AI Shocks Again: DeepSeek R2, Grok 3, OpenAI GPT 4.5, Helix Humanoid, Tülu 3... (February AI News)

This month's AI breakthroughs have been remarkable, with OpenAI releasing GPT 4.5, while Deep Seek's upcoming R2 promises efficiency levels that could leave OpenAI struggling to keep up. Meanwhile, Tesla, Optimus, Unitree and Figure's latest humanoid robots are blurring the line between machine and life, sparking shock across the Internet. Xai has just unveiled Grok3.3, rumored to be the most powerful AI yet. And China's relentless push with Tolo3 and ByteDance's new model is making waves in ways no one expected. And if that wasn't enough, the rise of AI humanoid soldiers is raising serious questions about where this technology is headed. The AI race is accelerating, and we're witnessing its most intense phase yet. One more thing before we dive in. I know many of you have been asking how to create an AI avatar like this. Well, I've got something for you. We just launched AI Content Lab, a school community where you'll learn how to build ultra realistic AI avatars and influencers, automate video creation, and monetize AI generated content. And this is just the beginning. Every month, we're adding new courses, tutorials, and strategies to keep you at the cutting edge of AI powered content. No complicated setups, no prior experience needed. Just plug into the future of AI content creation. If that sounds like something you'd want to be part of, check out the link in the description and join us inside AI Content Lab. Now let's jump back into this month's biggest AI breakthroughs. China just dropped a bombshell in robotics. Humanoid robots dancing at the Spring Festival Gala, perfectly in sync with human performers. Meanwhile, figure AI just walked away from OpenAI to build its own in house AI Tesla's Optimus is facing a new challenger in the robot hand game. And Nvidia is training humanoids to move like pro athletes. The race for the most advanced AI powered humanoid is heating up fast, and things are getting intense. Let's break it all down. First up, let's chat about China's Spring Festival Gala, where a group of 16 humanoid robots from a company called Unitree took the stage. They performed this traditional Yang ko dance alongside 16 human dancers, tossing and catching handkerchiefs, spinning around in sync and not missing a beat. The crazy part is that most humanoid robots out there struggle to stay balanced if you just give them a little shove. But these H Hun robots, they were not only dancing, but also flipping handkerchiefs in the air and catching them again, all while maintaining stability. That is no small feat. Now people are comparing them to Tesla's Optimus Robot if you remember, Optimus had some pretty shaky demos when it came to walking in a straight line or picking things up. The Unitree H hand stands about 1.88 meters tall, around 5 foot 11 and weighs 47 kilograms. That's about 104 pounds. They spent three months training with AI using laser slam for positioning. This helped them handle stage nuances like little gaps in the floor and the rapid changing of dance formations. These robots were officially rolled out in August 2023, even making an appearance at Nvidia's GTC conference in 2024. In each H1 robot sells for roughly 650,000 yuan. That's about US$90,000. Folks have been pointing out how China is stepping up big time in AI and robotics, especially after that AI assistant. Deepseek also made headlines. India, for instance, is keeping a close watch on Deep Seq's activities. Worried about data security and Elon Musk, he gave his own not so flattering opinion on Deep Seq, implying he wasn't super impressed. But here's the kicker. While the spotlight is on China's new AI and robotics achievements, other companies around the globe are making big moves too. Like Figure AI, they're the team building that commercial and residential humanoid robot called Figure 02. They raised around $675 million last year, boosting their valuation to $2.6 billion. And so far they they've raised a total of $1.5 billion. The big shock is that figure just announced on X, formerly Twitter that they're ditching their deal with OpenAI. Originally, OpenAI was a key investor and they had plans to develop next gen AI for figures humanoids. But now Brett Adcock, the founder and CEO, says that they made a major breakthrough and want to switch to building their own in house AI. AI. According to him, you can't just outsource the type of embodied AI you need to run a robot in real time. That's part of the reason they're going all in on an end to end system. Interestingly, OpenAI is also backing another humanoid robot startup in Norway called 1X. And on top of that, OpenAI just filed a new trademark application that references humanoid robots that can learn, communicate and even entertain people. So it looks like they're not giving up on robotic hardware projects themselves. Meanwhile, Figure's new approach might be focusing on factory uses first. BMW, for instance, began trying out Figure robots in a South Carolina factory, which is a pretty big test site. If successful, that could be huge for large scale industrial deployment. Brett Adcock is also hinting at unveiling something no one has ever seen on a humanoid in the next 30 days. So, yeah, definitely a lot of hype going on there. Now, in other robot news, Elon Musk jumped on X to talk about how intricate Tesla's optimus hand is, calling it more complex than a Faberge egg. That's when clone robotics chimed in, claiming that their own humanoid hand is actually lighter since they use artificial muscles instead of metal motors. Stronger and cheaper to produce. They even joke that it's soft enough to give comfy massages and hugs. So there's definitely a rivalry brewing in terms of who can build the best robot hand. Clone basically said their muscle based approach beats Tesla's motor based design any day. Fewer parts, less weight, more strength. It's a bold statement, but we'll have to see how that plays out in real world testing. Meanwhile, there's yet another big development in humanoid robotics, this time from Nvidia and Carnegie Mellon University. They're working on a new training framework called asap, which stands for Aligning simulation and real world physics for learning agile humanoid whole body skills. The researchers basically want humanoid robots to mimic top athletes. So they fed their system videos of big sports stars like Cristiano Ronaldo, LeBron James doing his Silencer celebration and Kobe Bryant's legendary fadeaway shot. They even taught the bot some dance moves inspired by K pop star Rose. A tool called Tram converted these normal videos into three dimensional motion data. After that, the robots learned in simulation, first through something called reinforcement learning, and then the team refined them to handle real life physics. One interesting challenge is the so called Real2 Sim2 Real Gap robots can do well in a computer sim, but when you throw them into the physical world, factors like motor heat and mechanical stress can cause them to fail. So the ASEP framework involves the robots practicing in a simulator, collecting data from real world attempts, even if those attempts are messy, and then adjusting the simulation to match what actually happened. They use something called a delta action model, which basically patches up the differences between the simulator's physics engine and and the real world. That way, the next time the robot tries that jump shot or that dance spin, the simulation is more accurate and the robot's moves become smoother and more lifelike. The big takeaway is that robots could be a lot more agile and expressive if we can handle all the physics quirks that show up when metal or muscle based actuators. If you're clone robotics meets real world friction, gravity and torque limitations. The study also pointed out that these advanced movements can be brutal on hardware. Overheating motors and stressed out metal or plastic pieces lead to frequent breakdowns. 2G line robots were damaged during the tests. The researchers also said that future approaches might integrate damage awares policies that adjust on the fly to keep the robot from blowing a motor. It's also worth noting how much time and money can go into making these humanoids truly human. Like Unitree's H1 is priced around $90,000, while figure is sinking billions of dollars into their broad vision. Elon's Tesla is doing the same, funneling loads of resources to develop Optimus. Some companies are focusing on commercial tasks first like factory work or warehouse jobs because businesses have a higher willingness and budget to pay for these futuristic helpers. Others, like 1x are already pushing toward making robots useful in the home, which is a whole other challenge because you're dealing with everyday random tasks, kids running around or pets underfoot. So basically China is pushing AI and robotics hard. Unitree's dancing humanoids are wowing everyone and shorter training times mean these robots are improving fast. Meanwhile, figure AI split from OpenAI so they can control every aspect of their humanoid's hardware and software. We've also got that friendly rivalry over robot hands. Musk's motor based design versus clone robotics muscle powered approach. On top of that, Nvidia and CMU are teaching humanoids to move like pro athletes using their ASAP framework which bridges simulation and real world practice. All this competition is great for speeding up advances in humanoid AI, whether it's perfecting robot hands or doing backflips while carrying fragile items. We'll see more big reveals soon. Figures Secret project, Tesla's next Optimus Update or whatever Nvidia and CMU come up with next. In a shocking turn of events, a new player has just taken the lead in what we're calling AI. World War I. AI2's latest model, Toloo 3445B, has come out of nowhere and outperformed both controversial deepseeks model and OpenAI's 01 on several major benchmarks. The AI wars began when Deepseek, a Chinese startup, released a model that could match or even surpass OpenAI's strongest offerings for free, sparking a fierce rivalry. Then Alibaba got in on the action with their QEN 2.5 model, upping the ante even further. To make things more intense, Microsoft and OpenAI accused Deepseek of stealing their tech, adding more drama to the war. But now with to Lou 3 405B on the scene. It's clear the stakes have been raised and the competition is more heated than ever. This is getting really interesting and I can't wait to see where it goes next. So here's the rundown AI2, short for the Allen Institute for AI, based in Seattle, is a non profit research organization. They're famous for doing cutting edge stuff in NLP and other AI research areas. This time they're stepping into the Spotlight by releasing Toulu 3.405B. The name 405B literally points to the insane scale of parameters the model uses. Bigger models often show better reasoning abilities, or so we've seen historically, and Thule 3.405B definitely follows that trend. It was apparently trained using 256 GPUs in parallel, which shows how big this project was in terms of compute resources. Now to lo3405b is special not just because it's massive, but also because it's fully open source, unlike many powerful models out there. Everything you need to recreate this model, like training code, data and instructions, has been freely released and is permissively licensed. The folks at AI2 are calling Tulu 3.45B a big step for the US in showing that America can still be a leader in building top tier open source AI without always relying on big corporate labs. It's part of a bigger conversation about OpenAI research. You know how a lot of AI breakthroughs get locked behind closed APIs or proprietary code? Well, to lo3.4 5B wants to break that pattern, but let's talk performance. How do we measure something like this? AI2 tested it on a bunch of popular benchmarks including pop, QA, GSM K and math and more. These tests cover everything from knowledge recall and factual correctness to advanced reasoning, math, word problems, coding tasks and instruction following and internal evaluations. To lo 3.405b beat deep seq v3 and gpt4o on bunch of tasks. For instance on an evaluation set called Pop QA which has over 14,000 knowledge questions from Wikipedia to LO3.4 5b did better than not only those two models but also Meta's Llama 3.1.405b version. Pretty nuts, right? Another place to LO3.5b shines is Math Tasks on a benchmark called GSM 8K which is all about grade school level math problems. Toulu 345b actually snap snagged the highest performance among models in its class. That's always interesting because math is for many models quite challenging, especially if they haven't been explicitly fine tuned in that domain. If you want to try to loo 345B out for yourself, AI too has a chatbot web app that showcases it. They also posted code on GitHub and on hugging Face. So if you're a developer or a researcher, everything is there. You can test it, adapt it, or even combine it with your own data. It's all free and open now. Tulu345B didn't just show up overnight.