we bought the product based on the promised features, we got the product, and a lot of those promised features aren't there yet. Saying all that, I totally agree with Marques. I find that really, really frustrating in general. Another YouTuber that I watch from time to time, Penguin Zero, aka Moisturizing Critical, weighed in on this as well and pretty much had the same thoughts about it. And I also hate this mentality of it's gonna get better. You're trying to shut them down before they ever even improve on their design. Improve the design and then release it. How about that? Why not just release it when it's fully done? Why charge hundreds of dollars for basically the beta, like, oh, just to be first to the market and then you just have to pay that upfront cost and hope that it gets better down the line. I know this is probably a hot take. I don't understand why it's a hot take. I feel like I'm a tech enthusiast, I'm an early adopter, I buy a lot of this stuff early, before it comes out, while it's still in the phase one. But I feel like if you're promising features in a Keynote and then selling the product based on all of these promised features, you probably shouldn't release the product until the promised features that people bought based on are ready. I don't know. It's a hot take, but I don't know why it's a hot take, honestly. And speaking of the rabbit R, 1, more criticism came out even after the Marques Brownlee video, where people basically figured out it's just an Android app running on its own little handheld device. Here's actually a video of somebody that managed to install it on their Android, and for the most part, it has all the same features that it would have had if you had it in a little handheld device. And here's yet another video. This time somebody who managed to get all of the features to work and the interface to work on iOS. You know, a lot of the criticism about the Rabbit and the humane pin is these features could just be apps. And, well, a lot of this that's going around now is showing that they really are just apps. They just put it into a different form factor. Now, I'm not saying all this because I don't like the Rabbit. I ordered one. I'm actually still excited to get it. I'm still excited to play with it. I've really lowered my expectations based on the reviews I've seen of what the Rabbit's actually going to do. And I do kind of like the idea of getting to this sort of post smartphone world where maybe we aren't taking our phones with us with all these notifications happening all the time. We have some sort of alternate device that we communicate with that's maybe easier to communicate with and uses voice. Personally, I still think glasses like the Meta Ray Bans and things like that are a better form factor. They make a little bit more sense to me, but I am still excited to play with these toys. I kind of ordered the Rabbit knowing that I was going to play with it and it's probably going to end up in a drawer anyway. If you're sick of using Midjourney in the Discord, they finally made it available on their website. For most people, if you've generated at least 100 images with mid Journey, then you can now use the Midjourney website to generate your images. And let's be honest, if you've used Midjourney and you're paying the 10 or 20 bucks a month for it, probably generated at least 100 images by now, right? But if you head over to MidJourney.com, sign into your account, you'll see that you'll have an Imagine box up at the top where you can type whatever you want and prompt. And then we've got a settings box here where you can set your aspect ratio, stylization, weirdness, variety, standard mode, raw mode, which version, what speed you want to use. And then once you actually generate an image, you have all sorts of features over on this sidebar your vary your upscale remix, pan, zoom, et cetera, to edit the image in any way that you can do it on Discord right now, you can just do it from their website without needing Discord anymore. Again, you do need a hundred generations, but if you haven't generated 100 yet, go make a bunch real quick and then you'll get it the tool Udo, which yep, I talked to the founder. He said internally they pronounce it Udio. Everybody else has been calling it Udio, which I actually think is funner to say, but it's actually called Udio. Apparently they just rolled out an update that now allows you to make longer generations. Here's an example song that they shared on Twitter or X for It's a three minute plus song. They claim that the verse and chorus structures are more consistent now. So if you like generating AI music and you've used Yudio, well, it just got a little bit better. This week, China unveiled a new humanoid robot called Tiangong. I'm not sure if I pronounced that right, but we have another humanoid robot to keep an eye on. At some point in time. I'm going to do another video about the advancement of AI robots and humanoid robots. I made one of those videos a couple of weeks ago, and I plan on making one of those videos every few months. So we'll check back in with these robots in a future video. Now let's move on to some of the weirder, more interesting news that came out. And I'm going to kind of rapid fire some of these. We're almost done here. DARPA released this video of what appears to be this, like, insanely fast moving tank type vehicle. This is actually operated by AI. This is an autonomous vehicle called racer, which stands for Robotic Autonomy in Complex Environments and Resiliency. This really makes me wonder if the future of warfare is just going to be our robots fighting other people's robots. Who knows? When countries go to war against each other in the future, it might just be a giant, more advanced episode of BattleBots. You want to see something really weird and interesting? This is an AI camera that will actually shoot paintballs at intruders. So it watches for people that might be intruding on your property. And if they're not supposed to be there, it actually fires paintballs at them autonomously using AI and computer vision. The demo video they made is actually pretty hilarious. I have no idea why the dude on the left is so dang sweaty, but apparently he's spotting an intruder and I guess it's like someone's boyfriend or something, I don't know. It says it's not just another security camera watching over your house. This thing packs a punch with paintball and tear gas projectiles to really give intruders a surprise they won't soon forget. It uses a sophisticated computer vision technology to identify human faces and animals even in low light and distinguish between friends and family foes. Apparently you use some sort of app interface and you categorize people, and when it recognizes people that are trained into the interface, it doesn't shoot them. And if it doesn't recognize them, well, they get fired at with paintballs or tear gas. And another interesting news story from this past week, there was a teacher who was arrested because they used something like 11 labs. I think it was actually 11 labs to train another teacher's voice and then create racist rants as that other teacher teacher that framed the other teacher using 11 labs got caught and was actually arrested this past week. But this opens up some questions because the laws on this kind of stuff are still very unclear and in a gray area there isn't really much precedent yet for when this type of thing goes to trial. So we will probably be seeing some new laws and some new precedents set of how some of these AI fakes and setups are going to be treated in the future. And finally, if you made it this far, you deserve this video that I found on Instagram of this wedding couple here who's using Photoshop generative fill to try to get rid of this rock on the image. And when it generates a new image to replace the rock, it's baby. For some reason. Yeah, AI can be weird sometimes. And that's all I got for you today. Before I wrap up, I want to point out something really cool. Myself and a handful of other AI youTubers did a fun live stream earlier this week over on a channel we're calling AI Community. This is a YouTube channel that's created by about 20 different AI YouTube creators. And every other week we're going to do a live stream and it's going to be a rotating cast of AI creators. This past week we had Matthew Berman, we had Igor Pognoy from the AI Advantage, myself, Tim from Theoretically Media, Matt Vid Pro, and by Cloud. And on the next live stream, it's going to be different AI creators jumping on this live stream and we're all going to kind of rotate in and everybody's going to be on one of these live streams once a month or so. Should be really, really fun to experiment with. A link up to our most recent live stream in the description below, if you want to go check out the hour and 48 minute conversation that all of us AI nerds had together.