AI News: The AI Arms Race is Getting Insane!

So just like spring of last year, AI news is really ramping up. There has been a ton of announcements this week. I'm really having to figure out what stuff to filter down that I think you'll find important because at the end of the day nobody really cares that like Walmart got a new AI chatbot or something. So I'm going to break down the stuff that I found important, interesting or just downright fun that I think you're going to enjoy. So let's get right into it.

This week Google had their Google Cloud Next event out in Las Vegas where they made a ton of new announcements. Plenty of them related to AI or the announcements were more relevant to like enterprise and developers that are building with AI models. There were some pretty interesting and fun announcements that I think you'll enjoy as well. So I'll be kind of sprinkling them throughout this video. The real story of this week is all of the news about new large language models becoming available or soon coming available. During Google's event this week they announced that Gemini 1.5 is now available in 180 plus countries with native audio understanding system instructions, JSON mode and more. Now we have talked about Gemini 1.5 in the past, but most people haven't had access to it until now. Of course the biggest factor about Gemini that people are most impressed by is the fact that it's got a 1 million token context window now. Now one token just to refresh your memory is about 75% of a word. So 1 million tokens means that between the input that you can give the model and the output it'll give back, you have a combined 750,000 words to work with. Gemini 1.5 is now available via the API. So if you're a developer and you want to build with this model, it's now available for you. My buddy Bill of Alsadhoo here has probably one of the best examples that I've seen of somebody actually using using Gemini 1.5. He shows an example here where he says I just dropped in an audio file of an hour long video interview and now it's helping me package it for YouTube. We can see in his screenshots here he actually uploaded the MP3 file, told it to analyze this audio recording for his interview. He then asked it to give the key takeaways and come up with 10 high click through rate YouTube titles based on the principles of Daryl eves and top YouTube creators. Keep each title to 50 characters or less. So from this audio file it then gave them the key takeaways. He also offered it two thumbnails and said which of these thumbnails is better suited for this YouTube video. It analyzed the two thumbnails and gave feedback on which thumbnail to use. And then once he picked the thumbnail, he had it suggest the 10 titles and it gave some pretty decent titles. Now, that's all pretty cool. We can kind of do that with Claude right now. The difference is he uploaded an audio file and it did this from the audio file. With Claude, you would actually have to get the text transcript, upload it, and you would pretty much get the same end result. Most likely where I thought this was the most impressive was when he asked it to generate timestamps and you can see it actually recommended these timestamps with an explanation of each of the timestamps and then there is a shorter version of the timestamps that it also generated. This to me is really impressive because I've tried to use Claude and I've tried to use ChatGPT to generate timestamps for these videos that you're watching right now. While it gets the various sections right and it knows what I'm talking about in the video, it just completely gets all the timings wrong and struggles to give an accurate timestamp to the actual chapter. This example that Billable shared is really, really useful in my opinion. In most of the past scenarios, OpenAI's announcement really overshadowed Google's announcement, but this time we just got API and rolling out inside of ChatGPT. We don't have a ton more details than that, but if we take a look here at the OpenAI documentation, we can see we've got the newest model here, GPT4 Turbo, and this is the April 9th edition of it. Vision Request can now use JSON mode and function calling. It's got the same 128,000 tokens we've been working with, and it's updated through December 2023, which is also what the previous model was updated through. Supposedly this new model is a lot better at coding and also a lot better at math. And for a while, Quad three Opus was the cream of the crop, the best model out there. But it seems now that according to the Chatbot arena here, the newest version of GPT4 Turbo, the April 9th edition, now took over Claude 3 Opus again as the strongest, most powerful model, as voted on by the people that rank this system here. Now, they do show Gemini Pro down here as being below Claude and GPT4 Turbo, but I don't believe I'm not 100% sure, but I don't believe that this is taking into account the newest 1.5 model. That's not the only news we've gotten. In the world of new large language models, the open source world is continuing to heat up as well. In fact, stability AI released stableLM2, which is a 12 billion parameter model and according to most benchmarks it just underperforms the mixtral 8x7B model. And although they kind of make it out to be like an open source product, it does say it can be used non commercially as well as commercially. But if you have a Stability AI membership. So if you do want to use it commercially, you got to pay. Which doesn't feel very open source to me. Well, Mistral said Hold my beer and watch this Releasing a new large language model using the mixture of experts architecture, but they released it in an interesting way. They released it as a torrent link directly on X with almost no context. In order to actually download the weights for this model you would need a torrent downloader, something like Views and then if you paste this URL into your address bar, it will start the download over inside of your torrent downloader. However, be aware it is a 281 gigabyte file. Now I don't have a ton of information about this new model from Mistral, however in the rundown newsletter they gave us a little bit more details this week. According to Rowan over at the rundown, the new model is Mixdril 8x22B, so the previous model was 8x7B. So it was eight separate models that the router called upon to get the prompt answered. And each of the models that it was calling upon was a 7 billion parameter model each. This new one has eight experts, but each expert, instead of being a 7 billion parameter model, is now a 22 billion parameter model. So it was just trained on a lot more data. According to this breakdown, the new model features a 65,000 token context window and a combined total of 176 billion parameters. I haven't used this model yet myself. Channels like Matthew Berman's channel does a really, really good job of testing these large language models. I have a feeling that this one is going to be the strongest open source model once some more of the tests come out around it, but we have even more large language model news. Even more coming from Google. Google released Gemini 1.5, their closed source model, but they also rolled out new versions of Gemma, which is Google's open source large language models. These two new models are code Gemma, a model fine tuned for using with coding, and recurrent Gemma, which is designed for more efficient research purposes. Now all of the articles and resources I make in this video I will share in the description below. So if you do want to dive deeper into how these Gemma models compare on the benchmarks against other models, check out the links below. But overall, Gemma appears to be pretty on par with the other open source coding specific large language models. And in the final bit of large language model news, we also learned this week that Meta is really close to releasing Llama 3. Llama 3 is expected to be roughly as good as GPT4, but open source and made publicly available for anybody to use and fine tune and build on top of. Now they did say with llama 3 they were going to release several different versions of the model. It kind of sounds like what Claude did with Haiku, Sonnet and Opus. There will be different models sort of more fine tuned for different reasons. That's my takeaway from this as well. According to TechCrunch, they announced that this is coming out sometime within the next month, hopefully sooner. So really looking forward to that because I truly am rooting for both the open source and the closed source side to keep on pushing the boundaries and each side is just making things better and better for us, the consumers. Now I know a lot of people that watch this channel have small to medium businesses and they're looking to use AI to improve their marketing in their business. That's why for this video I partnered with HubSpot so that I can get your eyeballs on their completely free free report all about how AI is completely redefining startup Go to Market strategy. I'm going to put the link in the description to where you can download this report, but this is a must read for any small to medium business that plans on using AI as part of their startup strategy. You're going to learn about the various strategies that startups are using to bring their products to market. You'll also learn about the most popular AI tools and best practices for scaling and and if you know me, I love me some AI tools. The free report also covers how AI is driving startup scalability and drawing the attention of investors, as well as the future of AI within Go to Market strategies. Of course, my favorite section as a guy who made a business around curating amazing tools is the section all about the tools and best practices that HubSpot recommends for your Go to Market strategy. Again, this free report was provided by HubSpot who is sponsoring this video so thank you so much again to HubSpot for sponsoring this. And once again the link should be right at the top of the description to make it easy for you to find. Now if the main story of the week was all of the large language models that have come out this week, the sort of B story of the week, the side story is it seems that all of these companies that are building these large language models are are all trying to release the reliance on Nvidia GPUs. At the moment, Nvidia owns the market on GPUs being trained for AI. But Google, intel and even Meta are all trying to bring that chip generation in house and stop giving so much money to Nvidia. At the Google Cloud Next event this week, Google introduced their Axion processors, Intel introduced their Gaudi 3AI chip, which is apparently a 40% better power efficiency than Nvidia's H100 GPUs. And Meta announced a chip as well. This new chip is called an MTIA or Meta Training and Inference Accelerator. This is the second generation of the chip and according to Meta's article here it is three times improved performance over the first gen chip. Now again I will link to all these articles that talk about all of these chips that were announced this week because a lot of this sort of technical stuff here is a little bit over my head. I don't personally understand how these chips work, but if you're somebody that wants to deep dive and truly understand what makes these chips better than what's available, feel free to read the articles in the description. However, back at Nvidia GTC earlier this year Nvidia announced their next iteration, the Nvidia Blackwell, which is supposedly already four times more powerful than than the H1 hundreds. Again, the chip that is sort of the industry standard for training AI models right now. So while all these companies are making their own chips to release Reliance on Nvidia, Nvidia is still way out ahead with their latest generation of chips, making it extremely hard for these other companies to catch up with the compute power that Nvidia is producing right now. Also during the Google Next event, Google revealed Imagen 2. This is sort of like Google's answer to OpenAI's Dall E or Adobe's Firefly. It's their own internal AI image generation model. However, what makes Imagen 2 different than tools like Dall E and Firefly is that it can actually generate animations, it can generate GIF files or GIF files, however you like to say it. Here's some Examples of the types of animations that it will make, you know, similar to what we get out of like Pika or Runway. But these are very short 2, 3 second clips and they seem to be designed to just make like short loops or like little gif GIF files. They described it as text to live image. So if you use an iPhone, you've got that live photo feature where when you take a photo it sort of saves like two seconds of video as well so that you can find that right spot in the photo. It seemingly is designed to generate that kind of little teeny short clip. That wasn't the only video announcement that they made at this event. They also revealed Google vids. Now this is something that we don't have a lot of information about. They put out a little like teaser video. It's a minute and 27 seconds here. And it appears to make videos that look almost like PowerPoint style videos. Using AI says let's choose a style and then you can see the styles of the videos almost look like you know, something you'd get out of Canva or PowerPoint or Keynote or something like that. You pick your style, you can give it a script or let AI generate a script and then it creates like a slide presentation style video. Very reminiscent of like a PowerPoint or Google Slides video. Again, not available yet, but it does say coming soon to Gemini for Google Workspace. But here's some research that came out this week about a new video generator. It's called Magic Time and it makes time lapse videos. This one is very specific to time lapses. So you can see some differences here. Where a prompt given to a normal animated video like bean sprouts grow and mature from seeds generates something like this. Where this Magic Time style generates this like time lapse version or construction in a Minecraft virtual environment. This one shows almost like a drone rotating view around it where Magic Time is making this time lapse. And you can see a couple other examples here. Again, it makes a very specific type of video which is amazing for people like me who love to make videos because I can use stuff like this for B roll. Their GitHub page, which will be linked below, has a lot of examples of the type of animations it can do. But the coolest part is the code is available on GitHub, so if you know what you're doing, you could run it locally or on a cloud. But they also have a hugging face demo that you can play around with right now. But for now, let's take a peek at One of their, like, cached models here. So I'll click on this. And this was a prompt of cherry blossoms transitioning from tightly closed buds, et cetera. And if I play it, that's what that looks like. And since we're talking about video, you remember last week I talked about how Neil Mohan, the CEO of YouTube, said that if OpenAI trained on their videos, that would be a clear violation of their policies, their terms. Well, according to a report from the New York Times, OpenAI transcribed over a million hours of YouTube videos to actually train GPT4. Now, there hasn't been a ton of confirmation. If you read the article, it's a lot of hearsay. And Google claims to have looked at the robots.txt file on YouTube and saw that OpenAI probably was scraping data, but there was no real, actual confirmation. But pretty soon we could have a law that forces the AI companies to to reveal what the models were actually trained on. So there was a new bill introduced to the Congress on Tuesday that intends to force artificial intelligence companies to reveal the copyrighted material they use to make their generative AI models. The bill would actually force companies to file a report about what copyrighted material they used at least 30 days before actually releasing that AI model. No clue whether or not something like this will get passed, but seeing as how some of the biggest companies in the world, namely Google, Microsoft and Meta, may not want to reveal what data they actually trained on. And being as powerful as they are, I would imagine there could be some lobbying going on behind the scenes to keep bills like this from getting passed. But I'm just speculating here. Adobe, on the other hand, is taking a completely different approach. They're actually willing to buy data off of creators to train on for their large language models. Adobe wants to create their own version of Sora, and in order to do that, they need a large amount of video training data. Well, in order to get that video training data, Adobe is offering to pay between $3 per minute and $7 per minute to purchase video content off of creators. Adobe is looking for everyday things like people riding a bike or walking down the street or the types of things you'd see in normal stock video. So if you like to go out into the world and film stock video footage, you might be able to make some extra money by submitting it to Google and letting them pay you a few bucks per minute of video. I think if a bill like we just looked at does end up getting passed, this is probably the way of the future for how these large language models will get trained. Meta also announced this week that they're going to take stronger measures to identify AI generated photos. Basically, when people upload photos to places like Facebook, Instagram or Threads, Meta is going to use their own AI to test detector to look for things that are commonly present in AI photos and try to identify when the photo or image was made with AI. Previously, Meta was able to label photos that was made with their own EMU AI image generator as well as identify AI photos that the uploader specifically marked as AI. But now it appears that they're actually going to use AI to try to spot AI. Another big story of the week was Udio. Udio is a really, really good AI music generator. You give it a prompt, you can suggest styles, you can decide whether you want to write your own lyrics, have AI generate lyrics, omit lyrics altogether, and the outputs are really good. Like a lot of songs I've heard, I would not have known that they were AI unless somebody told me that they were AI. They are that good. Now I'm not gonna go too deep into Udio in this video because the video that I released the the day before this video went live, was all about AI music generators with a very heavy focus on Udio. So if you really want to learn about Udio, check out this video here. But one really interesting thing about Udio before I move on is that it's actually being supported by other musicians, musicians like Will, I Am and Common. And it has backing from people like the co founder and CTO of Instagram and the head of Gemini at Google and and A16 is backing it. So lots of money getting pumped into this and musicians actually seem to be supporting this platform. So something really interesting to watch here. And since we're on the topic of AI music, Spotify is rolling out a new feature where you can have it generate a playlist for you using AI. You just give it a prompt like I need pump up music for the gym and it'll make a playlist of songs it thinks is pump up music for the gym. Or I need sad music because I'm painting a sad picture and I want to hear sad music and it will generate a list of sad music for you. Nothing super, mind blowing or groundbreaking, but I thought I'd share since we're on the topic of music anyway. Now let's talk about AGI for a second. Elon Musk thinks it's coming within the next year. In an interview on an Xbase, Elon Musk said that artificial intelligence will be smarter than the smartest human probably by next year or by 2026. So sometime within the next year and a half, he believes we're going to essentially have AGI. Yann LeCun, one of the godfathers of AI and Meta's cheap AI scientist, thinks otherwise. He believes that large language models will never actually reach human level intelligence. In his article, he doesn't actually say that he doesn't believe that AI can't reach human level intelligence. He just doesn't believe that large language models, the current standard that we're using for AI right now, isn't going to reach human level intelligence. Back in February we talked about how Meta is working on V jepa. This is Meta's new architecture for self supervised learning. This is the technology that Yann Lecun actually believes will at some point hit human level intelligence. This week the much anticipated Humane Pin started getting into the hands of consumers. If you don't remember, it's a little PIN device that goes on your shirt. It has a little projector that can project things onto your hand. It has a camera, it has a microphone, so it listens to voice commands. And it's designed to be sort of a replacement for a smartphone. And so far the reviews haven't been super favorable. Here's a quick recap. As of right now, the Humane PIN is an incredibly poor proposition. So no, I can't recommend the AI PIN in the form in which I received. I can't imagine a world or a use case where someone would prefer this over what already exists. The PIN is not worth the money. Not yet and probably not anytime soon. You definitely should not buy it. Planning to replace your phone now? The biggest complaints about the Humane PIN were that it really doesn't do anything beneficial over your smartphone. People complained about holding up your hand like this to get through menus and stuff actually gets fairly tiring on your hands. They also said that in bright light you can barely see the projection, so it's hard to use. People complained that the gestures were confusing and complicated. There's no privacy. If you are actually trying to prompt it. You have to speak out your prompts so in public it just feels awkward to be talking to a little computer on your chest. But the biggest complaint that everybody made about it was that it is a $700 product with a $24 per month fee. And then if you ever cancel that $24 per month fee, a product just stops working. It's a paperweight saying that. Every single person that talked about the Humane AI pin said that they Thought that the technology was really, really cool, just not very practical or usable yet, and that it still might get to something that people actually want that could replace a phone. It's just nowhere close yet. Finally, I thought this one was a kind of fun article about somebody making a lot of money using AI. A card game developer paid an AI artist $90,000 to generate card art because no one was able to come close to the quality that he was delivering. Now, I have a little bit of an issue with them saying they're an AI artist, because I think the better term is AI, AI assisted artist. I don't actually know if, like an AI artist is a thing, because if you're just pressing a button and letting AI generate the output, I don't really feel like you are much of the artist. However, I do really appreciate AI assisted art where these cards here. Somebody generated the images for every single card using AI, but then they went and touched it up using Photoshop or whatever image editing tool they use to make sure that the colors, the consistency of the characters, all the style match, there wasn't extra fingers. All the weird stuff that we get with AI, they actually had to manually go in and fix. So AI generated the initial sort of rough draft of the image, but then the artist actually then went and made it into what the card designer wanted. AI helped them make a lot more images at scale, but the artist still had to do work on all of the images. So it's more AI assisted art. Not really AI generated art. I don't know. That's a little bit of a soapbox. I'll step off it now. I just thought it was really cool that somebody was using an AI art generator to make the initial concept. Then they cleaned it up with Photoshop. They did this at scale and earned $90,000 from the company that hired them to do this. And finally, the last announcement. This is more of a selfish announcement, but our podcast, the Next Wave, is officially launched. You can find the first episode that we did with our event from Perplexity. It is available on YouTube, Spotify, Apple Podcasts, anywhere you listen to podcasts. I highly recommend the YouTube version because the editors did an amazing job of putting some cool graphics and overlays on it. It's a really fun video to watch, but if you're driving, going for a workout, doing whatever you do, and you like to listen to audio podcasts, it is available in pure audio form as well. HubSpot, who, yes, is the sponsor of today's video, but is also the producer of this podcast and is doing a really cool competition. They're giving away Apple Vision Pros and all sorts of cool stuff for subscribing, liking, reviewing, doing all of that kind of stuff with the podcast. I don't have all the details on the competition yet. Make sure you subscribe to the podcast and like it and maybe leave a review and the good people at HubSpot could potentially hook you up. It's not to mention I might be a little biased, but I also think the podcast is really good. With videos like this, I spend a very short amount of time per topic and with something like the Next Wave podcast, me and my co host Nathan Lands, we get to deep dive, have longer form conversations, talk more about the ethics and the implications and the long term timelines that we see with this technology. So it really gives us that platform to go much longer, much deeper and bring on really amazing guests who could help us and you better understand the stuff that we're talking about in this crazy, fast paced AI world that we're in. That's my pitch for the podcast. Check it out. I'll make sure again, it's linked up in the description. It's called the Next Wave Podcast.