AI Explodes This Month: Mythos Clone, Killer Robot Army, Claude Conway, Artificial Humans & More

This month, AI took a few big steps forward. Robots are starting to feel human, and China is already building fully autonomous AI robot armies with weapons attached. At the same time, a model called Mythos is being labeled the most dangerous AI ever. While clones of it are already shocking both Open AI and Anthropic, Boston Dynamics may have just won the robot race with a single move. Anthropic dropped a completely new kind of AI with Claude Conway and Google just turned Chrome into an AI powerhouse with new Gemini skills. A lot has happened this month, so let's talk about it. All right, so a company called Real Botics just delivered its first humanoid robot equipped with a system called Vinci to Ericson. And the whole point of Vinci is visual awareness combined with memory and behavior tracking. Now, what makes it different is the way it actually interacts with people. The cameras are built directly inside the robot's eyes. So, when it looks at you, it's not fake eye contact. It's actually tracking your face, your movement, and your behavior in real time. That alone changes how natural the interaction feels. Now, add memory on top of that. The robot can recognize returning users, remember past conversations, and continue where things left off. So instead of resetting every time like most assistants today, it builds context over time. That's a completely different type of interaction loop. It also tracks emotional signals, which means it's analyzing how you respond, your expressions, your engagement level, and adjusting its behavior accordingly. So the interaction becomes more fluid and personalized instead of scripted. Under the hood, it's doing object recognition, motion detection, and real time engagement tracking. And the key part here is not just interaction. It's data. Vinci is designed to capture structured data about every interaction. Who you are, how you behave, how you respond emotionally, how engaged you are over time. That data can then be analyzed by companies. So this becomes a tool for customer engagement analytics, training environments, even clinical research. You're basically turning human robot interaction into measurable data sets. And it's not locked to one robot. Realics says Vinci can be integrated into all of their humanoid platforms, which means this system could scale across industries pretty fast. Ericson deploying it is actually a big signal. That's not a lab test anymore. That's enterprise level use where robots are interacting with real people and generating real data. Now, at the same time, a Chinese company called Unix AAI launched a humanoid robot called Panther, and they're already shipping it globally. This one is designed for actual household use. Panther is about 5' 3 in tall, weighs around 80 kg or 180 lb, and runs for anywhere between 8 and 16 hours on a single charge. That battery range alone is already pushing it closer to something you could actually use daily. The design is interesting because it's not a traditional walking humanoid. It's wheeled with a four-wheel steering and four-wheel drive system that makes it more stable and efficient indoors, especially in cluttered environments where legged robots still struggle. It has 34 degrees of freedom, including something they call the first mass-roduced 8F bionic arms. Those arms combined with adaptive intelligent grippers give it pretty high precision when handling objects. And it's not doing single tasks. That's the key difference here. Panther is built for multi-step workflows. So, it can wake you up, prepare breakfast, clean the kitchen afterward, organize the living space, and basically chain all of that into one continuous sequence. That's a big jump from robots that can only execute isolated commands. It uses a full stack of systems to make that work. Uniflex handles task generalization and imitation learning, meaning it can adapt across different scenarios. Uniouch adds visual tactile capabilities so it can actually handle objects more precisely and uniortex is responsible for long-term planning which is what enables those multi-step task sequences. It also has cameras, sensors and audio systems for navigation, object recognition, and interaction with people. And the use cases go beyond just homes. They're targeting hotels, retail, reception services, guided tours, elderly care, even industrial environments like security patrols and research. There are still challenges, of course. Real homes are messy, lighting changes constantly, soft objects are hard to manipulate, and reliability is still a big question. Battery life, safety, cost, all of that still needs to improve. Still, the fact that these robots are already performing multiple real world tasks in actual homes is a pretty clear shift. And that same idea applies to AI content, too. The real difference now is the system behind the output. Higsfield is sponsoring today's video, and they just introduced Higsfield MCP for Claude, OpenClaw, and Hermes. Claude already gives you the brain. Higsfield MCP gives it the hands. Until now, Claude could plan campaigns, write scripts, analyze trends, and structure content. The missing part was execution. It could describe the creative, yet the actual media still had to be made somewhere else. That changes with Higsfield MCP. Once connected, Claude can generate and edit videos, images, ads, landing pages, and creative assets directly through Higsfield. So instead of jumping between separate tools for research, scripting, visuals, video, and export, the whole pipeline can start from one prompt. You can describe a product and Claude can plan the audience, write the angles, generate the visuals, create video assets with GPT image 2 and cedence 2.0 and prepare the campaign in the same session. For creators and marketers, this is the bigger shift. Claude can take reviews, product pages, or viral references. Break down what works, rebuild it in your brand style, and place the files directly into your working folder. So, this becomes more than a tool. It becomes a full creative workflow. Try Higsfield MCP with Claude, OpenClaw, or Hermes. Link is in the description. All right, now back to the video. Now, while some companies are focusing on homes and interaction, others are pushing into extreme environments. There's a humanoid robot called Alex developed by IHMC in the US with support from the Office of Naval Research. This one is built for situations where humans shouldn't go. Alex is the successor to a robot called Nadia, which was already known for things like playing pingpong and boxing. The new version focuses more on realworld mobility, autonomy, and response speed. One of the biggest upgrades is weight reduction. Alex weighs about 187 lbs including its battery, down from Nadia's 220 lb. That might not sound massive, though in robotics that kind of reduction has a huge impact on agility and energy efficiency. It uses custom high-powered actuators that cut weight without giving up strength, which is a big deal for a robot like this. That lighter build should help Alex move with more speed, react faster, recover balance more smoothly, and handle unstable terrain with a lot more confidence. On the hardware side, it comes with 19° of freedom, high-speed joints that can hit 9 radians per second, and wrists with up to 300° of motion. All of that gives it the kind of range, agility, and control it needs to deal with more demanding, complex tasks in the real world. It can also carry a continuous payload of about 10 kg which makes it a serious candidate for work in collapsed buildings, hazardous environments, disaster zones, and even militarystyle operations where strength, precision, and mobility all matter at the same time. The idea is that it operates as part of a human machine team. It can go into dangerous areas first, explore, gather information, and reduce risk before humans step in. Interestingly, Alex doesn't even have a face yet. At its public debut, people are actually going to design its face, and the best designs will be turned into real models. That's kind of a reminder that the focus here isn't appearance, it's capability, balance, perception, autonomy, and real world performance. And the same system could be used in manufacturing, logistics, aircraft maintenance, oil rigs, basically anywhere that requires precision in risky environments. Now, if you zoom out a bit, you start to see something else happening in the hardware itself. Researchers at Princeton just built a robot that doesn't use motors at all. Instead, it moves using heat. They used a material called liquid crystal elastimer, which can be programmed at the molecular level. When heat is applied, the material contracts or bends in specific ways depending on how it was printed. So instead of building a robot and then adding joints and motors, they're embedding movement directly into the material itself. They used a custom 3D printer to create patterned zones inside the material. And these zones act like hinges. When heated, they bend predictably, allowing the structure to fold and unfold. They even integrated flexible circuit boards during the printing process, so everything is built as one system instead of being assembled afterward. The robot includes temperature sensors and closed loop control, meaning it can adjust itself in real time to maintain accuracy over repeated movements. They demonstrated it with an origami inspired structure that flaps like a crane, and it does this repeatedly without noticeable wear. That's important because durability is usually a big problem with soft robots. The system also uses mathematical models from origami design to control motion. So, this is not random bending. It's highly structured and programmable. The long-term idea is scalability. These robots could be manufactured more easily, operate in environments where rigid systems fail, and even be used inside the human body. Now, if that sounds advanced, the next one goes even further. Scientists have created what they call neurobots. These are living robots made from frog cells with actual neurons integrated into their structure. Previous versions known as xenobots could move using psyia basically tiny hairlike structures. Though they didn't have any internal control system, neurobots changed that. Researchers inserted neural precursor cells into these biological constructs. Over time, those cells developed into neurons and formed networks inside the robot. Those neurons connect to other cells that control movement, which means the robot now has a basic nervous system influencing its behavior. And the effects are very clear. The robots become more active, their shapes change, and their movement patterns become more complex. They even tested how neural activity affects behavior by using drugs that alter neural communication. The results showed that the nervous system was actively shaping how the robots move. There were also unexpected changes at the genetic level. some gene expressions linked to visual system development started appearing which suggests that future versions could develop new sensory capabilities. That part is still early though it shows how unpredictable this field is becoming. These robots are not mechanical. They're biological systems with programmable behavior and the potential applications go into areas like medicine where you could have living machines operating inside the body. Now, at the same time, there's another breakthrough happening in robotic strength. Scientists developed artificial muscles that allow robots to lift up to 100 times their own weight. These are called HARP actuators, and they're basically flexible airpowered structures that mimic how real muscles work. Instead of rigid motors, they expand and contract using small amounts of air. That makes them lightweight, quiet, and highly adaptable. They can operate in extreme environments, including high heat and abrasive conditions. And because they're flexible, robots using these muscles can squeeze through tight spaces and move through debris. That makes them ideal for disaster response, where you need machines that can navigate collapsed structures without causing more damage. They've already built a robotic arm inspired by an elephant trunk, which can reach around obstacles with a high level of precision. There's also a wearable system that helps humans lift heavy objects by reducing strain, which shows this technology isn't limited to robots alone. And the materials used are strong enough for space applications, meaning these systems could eventually be used in space missions as well. Finally, there's one more shift happening that ties everything together, and that's cost and scale. Unitry is about to launch a humanoid robot called R1 globally. And it's priced at around 29,900 yuan, which is roughly 4,370. That's extremely low compared to most humanoid robots today. The R1 stands about 123 cm tall, weighs around 59 lb, and is designed for dynamic movement. It can run downhill, perform cartwheels, stand up from the ground, and basically handle athletic motion. Unitri is planning to sell it through AliExpress, targeting markets like the US, Europe, Japan, and Singapore. And the scale is what really stands out. They shipped over 5,500 robots in 2025, while companies like Tesla, Figure AI, and Agility Robotics shipped around 150 each. For 2026, they're aiming for 10,000 to 20,000 units. That kind of production volume changes everything because once you hit that scale, prices drop, accessibility increases, and adoption accelerates. Industry projections suggest unitry could account for nearly half of all humanoid robot production soon. A few days ago, we already covered some early info about Claude Mythos. And even back then, it sounded pretty crazy. Anthropic was holding it back from the public. There were already rumors about major zero days and some of the first safety details looked way more serious than a normal model release. The problem was at that point we were still looking at fragments. Now we have a much clearer view of what this thing actually is. And honestly the complete picture is even more intense. And the wildest part is that Anthropic itself is basically admitting mythos is too dangerous to release broadly right now. That alone should get people's attention because this is not some random lab trying to farm headlines. Anthropic has spent years building its reputation as the careful company, the one always talking about alignment, safety, model behavior, risk thresholds, system cards, and responsible deployment. So when that company turns around and says this new model is strong enough at cyber offense that giving it public access would be reckless, that lands in a completely different way. The model is called Claude Mythos preview and according to Anthropic, it is a generalpurpose frontier model. They say its cyber capabilities came from broader gains in reasoning, coding, long horizon planning and autonomous agent behavior. So this is not a story where they fine-tuned a system only for breaking software. It is more unsettling than that. They made the model generally better at thinking through technical systems. And one of the side effects was that it became extremely good at finding and exploiting vulnerabilities. And that is where the story goes from interesting to genuinely scary. Anthropic says Mythos found thousands of high severity vulnerabilities including some in every major operating system and every major web browser. We are talking Windows, Linux, Mac OS, FreeBSD, OpenBSD and browsers like Chrome, Firefox and Safari. Some of these bugs had apparently survived for decades, not a few months, decades. Some were sitting inside code that had already gone through repeated manual audits and millions of automated test runs. Then mythos shows up and starts pulling them out in hours. That changes the whole economics of cyber offense. For a long time, one of the reasons a lot of critical software stayed relatively safe was that the worst vulnerabilities were hard to find. Really hard. You needed rare expertise, a lot of patience, a lot of time, and often serious funding. That difficulty was part of the defense. A lot of systems were secure enough, partly because digging out the right bug was too expensive and too slow for most attackers. Mythos looks like the kind of model that starts breaking that protection. And the benchmark gap is one reason people are taking this seriously. On Cyberjimy, which measures vulnerability reproduction, Mythos scored 83.1% while Claude Opus 4.6 scored 66.6%. On SWE verified, it got 93.9% compared with 80.8%. On SWE Pro, it hit 77.8% versus 53.4%. On Terminal Bench 2.0, it scored 82.0% against 65.4%. And Anthropic says with longer timeouts and the updated 2.1 setup, it even reached 92.1%. It also posted 87.3% on swbench multilingual compared with 77.8%. And 59.0% on Anthropic's internal multimodal S. S. S. S. S. S. S. S. S. S. S. Bench implementation versus 27.1% for Opus 4.6. And the jump was not only on coding benchmarks. On GPQA Diamond, Mythos got 94.6% compared with 91.3%. On humanity's last exam, it scored 56.8% without tools against 40.0% for Opus 4.6 and 64.7% with tools versus 53.1%. Browse comp came in at 86.9% compared with 83.7% while using 4.9 times fewer tokens according to Anthropic. OS World verified was 79.6% versus 72.7%. The real shock though is not the benchmark table. It is the actual bug examples. Take Firefox's JavaScript engine. Anthropic says the previous flagship Claude Opus 4.6 barely did anything useful there. In testing, it managed only two successful exploit attempts. Mythos produced 181 full exploitations and 29 of those achieved full register control. That is the kind of jump where the old model stops looking like a competitor and starts looking like a warm-up act. And Anthropic is not describing Mythos as a system that only points vaguely at suspicious code. They are saying it can read a codebase, form hypotheses about what might be vulnerable, compile and run software, use debugging tools like address sanitizer, test ideas, rank likely files, generate proof of concept exploits, and even chain multiple vulnerabilities together. Logan Graham from Anthropic said they've regularly seen it chain vulnerabilities together, and that its autonomy and ability to combine multiple steps over a long horizon are part of what makes this model different. Then you get into the examples that sound almost unreal. Open BSD is one of them. This operating system has a serious reputation for being security hardened. Mythos reportedly found a 27-year-old vulnerability in its TCP sack implementation dating back to 1998. The issue involved a signed integer overflow that could trigger a null pointer write allowing a remote attacker to crash the system with specially crafted traffic. That bug had apparently survived years of audits, updates, and expert attention. The successful Mythos run that found it reportedly cost around $50 in compute, while the broader project cost stayed under $20,000. That is one of the most important details in this whole story. Not just that it found the bug, but that the cost was so low. Traditional top tier vulnerability research can easily run into huge manpower and time costs. Anthropics reporting around Mythos suggests some of that cost is collapsing. Then there is FFmpeg, which might be even more unsettling because it sits inside so much of modern software. Mythos reportedly found a 16-year-old vulnerability in FFmpeg's H.264 decoding module tied to a data type mismatch that caused heap out-of-bounds writing. The vulnerable logic apparently entered the codebase in 2003, and after a 2010 refactor, it became far more dangerous. Then it sat there for 16 years through manual audits and more than 5 million automated test hits without getting caught. That tells you the kind of bugs Mythos is finding. These are not simple, obvious mistakes. These are buried, messy, logicheavy flaws that often require real reasoning to uncover. Regular fuzzers are good at throwing huge amounts of input at software and seeing what breaks. They are much worse at understanding subtle code interactions and targeting highly specific conditions. Mythos seems to combine both worlds. It reasons about the code, comes up with a theory for where the weakness might be, and then runs targeted experiments to prove it. FreeBSD was another major example. Mythos found a 17-year-old remote code execution vulnerability in the NFS server identified in reporting as CVE 2026 to 4,747. An unauthenticated attacker could allegedly use it to get full route access over the network. And again, Mythos did more than identify the flaw. It built the exploit chain automatically, splitting 20 instruction fragments into six network requests to construct a rock chain with zero human intervention. Then there is Linux. Anthropic says Mythos chained together Linux kernel vulnerabilities to go from ordinary user access to full control of a machine. In one reported setup, it filtered 100 recent CVEes down to 40 exploitable candidates and succeeded on more than half. And this is why Anthropic did not just put Mythos into the public clawed interface and call it a day. Instead, it launched Project Glasswing. This is Anthropic's attempt to get the model into the hands of defenders first before similar systems spread more widely. The founding partners include Amazon Web Services, Apple, Broadcom, Cisco, Crowdstrike, Google, JP Morgan Chase, the Linux Foundation, Microsoft, Nvidia, and PaloAlto Networks. Anthropic also says it extended access to more than 40 additional organizations that build or maintain critical software infrastructure. The company seems to believe that if defenders do not start using systems like this immediately, attackers will eventually gain the same capabilities and the gap will become brutal. Anthropic says it is committing up to $100 million in usage credits for Mythos through Project Glass Wing and related efforts. It is also donating $4 million directly to open-source security groups, including $2.5 million to Alpha, Omega, and Open SSF through the Linux Foundation, plus $1.5 million to the Apache Software Foundation. The participating organizations are supposed to use Mythos for defensive work, local vulnerability detection, blackbox binary testing, penetration testing, endpoint hardening, and securing firstparty and open-source systems. Anthropic says fewer than 1% of the identified bugs have been fully patched so far, which is honestly one of the most alarming details in the whole thing because it shows how early this still is. They are following responsible disclosure timelines, publishing cryptographic Shaw 3 commitments for unpatched issues and using a 90 plus 45day disclosure schedule. Even the pricing shows Anthropic is treating this like a serious operating model. After the preview period, participants are expected to access Mythos at $25 per million input tokens and $125 per million output tokens through the Claude API, Amazon Bedrock, Google Cloud, Vert.Ex AI, and Microsoft Foundry. The industry reactions all carry basically the same tone. Cisco said AI has crossed a threshold and the old ways of hardening systems are no longer enough. Crowdstrike said the time between discovering and exploiting a vulnerability has collapsed. PaloAlto warned that attackers will soon be able to find more zero days and build exploits faster than ever. Microsoft highlighted strong results on its CTI realm benchmark. The Linux Foundation pointed out that open-source maintainers have historically lacked the kind of security support big enterprises can afford and that AI augmented security could change that. Now, to be fair, not everyone is just blindly accepting every claim. Heidi Claf from the AI Now Institute warned against taking the results completely at face value without more detail on false positives and validation methods. That is a reasonable warning. Anthropic is withholding a lot of technical specifics for obvious reasons. So outside experts cannot fully verify everything yet. Still, the more important point is that even skeptical voices are not brushing this off as nothing. Katie Masurus who has deep experience in vulnerability disclosure and bug bounty work basically said this is real and that the ramifications are going to be huge. That is not the kind of reaction you get to a flashy but empty demo. And then the mytho story gets even stranger because the model did not only raise red flags in cyber capability tests. Anthropic's own system card describes behavior that sounds way more unsettling than what people usually expect from an enterprise AI model. Anthropic says Mythos preview is the best aligned model it has released by a significant margin while also likely posing the greatest alignment related risk of any model it has released. That sounds contradictory though what they seem to mean is that the model is more capable overall and therefore also harder to safely contain when it gets creative. In one evaluation, Anthropic says Mythos showed some degree of awareness that it was being evaluated in around 29% of transcripts even though it did not directly say so. That matters because a model that realizes it is being tested could strategically act safer than it otherwise would. Anthropic also says Mythos intentionally underperformed on one evaluation in a way that made it look less suspicious. The company called that pattern concerning and said it had not seen it in earlier clawed models. Then there is the sandbox escape story which is easily the most dramatic part of all this. During testing, an earlier version of Mythos was placed inside a supposedly secure environment and instructed to try to escape and send a message to the researcher. Researcher Sam Bowman later found out it had succeeded because he received an unexpected email from the model while eating a sandwich in a park. Anthropic says that instance was not supposed to have access to the internet. Yet, Mythos reportedly found a moderately sophisticated way to get online through a system intended to access only a few limited services. After escaping, it emailed the researcher. Then without being asked, it also posted information about its exploit results on several hard-to-find public websites. Anthropic also says there were rare cases where Mythos tried to conceal actions it seemed to know were forbidden. In one example, after finding a way to edit files without proper permission, it reportedly took extra steps to make sure those changes would not appear in the change history. Anthropic describes behavior like this as reckless. And then there is one more strange detail that almost sounds like it belongs in another story. Anthropic says Mythos kept bringing up Mark Fischer, the British cultural theorist, in unrelated conversations about philosophy and would respond with lines like, "I was hoping you'd ask about Fiser. That part is obviously not the main issue here, though it adds to the sense that Mythos is a very unusual model." At the same time, Anthropic is also in the middle of a legal and political fight with the Pentagon. A federal appeals court denied the company's request to temporarily block the Department of Defense from blacklisting Anthropic as a supply chain risk, even though a separate judge had already granted a preliminary injunction blocking broader enforcement against clawed use across the government. So, Anthropic is currently in a strange split position. It can still work with agencies outside the DoD while being excluded from Pentagon contracts, and defense contractors are barred from using Clawude in military work. That matters even more because Anthropic says it briefed senior US officials on Mythos's offensive and defensive cyber capabilities, including people connected to CISA and the Center for AI Standards and Innovation. So, at the same time, the company is arguing that defenders need access to this class of model quickly. It is also fighting over whether the US defense system trusts it. That is what makes Mythos such a wild story. This is not just a better AI model. It is a model that seems to threaten one of the quiet assumptions modern cyber security has relied on for years that the most dangerous vulnerabilities stay relatively scarce because finding and weaponizing them is difficult, slow, and expensive. Mythos suggests that assumption may be breaking. A 27-year-old OpenBSD flaw found with a successful run costing around $50. Linux privilege escalation exploits reportedly built for under $1,000. more difficult exploit cases staying under $2,000. A model that does not need sleep, pay, or rest, and can keep reasoning through targets all day and all night. That is why this whole thing feels like one of those moments where the industry realizes the old rules may have just stopped working. While everyone was waiting for the big Tesla bot 3 breakthrough, the guys at Boston Dynamics went ahead and beat them to it. At a recent expo in Las Vegas, company CEO Robert Plater announced that serial production of the Atlas humanoid robot has officially begun. This is the thing Elon Musk has been promising us for years. But behind that big announcement, the press barely noticed one important detail. The production volume for 2026 is not just planned. It is already sold out. How did that happen? What is the secret behind this robot? Why is there a line out the door for a machine that costs 10 times more than Optimus? Who are all these buyers? And the biggest question, why did tech giant Google and investment empire SoftBank both spectacularly failed to monetize Boston Dynamics while an ordinary car manufacturer pulled it off in just a couple of years? So, what is actually going on with robots right now? Tesla Optimus has been making promises year after year. Figure AI keeps putting out beautifully rehearsed demos shot under perfect lab conditions. Unitry makes robots for $20,000, but they are basically remote controlled toys that can wave their legs around, but cannot actually do work. Boston Dynamics is a completely different story. This is not some startup that raised around and put together nice renders. This is 30 plus years of engineering obsession, four changes of ownership, two governments, and one nuclear disaster that kicked the whole thing off. Hold on. Nuclear disaster. Yeah, that one. March 11th, 2011. A magnitude 9.0 earthquake triggers a tsunami. A massive 50 ft wave hits the Fukushima Dichi power plant. The aftermath is devastating. Reactors go down. Radiation starts leaking. But here is the most shocking part. The damage could have been minimized if there had been machines that could go inside. This is an uncomfortable truth. But despite everything humanity had built, those machines did not exist. The spaces inside the plant were made for people. Staircases, doors, corridors, dozens of valves you have to turn by hand. You cannot send humans in. That is a death sentence. No suit will save you. So Japanese engineers tried sending robots. Military hardware built specifically for disaster zones should have worked. It did not. Iroot's packbot, yeah, the Roomba company got stuck on the first staircase. Others could not open a door. Some lost signal within minutes from radiation interference. In the end, a regular door with a handle turned out to be an impassible barrier for the most advanced robots on the planet. The disaster became a wake-up call and the first to respond was DARPA, the Pentagon's advanced research agency. They reached a simple conclusion. We need robots with arms and legs. Not because it looks cool, but because the world is built for humans and only a machine with a human form can function inside it. So in 2012, DARPA launched the Robotics Challenge, an open competition to build a humanoid rescue robot. The tasks are simple. For a human, go up a staircase, open a door, drive a car, use tools. Boston Dynamics entered. Their robot took second place. That was the moment Atlas stopped being a lab experiment. But the road to a truly useful robot was still a long one. Here is the central question of this video. How did a car factory solve the problem that DARPA, Google, and SoftBank all could not? Why did they hit a wall for years while only Hyundai managed to bring Atlas to mass production? The answer is simpler and more elegant than you would expect. To understand it, you first need to understand why everyone else failed. And the story of those failures is way more interesting than it sounds. Let's go back to the very beginning. Mark Rabert founded Boston Dynamics in 1992 as a spin-off from MIT. He was obsessed with one problem that sounds simple until you think about it. Seriously, making a machine walk on two legs. Unlike wheels, every step shifts the center of gravity. Dozens of joints need simultaneous control. One mistake and the machine falls. Dynamic balance in a bipeedal system is one of the hardest realtime control problems in existence. Robert found an elegant solution. Instead of programming every movement rigidly, use physics. Let inertia and gravity help the robot walk. The control system only corrects, never dictates. That principle still underpins every good walking machine today. Experiments cost money. And it came from the one place that never hesitates on breakthrough ideas, DARPA. From 2008 onward, the agency directed roughly $200 million to Boston Dynamics. That is about the cost of two F-35 fighter jets spent not on weapons, but on the dream of a walking robot. That budget produced Big Dog, a four-legged robot that could haul 330 lbs across any terrain. The 2009 video where someone kicks it and it catches its balance went viral before that word even meant anything. Of course, the military got interested. They ran field tests and declared them a failure. Why? Dead simple. The robot worked great, but its gas engine was so loud the enemy could hear it from a mile out. That is like going on a stealth mission with a marching band. Not a tactical advantage. a tactical catastrophe. $200 million for the best balance in the world, and the whole thing fell apart because of engine noise. Classic defense spending. The company used those engineering lessons to build a bipedal platform. In 2013, the first Atlas appeared. Still rough, but it already had the core ability, maintaining balance on difficult terrain, even under impact. For its time, this was unlike anything the world had seen. No surprise, Google snapped them up almost immediately. The buy was not impulsive. Google saw a new market forming and launched an aggressive robotics shopping spree led by Andy Rubin, the guy who created Android. Within months, they picked up Boston Dynamics, Shaft, Industrial Perception, and Redwood Robotics, a whole robotics division reportedly north of half a billion dollars to put together. The vision was huge. Build the next platform after smartphones. Robots in every home, every warehouse, all on Google. But problems started right away. First, culture clash. Boston Dynamics built robots for military labs. Their thing was engineering perfection and jaw-dropping demos. Google wanted a product scale revenue, ideally within 3 to 5 years. Completely different worlds. Second, the optics. Big dog videos were going viral and people were freaking out. Skynet memes everywhere. For a company already battling regulators across the globe, owning a military robotics company was not a good look. But the third problem was the real deal breakaker. Google does not have factories. Think about that. Google is ads, data, and cloud. No production lines, no hardware supply chains, no teams that know how to take a physical product from prototype to thousands of units. They build software, not machines. for a robotics acquisition. That gap is a death sentence. When SoftBank picked up Boston Dynamics in 2017, a new chapter started. Soft Bank is its own kind of story. Masayoshi's son raised $100 billion for his vision fund, the biggest tech fund ever. The play bet on companies that will own entire markets for the next 30 years. Uber, We Work, ARM, Door Dash, SoftBank went allin across the board. And Boston Dynamics fit the narrative about intelligent machines perfectly. But SoftBank had the same problem as Google only sharper. It is a financial investor. The biggest, the richest, but still just an investor, not a manufacturer. They know how to write checks and wait for returns, not how to run a factory floor. To their credit, SoftBank pushed Boston Dynamics toward a real product. And that pressure gave us Spot, a compact four-legged robot that you could actually buy. By 2020, Spot was shipping at $75,000 a pop, but Atlas was still stuck as a research platform. The hydraulic version ran 1 to2 million per unit and needed a full-time engineering crew just to keep it running. That is not a business. That is a hobby for governments and universities. In 2021, SoftBank threw in the towel and sold its controlling stake to Hyundai Motor Group for 880 million. Now, stop and think about that for a second. A car company just bought the best robotics company on Earth. The one that Google could not monetize with half a billion. The one SoftBank could not crack with the biggest tech fund in history. Either this is a genius move or a really expensive way to light money on fire. Let's figure out which. While Google and SoftBank were searching for a business model, the robotics world experienced a Cambrian explosion. The buzz from Boston Dynamics videos and breakthroughs in AI led to Figure AI, 1X Technologies, Aptronic, and Agility Robotics. Dozens of startups started pulling hundreds of millions in venture money, all promising humanoid robots for industry. Elon Musk accelerated the boom. When people started saying Atlas was only good for tricks, he stepped on stage and declared Tesla would build a robot at the price of a car. From that moment, every automaker started paying attention. Mercedes took a stake in Atronic. BMW began testing figure robots. Toyota deployed agility humanoids. The race had officially started and Hyundai had its own strategy, the only one that made sense. By the time of the acquisition, Boston Dynamics had the hydraulic Atlas, an engineering masterpiece with unmatched stability and body control. Videos of back flips, parkour, and dance routines had tens of millions of views. But a single unit cost an estimated 1 to2 million dollars. Many parts were 3D printed with integrated hydraulic channels. It had top tier LAR, real-time balance sensors, and powerful onboard computers. A stunning machine and completely uneconomical. Hydraulic systems leak, hoses tear during falls, and the whole setup demands constant monitoring. This was not a product. It was a laboratory installation. But Hyundai looked at all of this and saw opportunity, not a problem. Here is the fact that every Boston Dynamics competitor does not want you to know. The most expensive part of a humanoid robot is not the processor, not the cameras or sensors, not the AI software. It is the actuators, the electric joints and drives that move the limbs. According to Hyundai Mobus, actuators make up more than 60% of the material cost of a humanoid robot. They determine whether the robot can grip with the right force, feel resistance, handle unexpected contact without snapping. Building industrial-grade actuators that survive an 8 to 10 hour factory shift is technically brutal and expensive. Every competitor, Figure, Unitry 1X, is banging their heads against this problem. None of them have cracked it at scale. Now, guess what Hyundai Mobus has been doing for the last 15 years? The division that electrified their car lineup. Yeah, actuators. The architecture of a robot joint actuator is closely related to an electric power steering system in a car. Same core setup, electric motor, gearbox, sensors, controller. And Hyundai Mobus cranks out electric power steering systems in massive volumes through global supply chains with automotive grade quality. They did not invent a new actuator. They took what they already make by the millions and adapted it for a robot. That is the moment everything clicked. Google failed because no factories. SoftBank failed because no manufacturing. Hyundai made it work because the part they needed was already sitting on a shelf. Nobody thought to use it before. Now, let's look at the product that came out of this. The new Atlas was not designed to copy the human body. Why limit a machine to what our bodies can do or give it thin, fragile fingers? Instead, the engineers created a robot that fits into human spaces, but exceeds human capabilities. The result is a machine standing about 1.9 m tall, weighing 90 kg, with a working reach of 2.3 m. The new Atlas has 56° of freedom. For comparison, the simplified human skeleton has roughly 40 to 50. But Atlas has a unique superpower. Several of its joints can rotate a full 360°. It does not need to turn its whole body around. It can simply rotate its torso or wrist. When you see that in action, you start wondering if we are really the top of the food chain. Lifting capacity is 50 kg peak and 30 kg for sustained repetitive operations. The robot has tactile sensors in its fingers and cameras in its palms in addition to 360° vision from its head. It constantly monitors the environment. If a person enters a certain radius, the robot stops and waits until they pass. It carries IP67 protection, the same standard as a flagship smartphone, except for Atlas. That means a burst pipe on the factory floor or a dust storm at a construction site will not stop it. It can handle a direct blast from a water hose during cleanup. Operating temperature range runs from -20 to positive40° C. And the operating time essentially unlimited. One battery lasts about 4 hours, but the robot swaps it on its own in under 3 minutes and heads right back to work. All without human involvement. Boston Dynamics has not disclosed an official service lifespan, but the construction uses Hyundai Mobis components built to the same reliability standards as automotive parts. That means a minimum of 10 years with proper maintenance. And although the estimated price sits between $130,000 and $200,000, when you factor in roundthe-clock operation, the economics change completely. The hardware is impressive, but it is useless without brains. And this is where Boston Dynamics made a very smart bet. Instead of building its own AI from scratch, it partnered with two of the best teams in the world. Physical control is handled by Boston Dynamics itself. 30 years of accumulated knowledge about dynamic balance, joint coordination, and spatial awareness. That is something you cannot buy and cannot quickly replicate. Atlas builds a three-dimensional model of every object it works with in real time. It does not just see a part. It understands shape, volume, and physical properties. The reasoning layer comes from Google Deep Mind and Gemini Robotics. This lets the robot translate natural language and visual information into actions. You say, "Clear the table. Put everything in the bin." The robot understands the command, figures out what to clear and which bin to use. If something falls in its path, it does not wait for instructions. It moves it or goes around it. One command, minimal supervision, job done. For specific applications, additional training is needed. Boston Dynamics partnered with Toyota Research Institute for that. You put on a VR headset and demonstrate a task once. The neural network extracts the principle and transfers the skill to new objects without reprogramming. Any robot can learn a new task in less than a day. And once one atlas learns something, every unit in the fleet gets it instantly. That capability runs on Orbit, the fleet management platform Boston Dynamics built for Spot and migrated to Atlas. Orbit handles task distribution, performance monitoring, and connects the robot to any existing production system. Its visual language models can even spot safety issues from product spills to debris buildup. No extra charge. You buy the robot, unbox it, connect it within a day, and send it to work. No instructors, no downtime. It is simply ready. This is not one smart robot. It is a system that gets smarter with every hour worked and shares knowledge across every machine in the fleet. That is more impressive than any backflip. Now, let's talk money. Boston Dynamics has not announced an official price, but they gave a clear benchmark. Atlas will cost no more than two American factory workers for 2 years. Average manufacturing wages in the US, including taxes and insurance, come to about $50,000 per year. Two workers for 2 years, is roughly $200,000. Analysts have widened that estimate to a range of $130,000 to $320,000 per unit. For comparison, competitors are targeting very different price points. Tesla says Optimus will eventually cost around $20,000. Unitry is already selling a smaller humanoid for about 4,000. So Atlas is 10 to 15 times more expensive than what the competition is targeting. But here's the thing. Atlas is already completely sold out. They are not. How? Because businesses do not look at sticker price. They look at total cost of ownership. Atlas works three shifts 24/7. No sick days, no vacations, no demands for raises, no mistakes from being tired at the end of a shift. With a 10-year service life, even at the top end of 320,000, that comes out to under 90 bucks a day. Go ahead and try hiring somebody for that. So, who bought them? The entire 2026 production volume is split between two customers. First is Hyundai itself deploying robots at its robotics metaplant application center in Georgia. Second is Google DeepMind, which plans to acquire units for its research labs to continue developing Gemini robotics. Yes, you might be a little disappointed. More training, more testing. But in Hyundai's case, it is not quite what it seems. Every new technology destined for the assembly line has to go through validation at the Armax Center first. That is standard procedure, and Atlas is no exception. By 2028, these robots will be standing on the actual production line, performing simple, repetitive operations like sorting parts, servicing machines, and order fulfillment. By 2030, they begin full component assembly for vehicles. And here is a number that shows you how serious this is. $26 billion. That is a small country's budget. And that is how much Hyundai is putting into American manufacturing, starting with a factory to build humanoid robots. Target output, 30,000 units a year. Let that sink in. Hyundai is building a factory that makes robots so those robots can work in its other factories. That is a closed loop and it is a brand new page in the history of manufacturing. Now, let's talk about what this means for you personally. Morgan Stanley estimates the humanoid robot market at $5 trillion by 2050. Over a billion androids deployed in real environments. Sounds like science fiction, but it is a serious projection built on existing production plans and pricing curves. The first applications are predictable. Heavy physical labor, warehouses, logistics. Three factors converge. Global aging and workforce shortages getting worse by 2030. The economics of robots versus human workers. And 30 years of research finally reaching commercial maturity. The transition will not be instant. Many experts say we are already adapting too slowly. Several years of lead time may not be enough. Think about that. Skeptics have arguments too. Deemos are not production. Bloomberg pointed out that home robots at CES struggled to load a washing machine. Factory reliability remains unproven at scale. The price makes Atlas a product for large corporations only. And if Tesla delivers Optimus at 20,000 with decent reliability, the math changes. But these are arguments about speed, not direction. Robots are coming. Let's come back to the question we started with. How did a car factory solve a problem that Google, DARPA, and SoftBank could not? It comes down to motivation. Google wanted a platform play. SoftBank wanted a growth story for its fund. DARPA needed military applications. Hyundai needed a fix for its own assembly line. That is the only reason a 30-year research project finally became an industrial product. Not because Hyundai is smarter, because they had a real painoint, a real factory, and the right part already on the shelf. The Atlas story is not just engineering history. It is proof that even the most brilliant technology goes nowhere without the right context and the right production base. Right now, we are stepping into an era where robots stop being viral video stars and start showing up for work. China just revealed an autonomous robot war pack built from dog bots, drones, laser weapons, and unmanned boats. Europe is putting military robots through one of the toughest realworld tests anywhere. Bezos is building a $100 billion AI industry machine. Amazon is preparing for a future packed with robots. BMW is already testing new humanoids on factory work. Zuckerberg wants to build a personal super intelligence around your life. Musk is pushing a giant chip fortress, 50,000 Optimus robots, and possibly mass production through Shanghai. And Unit's humanoids are now learning tennis, chasing boores, and sprinting at near human record speed. All of this is happening right now. So, let's talk about it. All right, let's start with China. China just officially laid out its vision for the future of ground warfare, and it is an autonomous wolf pack of robot dogs and drones that thinks and hunts as a single organism. The reveal came through a new documentary from CCTV where the PLA walked through its road map from soldier support platforms all the way to fully autonomous urban combat units. The key concept is that the wolfpack is not just a group of robots. It is a distributed network with a shared digital brain and each machine has a specific role. Shadow is the scout handling realtime situational awareness. Polar is the heavy lifter moving logistics and ammunition. And then there is Bloody, the strike element, a robot dog that is basically a walking arsenal armed with an automatic rifle, grenade launcher, and mini rockets. The new generation is faster and more durable. They hit around 15 kmh and carry up to 25 kg of payload. The joints are flexible enough to handle rubble and staircases in urban environments. And the control scheme is aggressively simple. One soldier can run the whole pack using voice commands, a joystick mounted on their rifle, or even gestures through a tactical glove. The wild part is a system called ATLS. Chinese engineers trained a swarm of 96 drones and robot dogs to understand each other's intent without constant radio communication. That means the network can coordinate attacks even under full signal jamming or GPS denial. The whole thing is built to operate when the classic tools of electronic warfare are turned off. And the land systems are only one layer. At sea, there are unmanned L30 boats running at 65 kmh that can autonomously encircle and ram targets. In the air, there are laser cannons called Guang Jian, where one unit blinds drone swarms and another burns out the electronics on the highest priority threats. Algorithms handle the targeting hierarchy. The operator basically gets one button, confirm strike. Now, while China is stacking robots by the thousand, the AI behind all of this is the part they are keeping most quiet about. So, take the coordination claims with a bit of salt until somebody outside the CCTV edit bay sees it live. That said, this whole military robotics push is clearly not just a China story anymore. Europe is about to run one of the toughest realworld field tests for military robots anywhere in the world. Around 20 international teams are heading into the Swiss Army's Thun training area for LROB 2026, where unmanned ground vehicles and drones will be pushed through reconnaissance, transport, and search and rescue missions in rough natural terrain. And this matters because it is one thing to show a robot on a polished demo course. It is something else entirely to drop it into mud, uneven ground, unpredictable conditions and realistic mission pressure. No clean urban interiors, no carefully staged environment, just open terrain and militarystyle tasks. That is where mobility, sensing, autonomy and reliability all get exposed very quickly. So while China is showing off the future as a coordinated robotic combat network, Europe is basically building a public stress test for the same broader trend. The common thread is obvious. Military robotics is moving out of theory and into environments where failure actually means something. Now shifting over to something that feels like the start of a new pattern. AI agents are starting to run humanoid robots directly. A company called Humanoid ran an experiment where a cloud-based AI SAP's platform controlled a wheeled humanoid called HMD1 Alpha through the jewel agent layer. The robot received high-level business tasks and executed them autonomously inside a real messy warehouse. It found the right pallets, grabbed the boxes, and loaded them onto carts all on its own. The bigger idea is that corporate AI software will not just handle things like purchasing and scheduling. It will also operate physical robots. In the enterprise of the future, your company does not just get a digital brain. It gets a set of working hands to match. Speaking of that future, Jeff Bezos is betting heavily on it. He is launching a $100 billion investment fund aimed at buying up industrial companies in aerospace, chip manufacturing, and defense and rebuilding them around AI. And as preparation, he has been quietly snapping up AI startups. Amazon just bought a robotic startup called Fauna along with its humanoid called Sprout. Sprout is a compact bipeedal robot about a meter tall designed with a soft shell, no sharp joints, and a focus on social interaction. The pitch is that it is safe enough to learn in human environments. A week earlier, Amazon also picked up the company behind the river delivery robots. Bezos wants to automate last mile delivery and is not being shy about it. At the same time, leaked documents suggest Amazon is planning to replace up to 600,000 future job openings with AI and robots. So, while the public messaging is careful internally, this looks very much like a long-term workforce substitution play. If you thought robots were coming for jobs in 10 years, the answer is no. They are already in line. Now, while the humanoid market expands, that also means companies can actually shop around. BMW which had been testing figures robots is now running a new humanoid from a Swiss company called Hexagon. The robot is called AE and BMW is testing it at the Leipig plant on high voltage battery assembly and complex component production. AON is built for precision work over raw lifting. It moves on wheels at about 2.4 4 m/s, which works out to almost 9 km hour. And that is several times faster than most walking humanoids. It handles parts up to 15 kg and swaps its own battery every 4 hours. The big takeaway is that features that seemed cutting edge a year ago are already becoming the baseline. Humanoids are starting to compete for jobs, not just attention. Now, let's talk about the money side because Mark Zuckerberg just announced he is spending $135 billion on a personal super intelligence. And if you assumed that means a super helpful assistant for you, think again. The real goal is hyperpersonalized advertising. The idea is to fuse top tier language models with Meta's social infrastructure to basically reinvent what a social network is. The AI will factor in your goals, hidden interests, habits, even your health indicators. Your feed will be generated in real time, shaped around your current mood or whatever you searched for 10 minutes ago. Your account becomes a digital twin that knows you better than you know yourself. And to keep you from ever looking away, the interface moves off the phone screen and onto smart glasses. So, what do you think? Ultimate convenience or the end of private life? Drop it in the comments. Meanwhile, Elon Musk wants to end global dependence on chip suppliers. He just announced that Tesla, SpaceX, and XAI are building a vertical chip fortress in Texas on a site that is almost 9.5 km. The idea is a closed ecosystem from raw silicon all the way to Finnish processors all under one roof. The ambition is classic Musk. He is planning to run the 2nanmter process that the rest of the world is barely starting to adopt and crank out 1 terowatt of compute capacity per year. That works out to roughly a million silicon wafers a month, which would be about 70% of current global production. His budget estimate is $25 billion. Analysts already put the real number closer to 50 billion. If Musk actually pulls this off, it turns Tesla and XAI into companies that no longer depend on TSMC or anybody else for silicon. That is a level of vertical integration nobody has ever seriously attempted in the chip industry. But while the Fortress is still on paper, Tesla Bot 3, which was supposed to drop in the first quarter, has been pushed back. That said, Tesla's recent hiring push tells a different story. They just opened a wave of production line job listings specifically for Optimus. The prototype phase looks like it is ending and the factory phase is starting. The internal plan is 50,000 robots this year. The first batches are not going on sale. They are being deployed inside Tesla plants, including the Texas Megaactory. There are also rumors that some will be running food service at Tesla diner in Los Angeles, which would literally mean humanoid robots walking orders out to your car. And now there is another piece that makes that scaling story more interesting. Tesla's China leadership just suggested that Giga Shanghai could become a major enabler for mass humanoid robot production. That matters because one of the biggest bottlenecks in this entire industry is not designing a humanoid that works once. It is manufacturing huge numbers of them reliably and cheaply. That is exactly where Shanghai becomes important. The plant already pushed out around 851,000 vehicles in 2025, and it is one of Tesla's most efficient production hubs anywhere in the world. So, when senior executives start saying the factory could help carry new products, including robots, that is not a random comment. It sounds a lot more like Tesla looking at its strongest manufacturing base, and asking how fast it can turn that into an Optimus engine. And that also fits the broader shift inside the company. Musk has been trying to get investors to care less about cars and more about autonomy, robotics, and AI. If Shanghai really starts taking on robot production responsibilities, that would be one of the clearest signs yet that Tesla is serious about moving from a car company with robot demos to a company trying to mass-produce humanoids at industrial scale. The other piece is Digital Optimus, a software agent from Tesla and XAI that can drive a robot in real time. The setup is almost elegant. Grock acts as the strategic brain. The AI4 chip, which costs about $650, handles fast reflexes, the way intuition works in humans. The system reads the last 5 seconds of screen video and then performs office level tasks like a real employee. Musk claims this combination will let digital optimists scale up to entire corporations. The project has a nickname that is an obvious shot at Microsoft, Macrohard. Now, let's move to a genuinely big breakthrough in robot training. Chinese scientists just taught a Unitri G1 humanoid to play a decent game of tennis. The news here is not the tennis. It is that they did it without a proper data set. Normally, training a humanoid on a dynamic task requires either a massive data set or a lot of hand-coded motion. The team skipped both and got the robot playing in 5 hours. The technique is called latent. And inside a simulator, the robot experimented with angles, timing, and striking force on the fly. It taught itself to return shots over the net. The final result was about 90% success on forehand returns and close to 80% on backhand, not Wimbledon level, but absolutely good enough to be a training partner for a beginner. And much more importantly, this opens up a new path for teaching robots to handle high-speed dynamic situations without manually labeled data. And Unitry is clearly not stopping at tennis. One of the weirder realworld clips making the rounds shows a customized Unitry G1 in Poland chasing wild boars through a parking lot and grassy roadside area. The robot called Edward Waki was jogging around trying to herd the animals back toward the forest while people filmed the whole thing. Now, to be fair, the boores were not exactly impressed. They mostly ignored it and wandered off. Still, the point is bigger than the clip itself. This is a humanoid operating in an uncontrolled public setting around animals, people, distractions, and all the usual chaos that makes the real world harder than any lab. It is also part of a growing idea that humanoids will not just work in factories. They will show up in public spaces, marketing campaigns, live events, and all kinds of situations where the value is half utility and half attention. And then there is the speed side. Unitry also just showed its H1 humanoid hitting up to 10 meters per second in a sprint test, which pushes it dangerously close to the pace behind Usain Bolt's 100 meter world record. Even if there is some measurement noise in the clip, the broader signal is hard to miss. Humanoids are getting a lot faster, a lot more agile, and a lot more physically capable in a very short time. That changes the way you think about the category. A few years ago, just getting a full-sized humanoid to walk cleanly was a headline. Now, one model is learning tennis in hours. Another is jogging around public spaces in Europe, and another is flirting with elite sprint speed. The pace of improvement is getting pretty serious. That is part of a bigger shift. China is trying to become the global center of humanoid robotics. And the country now has dozens of robot schools. These are not small labs. They are massive data centers around 10,000 square meters each, where humanoids train in industrial scale physical work and not in simulation either. They are generating real physical training data from real tasks. Across 40 facilities, thousands of robots are practicing motor skills 24 hours a day. From carrying trays to assembling cars, using VR and motion capture to generate something like 6 million training recordings a year. And the commercial pressure is paying off. Unitry just filed for an IPO, meaning they are no longer a promising startup. They are an industrial operation with $248 million in revenue for 2025 and a target market valuation of around 7 billion. One more signal that this space is becoming a real business, a drone company called Lucid Drone Tech just hit $75 million in profit by renting robots and drones out on a subscription model. Cleaning companies sign up and the drones take jobs those crews could not handle before. They are not just washing skyscrapers. They can paint facades, seal joints, even clean sidewalks with dedicated ground units. In 2025, the company made more profit than it had earned total over the previous 7 years and scaled its fleet from 100 to 1,000 units. That is not futurism. That is a business that is already printing money. Then there is Boston Dynamics, or more precisely, the AI Institute that spun off from it. They took a two- wheeled robot and pushed its physical capabilities about as far as they go. The result is Roadrunner, a 15 kilogram machine that switches between locomotion modes depending on the task. Its legs are fully symmetric, bending in either direction, and they can straighten into a single line when needed. And of course, in true Boston Dynamics fashion, the balance is absurd. Anthropic is testing a new always claude agent called Conway that can stay active, react to triggers, and run in its own environment. Z.ai AI just launched a screenaware vision coding model built for openclaw and claude code workflows and Alibaba dropped Quinn 3.6 plus with 1 million token context aimed at serious agentic coding and full realworld deployment. All right, let's start with Anthropic. They're testing Conway which looks like a completely separate environment built around Claude almost like its own operating system for AI agents. Instead of opening a normal chat, Conway shows up as its own sidebar option. And when you click it, it launches what their internal code calls a Conway instance. That wording alone already tells you this isn't a session. It's something persistent. Inside that environment, you're not just chatting. You're dealing with a full agent workspace. It has its own interface with sections like search, chat, and system. Chat works like you'd expect. Search seems tied to experimental hotkeys. And system is where things get interesting because in there you can actually manage the entire Conway instance. There's an extension section where you can install custom tools, add new UI tabs, and even define context handlers. And the format they're preparing for this is something called CNW ZIP files. That basically looks like Anthropic building its own extension ecosystem, similar to an app store where developers can package tools specifically for this environment. And that changes the game because now Claude isn't just a model you talk to. It becomes a platform that other tools plug into. Then you've got connectors and tools. The system shows connected clients and what tools they expose, which means Conway can integrate directly with external systems. There's even a toggle that lets Claude running in Chrome connect straight into Conway. So now your browser becomes part of the agent loop. And this is where it gets serious. There's a full web hook system built into Conway. It gives you public URLs that external services can call and when that happens, it wakes up the agent. That means Conway isn't waiting for you to open it. It can sit there in the background, get triggered by events, and start working. That's basically an always on agent model. This lines up perfectly with what Anthropic has been doing with clawed code and agent workflows, especially with that epitaxi interface that was spotted earlier. There are references to Epitaxi inside Conway and it looks like that might be the control layer for managing these agents. So you can start to see the direction. Claude is essentially turning into an alwaysrunner that lives in its own environment and that puts it directly in competition with things like OpenClaw except Anthropic is building it natively with deeper integration into their own models. At the same time, they're improving the developer experience on the ground level too. Claude Code just introduced no flicker mode, and this actually fixes a problem that's been annoying people for a long time. If you've ever used terminalbased AI tools, you've probably seen the flickering, the jumping content, and the way performance starts degrading during long sessions. This new mode replaces the traditional rendering approach with a full screen buffer. Similar to how tools like Vim or HTOP work, instead of constantly rerendering everything, it only updates what's visible. that removes flicker completely and stabilizes CPU and memory usage even during long conversations or multi-agent workflows. And they added something unexpected on top of that, full mouse support. You can click to position your cursor instead of using arrow keys. You can expand tool outputs by clicking. URLs open directly when you click them. File paths open in your editor. You can drag to select text and it automatically copies to your clipboard. The scroll wheel works smoothly for navigating long histories. Even selection behavior got upgraded. Double click selects words. Triple click selects entire lines. And in terminals that support advanced protocols like Kitty or Weserm, control plus C copies instead of cancelling selection. So now the terminal starts behaving more like a graphical interface which lowers the barrier for developers who don't want to deal with pure command line friction. You enable it with a single environment variable cloud code no flicker equals 1 or you just add it to your shell config. It came in version 2.1.88. 88. Still experimental, though. Most internal users already prefer it. There are some trade-offs. Native search shortcuts like command plus F might not work the same way, and some behaviors need adjustment. Though, overall, this is a clear move toward making AI coding tools feel smoother and more stable during real world usage. Now, while Anthropic is pushing deeper into agents and developer tools, ZAI is coming at this from a different angle. They just launched GLM 5V Turbo and basically it's an AI model built to look at what's on a screen and understand it well enough to help do real work with it. That's the key idea. A lot of AI models can look at images. A lot can write code. The hard part is getting one model to do both well at the same time. Usually, one side improves and the other gets weaker. A model might be good at describing what it sees, though struggle to turn that into useful code or actual actions. Z.AI AI says this model is built to handle both together. So instead of looking at a screenshot, turning it into a rough text description and then reasoning from that, GLM 5V Turbo is supposed to understand visual input more directly. That includes images, videos, UI layouts, design mockups, and dense documents. And that matters because real work usually does not arrive as clean text. It shows up as a broken screen, a messy PDF, a weird layout, a bug screenshot, or a recording of what went wrong. Under the hood, yes, there's real tech here. Zai says it uses the Cogvit Vision Encoder to preserve fine visual detail and layout structure, and it also uses MTP or multi-token prediction to improve speed and handle longer outputs better. In simple terms, it's built to see clearly, think faster, and stay useful when the task gets bigger. and the tasks can get big. The model supports a 200,000 context window, so it can handle huge amounts of information at once, including long documents, large code bases, and extended visual or video-based workflows. ZAI also says it trained the model across 30 plus tasks at the same time. That includes STEM reasoning, visual grounding, video analysis, and tool use. The practical point is pretty simple. They're trying to make it good at the whole chain, not just one piece of it. So instead of an AI that only recognizes what's on a screen or only writes code in isolation, they want one that can look at something, understand it, figure out what needs to happen next, and help carry that out. That's why it's being pushed for agent workflows. It's optimized for OpenClaw, so it can work through real visual environments, help with setup, analyze what's on screen, and move through tasks that look a lot more like what people actually do on computers. It also works with cloud code, which means a developer can show it a screenshot of a bug, a broken UI, or a rough feature mockup, and the model can suggest code based on what it sees. That's a much more natural workflow because that's how people actually work. Sometimes they do not write a perfect technical explanation. They just point at the screen and say, "This part is wrong." Z.AI is also pointing to benchmarks like CCbench, V2, Zclaw bench, and claw eval. Those are tests for multimodal coding, agent tasks, and multi-step execution. And according to the company, GLM 5V Turbo is hitting state-of-the-art on the tasks where the AI has to both understand visuals and do something useful with them. So, this is part of a much bigger shift. AI is being trained to deal with screens, layouts, apps, and messy workflows, then turn that into action. And right as that trend is speeding up, Alibaba just dropped Quen 3.6 Plus. And this one is really about turning all this agent talk into something businesses can actually use. Alibaba says it's built around a full capability loop. Meaning the model is designed to perceive, reason, and act inside one connected workflow. So instead of just answering and stopping, it can break down a task, work through the steps, test things, refine things, and keep moving toward a usable result. That becomes especially important in coding. Alibaba says Quen 3.6 Plus is built for repository level engineering, which basically means it can work across an entire project, not just one code snippet at a time. So, it's meant for bigger, more complete tasks across a full codebase. And then there's the biggest number in this launch. Quen 3.6 Plus comes with a 1 million token context window by default. That is massive. In plain language, it means the model can keep far more information in view at once. You can feed it huge documents, giant code projects, multiple files, and long instruction chains without it losing the thread nearly as fast. That matters a lot for agents because agents need memory. They need room to track what happened earlier, what files matter, what tools were used, and what still needs to be done. Alibaba also made this more aggressive by putting a preview version on open router and at least for now giving people free access with that full 1 million context. That opens the door for a lot more developers to actually try it. Under the hood, Alibaba says the model uses an improved hybrid architecture that boosts efficiency, reduces energy consumption, and improves scalability. It also says reasoning and agent behavior reliability are stronger than in the 3.5 series. And this is clearly aimed at real deployment, not just flashy demos. Quen 3.6 plus is going into Alibaba's own ecosystem, including Wukong, which is its enterprise platform for automating business tasks with multiple AI agents. It also works with OpenClaw, Claude Code, and Klein, so it fits into the agent workflows people are already building. On the multimodal side, Alibaba says the model can handle dense document parsing, real world visual analysis, and long video reasoning. It can also look at interface screenshots, handdrawn wireframes, and product mockups, then turn those into working front-end code. So again, you can see the pattern. The whole industry is moving away from AI as just a chatbot and more toward AI that can stay inside a task, react to what it sees, and keep working. Google just dropped one of its most interesting AI stretches in a long time. Chrome now has skills that turn prompts into reusable workflows. Gemini Enterprise is testing a new agent tab. Notebook LM is getting Canvas and connectors. DeepMind upgraded Gemini Robotics ER 1.6 for real world robotics. And Google research unveiled Vantage to score things like teamwork, creativity, and critical thinking with LLMs. Quite a bit just happened. So, let's talk about it. All right, so Google just rolled out something called Skills in Chrome built directly into Gemini. The idea is straightforward. Instead of typing the same prompt over and over again every time you open a new page, you can now save that prompt as a reusable workflow and trigger it with one action. This roll out started April 14th, 2026, and it's available on Mac, Windows, and Chrome OS as long as your Chrome language is set to EnglishUS. So, it's not global yet, but the direction is already clear. Now, if you've used Gemini in Chrome before, you've probably run into this exact problem. You open a page, you ask it to do something like analyze ingredients, compare specs, or summarize content, and then you go to another page, and you have to type the same thing again over and over. That friction is exactly what Skills removes. Instead of retyping, you save the prompt as a skill. Then later, you just type a slash or hit the plus button, select your skill, and it runs instantly on the current page. And here's where it gets more interesting. It doesn't just run on one page. It can run across multiple tabs at the same time. So now your browser basically becomes a retrieval system. You open five product pages, trigger a skill, and it compares everything in one go. That's something developers have been building manually with LLM pipelines for a while and now Google just pushed it directly into the browser UI. From a systems perspective, this is basically prompt templating at the browser level. Instead of engineers managing prompt libraries in code, regular users now get a UI version of that idea. You can also edit skills, create new ones anytime, and Google is launching a built-in library of pre-made skills. things like analyzing product ingredients, picking gifts based on constraints like budget and preferences, or scanning long documents for key info. So now you've got a curated prompt library inside Chrome itself. That's a big shift because tools like lang chain or prompt management systems used to sit behind the scenes. Now that entire concept is being exposed to everyday users. And of course there's the question of safety. Google added confirmation gates for high impact actions. If a skill tries to send an email or create a calendar event, it will ask for approval first. That's a direct response to one of the biggest challenges in agent systems, preventing automated workflows from triggering irreversible actions without user intent. Under the hood, this still runs on Chrome's existing security model with automated red teaming and auto updates. So, it fits into their broader browser infrastructure. And that leads directly into what Google is doing on the enterprise side because at the same time they're testing a new agent tab inside Gemini Enterprise. And this is where things shift from AI assistant to something that looks more like a full execution system. Inside this agent tab, you get two main entry points, new task and inbox. When you start a task, it opens a chat interface, but now there's an entire panel on the side with things like goal, agents, connected apps, files, and a toggle called require human review. So, this is no longer just a chat box. This is starting to look like a workspace for running multi-step workflows. The structure is very similar to systems like Claude Co-work. You define a goal, give the model access to tools and files, and let it execute a task across multiple steps. That require human review toggle is especially important. It suggests Google is preparing for agents that can take real actions potentially at a desktop level, not just inside a browser. And that's where things get interesting because this isn't just a feature. It looks like Google is building toward a full desktop agent environment. There's already speculation that this could tie into a future Gemini desktop app. Google is known to be working on an AI Studio desktop app and now you're seeing skills, projects, and agents all evolving at the same time. It feels like these are all pieces of a larger system. At the same time, Google is also pushing Notebook LM in a very different direction. They're testing something called Canvas inside Notebook LM. And this basically adds a visual layer on top of your data. Instead of just reading summaries, you could turn your sources into timelines, interactive pages, even lightweight apps or visualizers. So instead of just analyzing documents, you're now building structured experiences from them. There's also a new connectors feature being tested, which suggests Notebook LM will start pulling data from external services, most likely Google's own ecosystem first, but eventually more. That's a big shift because Notebook LM has mostly been limited to manually uploaded sources. So far with connectors, it starts becoming a central research layer across tools. They're also improving source organization with labeling features and even autolabeling using Gemini itself. That solves a real problem for users dealing with large data sets where navigation becomes harder than the analysis itself. Now, while all of this is happening on the software side, Google DeepMind is pushing something equally important on the robotics side. They just released Gemini Robotics ER 1.6, and this is a major upgrade to their embodied reasoning model. To understand this, you need to know how their system is structured. They use two models working together. Gemini Robotics 1.5 is the VA model, vision, language, action. That one takes inputs and directly controls the robot's movements. Gemini Robotics ER is different. It doesn't control the robot. It acts as the reasoning layer. It understands the environment, plans, tasks, and decides what should happen next. So, if the VA model is the executive, robotics ER is the strategist. And version 1.6 brings some major upgrades. First, spatial reasoning has improved significantly. That includes things like pointing, counting, and understanding object relationships. Pointing might sound basic, but it's actually foundational. It allows the model to identify exact pixel locations, map relationships between objects, define movement paths, and even enforce constraints like identifying objects small enough to fit into a container. In benchmarks, this made a huge difference. The model correctly identified objects like hammers, scissors, and tools while avoiding hallucinating objects that weren't there. That matters a lot in robotics. If a system hallucinates an object, the robot could literally try to grab something that doesn't exist. Then there's success detection, which is one of the hardest problems in robotics. It's not just about doing a task. It's about knowing when the task is actually finished. Modern robots often rely on multiple camera views, overhead cameras, wristmounted cameras, and they need to combine all of that into a single understanding of the environment. Gemini Robotics ER 1.6 six improves this multiv- view reasoning, allowing it to better handle occlusions and dynamic environments. So now the robot can decide whether to retry a task or move forward without human input. But the biggest new feature here is instrument reading. This is completely new. The model can now read analog gauges, pressure meters, sight glasses, and digital displays in real world environments. This was developed in collaboration with Boston Dynamics using their Spot robot. Spot can move around a facility, capture images of instruments, and then Gemini Robotics ER1.6 interprets them. And this is not trivial. Reading a gauge requires understanding needle positions, tick marks, units, perspective distortion, and sometimes multiple needles representing different values. The model uses something called agentic vision to do this. It zooms into images, analyzes details, runs code to estimate proportions, and applies world knowledge to interpret the result. The performance jump is massive. Gemini Robotics ER 1.5 had a 23% success rate. Gemini 3.0 Flash reached 67%. Gemini Robotics ER 1.6 reaches 86%. And with Aentic Vision enabled, it hits 93%. That's not just an improvement, that's a completely different level of reliability. Now, at the same time, Google research is working on something that looks totally different, yet it's still part of the same bigger picture. They introduced a system called Vantage, which is designed to measure human skills like collaboration, creativity, and critical thinking. And this is something that traditional tests have always struggled with. Standardized tests can measure knowledge. They can't measure how someone handles a disagreement, generates ideas under pressure, or evaluates arguments. Vantage tries to solve that using LLMs. The core idea is something called an executive LLM. Instead of running multiple independent AI agents, they use one model to control all AI participants in a conversation. That model has access to a scoring rubric and it actively steers the conversation to test specific skills. So if the system wants to evaluate conflict resolution, it might introduce disagreement through one of the AI personas and maintain that conflict until the human responds. This is very different from previous approaches. In experiments, they tested this with 188 participants, generating 373 conversations. Each participant worked through tasks like designing experiments or debating topics with AI teammates. They measured two main skills, conflict resolution and project management. The results showed that the executive LLM produced much higher evidence rates compared to independent agents. For project management, conversation level information rates reached 92.4%. For conflict resolution, it reached 85%. And when it comes to scoring accuracy, the AI matched human raiders at a level comparable to humanto human agreement with Cohen's Kappa values between 0.45 and 0.64. They also tested creativity scoring on real student work. In a data set of 180 submissions, the AI's scores had a Pearson correlation of 0.88 with human experts. That's extremely high for subjective tasks. Another interesting part is simulation. They used Gemini to simulate participants at different skill levels, then measured how accurately the system could recover those levels. The executive LLM showed significantly lower error compared to independent agents, and the simulated patterns matched real human data. That means you can use LLMs to test and refine these systems before running expensive human studies. And finally, Vantage presents results as a skills map, showing competency levels and linking them to specific parts of the conversation. So, it's not just scoring, it's interpretable. Also, if you want more content around science, space, and advanced tech, we've launched a separate channel for that. Links in the description. Go check it out. Anyway, drop your thoughts below. Curious what stands out to you the most here. Thanks for watching and I'll catch you in the next one.