Xiaomi Humanoid Robot w/ 8200 mm Tactile Sensing SWEATS + Anthropic Leaks Mythos (AI NEWS)

In fact, this upgrade transforms what was already an ambitious humanoid into what could become a paradigm shift in robotic labor. So here’s what changed. Firstly, it’s 60% volume reduction. So the entire hand assembly has been shrunk down dramatically, achieving near-perfect one-to-one scale with an average adult human hand. And this human-sized form factor is critical because it allows much easier transfer of human-teleoperation data, and it makes the robot far more compatible with existing tools, workstations, and environments that are designed for people. The second change is its full-palm tactile sensing. With over 8200 square millimeters of tactile sensors now covering its fingertips, finger pads, and the entire palm, this robot can feel texture, pressure, and even slips when its vision is blocked or it’s in low-light conditions.

And this is a huge advantage for delicate or occluded manipulation tasks. Third, there’s been an endurance breakthrough, where the hand has been endurance tested through more than 150,000 continuous grasping cycles, over a 61-hour non-stop test, and earlier prototypes often failed here, right around the 10,000 cycle mark. So this leap in durability is essential for 24-7 production line use. And finally, its degrees of freedom. Now, the hand features an increase of 50% in total degrees of freedom and an 83% increase in active degrees of freedom, which enables fluid, human-like movements for precise tasks like turning screws, pinching small objects, or handling fragile items.

But one of the most striking engineering details is the bionic sweat gland cooling system. Using 3D printed liquid channels and evaporation, the hand now actively dissipates up to 10 watts of heat, mimicking how human palms sweat to regulate temperature during intense work. And this prevents overheating during long shifts and directly contributes to the massive jump in the life cycle. And Xiaomi has already begun piloting the upgraded Cyber One hand in its own electric vehicle assembly plants, where the robot is currently tackling repetitive and precision-heavy tasks, with the CEO stating that their goal is to have humanoids working at a large scale inside of their Xiaomi factories within the next five years.

And the timing is no coincidence here, because by integrating its own humanoid technology into its manufacturing lines, Xiaomi isn’t only solving the labor challenges, but also its creating a closed-loop testing ground to rapidly iterate on both hardware and embodied AI software. And with this new human scale, tactile-rich, heat-managing hand, Cyber One is moving beyond flashy demos and into the realm of production-grade reliability, as the robot can now perform intricate assembly steps, adapt to uncertain conditions through touch, and even keep working hour after hour without thermal throttling. And while commercial deployment details remain under wraps, this latest upgrade signals that Xiaomi is treating humanoid robotics with the same aggressive, vertically integrated approach that turned it into the smartphone and EV powerhouse that it is today.

But while Xiaomi is making robots that exceed us physically, what happens when AI exceeds us mentally? That’s the question hanging over the industry right now, as what started as an internal slip at Anthropic exploded into one of the most revealing AI leaks in recent memory. That’s because around March 26, a misconfigured content management system left roughly 3,000 images, PDFs, audio files, draft documents, and videos, all publicly searchable and downloadable by anyone who stumbled across them, with no authentication required. So Fortune first spotted the exposure and alerted Anthropic, which afterwards they swiftly took it down, calling it a straightforward human error configuration mistake.

But buried in that the other used a placeholder saying Claude Capybara, but both described the identical breakthrough of a finished model that Anthropic calls, quote, by far the most powerful AI model that they’ve ever developed, which is a true step change beyond anything in their lineup. And it’s explicitly pitched as a new tier, which is larger and more intelligent than their Opus models, which were until now, the most powerful. So the leap in mythos doesn’t seem to be incremental, as it’s described internally as dramatically higher performance across software coding, academic reasoning, and cybersecurity benchmarks.

But after some independent analysis of the leaked benchmarks and early API signals, an even more vivid picture was painted as to what the step change actually unlocks in practice. So on coding tasks, mythos isn’t just faster, it excels at multi file code generation, reliable debugging across entire code bases, reasoning through unfamiliar or messy projects, and the tackling of complex software engineering challenges that leave today’s Opus model struggling. And when it comes to reasoning benchmarks, it showed that it dominates multi step problems, handles ambiguous or incomplete information, and can draw logical conclusions under uncertainty, while sustaining extended chains of thought far beyond previous limits.

And in terms of cybersecurity, the model stands out for identifying vulnerabilities with precision, dissecting malicious code, and can navigate intricate security scenarios with capabilities so advanced that leaked drafts themselves flag them as being a double edged sword. And in anthropics own wording in the drafts, the model was quote, currently far ahead of any other AI model in cyber capabilities, which could foreshadow a coming wave of systems that could exploit vulnerabilities faster than defenders could patch against them. And this is probably why anthropic is proceeding with such caution. Despite their training being complete, early access testing is already underway with only a select group of customers prioritizing cybersecurity applications so defenders can start hardening their systems now.

And a full public rollout is being handled carefully with the model being compute heavy, and quite expensive to run. And anthropic wants extra time to map out the risks, even beyond what they learned in their own testing. And this leak also pulled back the curtain on supporting materials for an exclusive invite only two day CEO summit at a Manor Hotel in the English countryside, with European executives being set to hear directly from anthropic CEO about the unreleased capabilities of mythos, which is another clue that the company is laser focused on enterprise trust, and high stakes adoption while keeping the most potent version under wraps.

And the company’s constitutional AI framework remains front and center embedding safeguards directly into training to manage the dual use tension in areas like cybersecurity without artificially capping the model’s potential. So as of now mythos remains unreleased to the public with no official product page, no API access for most users. But anthropic has confirmed that it’s already in controlled testing. Anyways, like and subscribe and check out this video here for more of the latest in AI and robotics news. [tr:trw].