AI News: Sam Altman Reveals 2025 AI Roadmap

Well, it's Christmas week, which means there's not typically a lot of AI news this week. However, I record these videos on Thursdays so that I can publish them on Fridays, meaning I usually miss the Friday news from the previous week. And, well, last Friday was quite a big day in the world of AI. On the final day of OpenAI's 12 days of announcements, they announced their latest model called O3. Now, most people don't have access to this model yet. I don't have access to this model yet, but they did show it off and show what it was capable of and show some of the benchmarks. They showed that in software engineering, it was 71.7% accurate compared to 01, which was less than 50% accurate in competition code. It also outscored 01 Preview and 01. They don't even put the other models from other companies on their comparisons anymore because quite honestly, these models are still so far ahead of those other models when it comes to logic and reasoning. When it comes to competition math, it got a 96.7% accuracy compared to the OH1's 83.3. In PhD level science, it scored an 87.7% accuracy versus OH1's 78%. In research math, it got a 25.2% versus. The previous state of the art model is only getting a 2%. Now, this research math is like insanely complicated math problems like these are problems that take multiple mathematicians working together for, you know, days or weeks to solve them. It is like some of the world's most difficult math problems. This OpenAI03 is able to solve them at a 25.2% rate. Now we do see a dark blue line and a light blue line here. The dark blue line means it was able to do these in a single shot, meaning sort of one prompt and it got the answer right. The light blue from my understanding, is the multi shot, meaning that it prompted, it wasn't right. Prompted. It again wasn't right. Prompted it wasn't right. And it kept on working at it until it got it right. That's my understanding. And then you have what they call the ARC AGI benchmark. And it shows the model these sort of visual puzzles. This one on the left is an example. You've got three light blue blocks here and then on the right, the AI would be expected to put a dark blue block in these areas to complete the pattern. So this was an example that it would give the AI and then it would tell it to fill out this one. Obviously us as Humans, we can figure it out pretty easily. We would just put these light blue blocks in the same spot on the right grid with dark blue blocks to complete the square. Notoriously, AI systems have been really, really bad at this kind of test. And OpenAI's 03 is one of the first ones that has actually been able to do this test pretty well. Here's some more examples. We see like a yellow box here with two red dots in it. And then the output shows that we have a border around the box that is the width of the amount of dots in it. So it is a two square width around the box. Same with this one. We have three blue squares inside the yellow box. So we'd get a three square thick border around the result. And we can see a handful of examples here. And then the AI would be asked to complete this final one based on the examples that it saw prior to it. On this benchmark test, the ARC AGI benchmark test, the previous 01 models all scored, you know, pretty low on it. The 01 mini model got an 8%. The 01 high model, the best of the 01 models, got a 32% on it. Well, this new O3 series got a 75.7% on the lower compute per task model and an 87.5% on the higher compute model. Now, for context, I read that a human would typically get around a 76% on this test. So this high model outperforms most humans on this test. And the low compute model actually performs about what a human would perform on this test. Now, one thing that they sort of hid in this video, but you can actually see it over on the ARC AGI website, is that if we look at this same chart here, it's pretty much the same information. They rounded a little bit. So this is 76%, this is 88% instead of 75.7 and 87.5. But it's pretty much the same chart. But we can see down here that instead of compute per task, it shows cost per task. And one interesting thing about this chart is that it is a logarithmic scale. So this axis right here is $1, this one is 10, this one is 100, this one is 1000. And if we're continuing along that same line, then if there was another axis right here, it would be $10,000. So the amount of compute it took to actually get a 76% on the ARC AGI benchmark was somewhere around maybe $30 per task to get there. Right. Remember, this isn't a linear scale. So this is probably somewhere around 25, $30 here on this graph. Now, if we look at the high compute model, remember that this line that would be right here on the side would be roughly $10,000 to get this. So to get this 88%, it is on this logarithmic scale between 1000 and 10,000, meaning that this was probably in the range of 5000 to $6000 per task to get this outcome. All that being said, it is really, really, really expensive to get.03 to pass these marks right now. These aren't at levels where the normal consumer would just get access to these tools and get to prompt to their heart's desire to actually get these kinds of outcomes. They are still insanely expensive to run. I've also seen a lot of videos on YouTube and a lot of people on X claiming that this O3 is AGI and we have reached AGI. I guess that's really up for interpretation in my mind. I don't see this as us reaching AGI. I feel like there's still a lot of tasks that humans would do better than in AI. But I do believe that this was a significant leap forward getting us closer to AGI. But if we actually want like a practical AGI that's usable by the most amount of people, they need to figure out how to get these compute costs way down. Because right now it is actually really insane to use this O3 model. Now, they did say that the O3 mini model, they're expecting to release it early 2025 and the larger O3 model will be a little bit later than that. But obviously the costs have to come down quite a bit or OpenAI's got crank up the price and add an even more expensive pricing tier to make it economical for them. And since we're talking about OpenAI and the economics of IT and AGI, this article came out on the Information this week on December 26th about how OpenAI and Microsoft have their own internal definition of AGI. Now, if you're not familiar with the deal between Microsoft and OpenAI, when Microsoft invested in OpenAI as part of the agreement, they decided that if AGI is achieved by OpenAI, then the deal with Microsoft and OpenAI sort of changes, restructures and Microsoft has a lot less control over OpenAI. And OpenAI is free to go and do what it wants once it reaches AGI. However, anybody that's been following AI for a while knows that it is very complicated to decide what actually constitutes AGI. Different companies have different definitions the goalposts feel like they keep on moving. Nobody really has like this general consensus on AGI. Apparently Microsoft and OpenAI put a financial number on it. We can see here in this article. Without a new deal, Microsoft will not be able to use OpenAI technology developed after reaching that point, AGI that's led to speculation that OpenAI has threatened to declare it has achieved AGI to get out of its obligations to Microsoft. And this is their target here, $100 billion. And right now this is sort of rumor, speculation. This is kind of like a leak, but it can't be 100% confirmed. But apparently the deal between Microsoft and OpenAI stated that AGI would be achieved only when OpenAI has developed systems that have the capability to generate the maximum total profits to which its earliest investors, including Microsoft, are entitled. According to documents OpenAI distributed to investors. Those profits total about $100 billion, the document showed. Now, the document still leaves some things open to interpretation. And the declaration of sufficient AGI is in the reasonable discretion of the Board of OpenAI. So they may still have differing views, but apparently they'll consider AGI achieved once AGI is capable of generating $100 billion. And for context, those kinds of profits are not happening anytime soon. OpenAI currently loses billions of dollars a year, and around September this year it told potential investors it didn't anticipate turning its first annual profit until 2029. So if anybody is expecting OpenAI and Microsoft to sort of get divorced anytime soon, it doesn't sound like that's going to happen. If you are a fan of OpenAI's previous generation O1 model, which is still their current sort of state of the art model that we have access to, but after seeing O3, it now almost feels like an older model. Now you can use it inside of GitHub Copilot, which as we announced last week now actually has a free tier. So if you Want to code with 01, you can now use GitHub's copilot to do that. Now, it does say in this article that 01 is included in your paid subscription to GitHub Copilot, allowing up to 10 messages every 12 hours. So if you are on their new free plan, it's unlikely you can use 01. But if you're on their paid plan, which start at $10 a month, I believe you now get access to 01 to help you with your coding. And in the final bit of OpenAI news, Sam Altman took to X on Christmas Eve to ask, what would you like OpenAI to build slash fix in 2025 and I'm going to kind of pay attention to some of the answers that Sam Altman actually commented on. Like McKay Wrigley here said the vector store you guys have for Assistant API is really good. Make that its own retrieval API would instantly become the top retrieval product. Video input modality would be nice. You're already doing agents but excited for these unexpected let's see a hardware play. And then Sam Altman gave it some eyes, making me think that they're probably thinking about all this stuff. In fact, another recent article on the information claims that OpenAI is getting back into robotics and about to start building their own robots soon. Pliny the Liberator said, lose the guardrails, it's cleaner. Sam Altman said, definitely needs some sort of grown up mode, so maybe we'll see some ability to remove guardrails. On the flip side of that, Jace here said, family accounts. Let me create accounts for my kids with guardrails. Let their curiosity take off, but within reasonable limits as determined by the parent. Maybe we can even get insights about our kids from their usage. Sam said, that's a good idea. So sort of two opposite ends of the spectrum, but would love to see both of Those built in OpenAI especially as someone who has kids who are very fascinated by AI as well. Leo says a strong 4o replacement GPT O Sora all seamlessly integrated into ChatGPT, a 50 to 70 buck plan that's like more of a middle ground between their 20 and $200 plan. A longer context knowledge cut off aggressive updates please. And Sam Altman said hopefully he'll be quite happy with us next year overall, obviously making us think we're probably gonna see a lot of this stuff next year. When I talk to Advanced Voice mode, I wish it had better memory of my previous conversations, both verbal and text, even if it were just a rag system. I enjoy conversing and brainstorming during my commute, but I find myself having to rehash things in great detail and Sam said I really want to figure this one out. I think that would be great as well. That's something that when I got to play around with Project Astra, that does really well. It actually remembers previous conversations where Advanced Voice mode only kind of knows what you put into its memory specifically and what you talked about in your current conversation. Better turn detection for voice chat. It sucks that you can't just silently think for five seconds without being interrupted. Good point. That is another thing I think Astra is really good at. If you're just kind of silent for a few minutes it lets you think. It doesn't try to ask you oh are you still there or anything like that. Honestly a good non thinking model. GPT4O kind of sucks right now. Even Mini isn't that good. I think having both is important, Sam said. Definitely do a deep research feature like Gemini but better, he said. KK release GPT4O image generation support Hope to Sora please. Sam would like to see more adherence to image and text prompts. Also a more reasonable content restriction policy. Lots of SORA improvements coming, he said. A drag and drop UI where multiple separate chat models can be strung together to generate and refined outputs, he said. This is an interesting idea I haven't really thought about before. Maybe we'll see it and I'll end on this one. Sam Sheffer says AGI Sam Altman gave him the salute so just a little sneak peek of what we can expect from OpenAI and ChatGPT in 2025. Really cool that Sam did that and actually responded to people. Usually Sam is more of a post and ghoster. He puts a post up on X and then doesn't respond to anybody. This time around it looks like he's engaging with most people and kind of confirming that a lot of the cool things that we want to see coming out of OpenAI they're probably already working on. All right, like I mentioned, there wasn't a whole lot of huge news this week. The O3 from last week was a big deal, and Sam getting on and sort of giving us hints to what's coming seemed pretty cool to talk about. Other than that, we have a lot of smaller announcements that I want to cover really quickly, like the fact that XAI just raised $6 billion in a Series C from A16Z, BlackRock, Fidelity Management and research company Kingdom holdings, and, you know, a bunch of other venture capital firms. I do feel like X is going to be a much bigger player in 2025 in the AI world. I think burying Grok inside of X.com was not their best move, but it does sound like they're going to break Grok out and sort of do some more standalone products outside of the X platform next year. Case in point, this TechCrunch article claims that Xai is testing a standalone iOS app for its Grok Chatbot. It's actually available in the App Store right now if you're in Australia, but doesn't appear to be available in countries yet. So I guess if the Aussies like it Then the rest of us might get it as well Soon. There's a new open source leader in the world of large language models in the form of deep seq v3. It can generate at 60 tokens per second and it's fully open source. And we can take a peek at some of the benchmarks here. Between some of the other open source models and some of the closed models, we have Claude 3.5, Sonnet and GPT4O here. And this new deep seq v3 open source model beats out almost all of these other models in including the closed models in almost every single benchmark, including English, code, math and Chinese. It's got 671 billion parameters using a mixture of expert model and it appears to be both open source for anybody to sort of download and use and also available via an API that you can use if you don't want to use the open source version. And one thing that's interesting about this model, and I believe my buddy Bill of all pointed it out, he said China just dropped an open source AI model that absolutely mogs using a fraction of the compute US labs burn through US Restricting China's chip access will slow them down. China, cute. Watch us flex with a $5 million training run while you need 150 million. So the US is actually restricting chips and GPUs from getting to China and China is figuring out how to work within the constraints and still building really, really good models. All right, moving on to Google real quick. Apparently Google search is going to have a dedicated AI mode. We don't exactly know what this looks like yet, but seemingly you'll have the to kind of search the traditional way or switch to AI mode and get AI assisted search. That's sort of what I gathered from this article, but we'll learn more as more details come out about it because this is sort of just rumors and leaks right now. Here's another thing that happened at the end of last week but didn't make last Friday's video because it actually happened on Friday. And that's that LTX Studio just made a big update to their LTX video model which is an open source video generation model. Now in full disclosure, LTX has sponsored past videos and past tweets and things like that. They're not sponsoring this video, but I do want to shout out the news because it did happen within my time window of making this news video last Friday. So the text to video and image to video workflows now deliver even better results. Advanced training and new data ensure smoother more polished motion, they've eliminated flickering and jittering effects. With a new VAE decoder, videos now look cleaner and with no distracting artifacts. Image to video is now optimized, creating more natural and seamless results that make it easier to bring your ideas to life with precision and creativity. And there'll be building it into LTX Studio, their front end app pretty soon. But again it is open source, so if you've got a strong enough computer or you're good at putting this kind of stuff on cloud GPUs, you can use this right now and generate videos with it with pretty much no restrictions. In other video news, Viggle AI released kind of a fun novelty app inside of their platform where you can upload an image and turn it into a wrap. Here's an example. Santa Rapa kids they nice never known.

It coming through your chimney like a holiday party.

You can actually use this for free. Head on over to Viggle AI, click on Try on web and then after logging in you can see there is a rap feature over on the left. I could drop in a character, choose a background here and then for Describe My Song, let's just do a rap about making YouTube videos about AI news. We'll click create and it's actually going to use the UDO platform to generate this song. And here's what we got out of that.

Stories fly like headlines Tech fame in the cyber world I stake my claim Yo, I'm the boss in the the AI game Spitting facts every vids never lame Stories fly like headlines Tech fame in the cyber world I stake my claim.

I mean it's not bad, but I think a lot of the words in there were just gibberish. I have no idea what it was saying, but that's what I got out of it. Again, it's free to play with. You got nothing to lose over at Viggle AI. I'm sure with a couple more rerolls or some different images or some different prompts, I could get a video a little more impressive, but I don't know what language this one was speaking. Another tool that I came across recently that I thought was pretty fun to play with was this backflip tool. We actually played around a little bit on my recent livestream. I do them every Mondays at 11am Pacific on this channel. This was one of the ones we played with and it's really cool. You can make 3D objects that are 3D printable or you could pull them into Blender or you could pull them into Unreal Engine. These are the types of characters that you can create. And if you wanted, you can actually export it as an STL file for 3D printing. On my recent live stream, we exported it and pulled it into mixamo and actually made our character dance. If I look at my generations here, you can see I was trying to create some different, like wolf characters and some different orcs. And we actually took one of these orcs into mixamo and actually had it doing like the twist for us and stuff. It was pretty fun. And it works by going to part creator, entering a prompt like an eight legged alien creature. And then we have a handful of presets here. I'm gonna use the character preset and then we can generate images. And it's gonna start by giving us four starting images. And then once we pick one of the starting images, like this one right here, we can click on this button to make it 3D. And in about a minute we have four different 3D objects to choose from that follow that original model that we created. And these are all pretty awesomely creepy. And you can see we've got this 3D object now. And if I come up here to download, I can download it as an STL, a 3D, print it, an OBJ, a GLB, a PLY, and you can pretty much do whatever you want with this 3D object that you just generated. And it's pretty fast and it's pretty accurate. So that's backflip AI and I'm still on a free plan. Everything I just showed you I did for free. Everything I did on the live stream, I did for free. All right, moving right along here, Arizona announced that it's getting an online charter school taught entirely by AI. So the newest online only school was greenlighted by the Arizona State Board for charter schools, and it will be taught entirely by AI. Academic instruction is whittled down to just two hours. The remainder of the student's day will include life skills workshops covering areas such as critical thinking, creative problem solving, financial literacy, public speaking, goal setting, and entrepreneurship. The online only school targets students from fourth to eighth grade. This one should be pretty interesting. Now there will be a human in the loop, so they will actually be able to talk with teachers as well. But for the most part, they'll be working with AI back and forth to learn what they're trying to learn. We got a preview of the new Asus AI PC. It almost looks like a Mac mini, but it's designed for AI inference. It's got an Intel ARC GPU in it. I Have no idea how this will do with games, but it should do pretty good for whatever your AI inference needs are. And it's gonna have like the whole Microsoft Copilot stuff built into it as well. In other hardware news, apparently the Ray Ban meta glasses are going to get a display sometime in 2025. I got to demo some of the Astra glasses where you wear, they feel like you're wearing the Ray Bans that can talk to you and see what's going on around you, but they also have a little heads up display that you can see in one eye while you're moving around. It sounds like the meta Ray Bans are going to get that as well, which is really exciting. I would love to see, you know, subtitles of somebody speaking in a different language, have it subtitle them in English or you know, show me map directions or things like that right in front of my eye as I'm trying to navigate. So I'm excited to see that in the meta Ray Bans because I already love my meta Ray Bans and I will probably buy a new pair if they add that feature. And finally, in robotic news, China unveiled ultra acrobatic humanoid robot with a 320 degree waist spin power. So here's a look at what that robot looks like and one of the big selling points is that its gait, the way it walks, feels a lot more human. We can see in some of this video what it looks like when it's walking around. If you ever watch Battlestar Galactica, it reminds me of a Cylon if I'm being honest, which is not a good thing, but pretty cool. I always love ending on robots and robotics. And this is another company Engine AI that we'll have to pay attention to and see how their robot efforts go. And that's what I got for you this week. Again, it was a week where we had Christmas and not a lot of AI announcements, but we did have 01 to talk about. We did have Sam Altman getting feedback on where ChatGPT should go and we had a bunch of smaller announcements from a lot of companies. And now you should be pretty looped in on what happened the week of Christmas. While you probably weren't paying attention to AI, I still was. So I hope you enjoyed this video. If you like staying in the loop on AI news, check out FutureTools IO. I do keep the news page up to date on a daily basis with all the news that I even can't include in the videos because there's too much of it. And I also share all of the coolest AI tools that I come across right here on the homepage. And there's a free newsletter where I will send you just the coolest tools and the most important news directly to your inbox box. It's totally free. When you sign up I'll hook you up with the AI Income Database, a database of cool ways to make side hustle money using various AI tools. It's all completely free over at FutureTools IO. And again, if you enjoyed this video, give it a thumbs up and subscribe to this channel and I'll make sure more cool AI and tech videos show up in your YouTube feed. I really really appreciate you tuning in this week and I hope you had a happy holiday, a merry Christmas, all of that good stuff. Really really appreciate you taking the time time and spending it with me watching this video. Really really appreciate you. I'll see you in the next.