Simulation + Egocentric Data: Why Physical AI Needs Both? How do we teach a robot to operate in the real world when neither simulation nor real-world demonstrations can capture every possible scenario on its own? I think this is one of the most important data challenges in Physical AI. Simulation offers scalability. We can generate diverse scenes, vary object positions, modify environmental conditions, and reproduce scenarios that would be expensive or time-consuming to collect in the physical world. It provides a controlled environment for generating large volumes of structured training data. But simulation has its limitations. Even a sophisticated simulator cannot perfectly reproduce every nuance of real-world interaction, human behaviour, or unexpected physical conditions. This is where egocentric data becomes valuable. First-person recordings of human activity can provide rich information about how people interact with objects, navigate environments, and perform everyday tasks. Rather than relying exclusively on synthetic scenarios, models can learn from demonstrations grounded in real human behaviour. Yet collecting real-world data at scale introduces its own challenges, including cost, coverage, consistency, and annotation. Neither source is sufficient on its own. Simulation can expand the range of scenarios we explore. Egocentric data can provide complementary evidence of how tasks unfold in the real world. The opportunity lies in combining their strengths while addressing their respective limitations. This is why I find the approach of @axisrobotics interesting: its broader data infrastructure brings simulation-based collection and egocentric data into the same Physical AI conversation. The goal is not simply to accumulate two different datasets. It is to create a more comprehensive foundation for training models that must operate beyond controlled demonstrations. My takeaway: Simulation helps us scale the scenarios we can generate. Egocentric data helps us learn from the world humans actually navigate. The next challenge is making these different sources complementary, consistent, and useful for model training. Because building Physical AI requires more than teaching robots what success looks like. It requires preparing them for the complexity of the world in which they must succeed. Connect X with @KaitoAI / Yaps: s.kaito.ai/jd7nwhO

Physical AI, built by all and owned by all. The time you give your rose is what makes it precious. Your contributions built the Physical AI data engine. $AXIS is now LIVE.

Simulation + Egocentric Data: Why Physical AI Needs Both? How do we teach a robot to operate in the real world when neither simulation nor real-world demonstrations can capture every possible scenario on its own? I think this is one of the most important data challenges in Physical AI. Simulation offers scalability. We can generate diverse scenes, vary object positions, modify environmental conditions, and reproduce scenarios that would be expensive or time-consuming to collect in the physical world. It provides a controlled environment for generating large volumes of structured training data. But simulation has its limitations. Even a sophisticated simulator cannot perfectly reproduce every nuance of real-world interaction, human behaviour, or unexpected physical conditions. This is where egocentric data becomes valuable. First-person recordings of human activity can provide rich information about how people interact with objects, navigate environments, and perform everyday tasks. Rather than relying exclusively on synthetic scenarios, models can learn from demonstrations grounded in real human behaviour. Yet collecting real-world data at scale introduces its own challenges, including cost, coverage, consistency, and annotation. Neither source is sufficient on its own. Simulation can expand the range of scenarios we explore. Egocentric data can provide complementary evidence of how tasks unfold in the real world. The opportunity lies in combining their strengths while addressing their respective limitations. This is why I find the approach of @axisrobotics interesting: its broader data infrastructure brings simulation-based collection and egocentric data into the same Physical AI conversation. The goal is not simply to accumulate two different datasets. It is to create a more comprehensive foundation for training models that must operate beyond controlled demonstrations. My takeaway: Simulation helps us scale the scenarios we can generate. Egocentric data helps us learn from the world humans actually navigate. The next challenge is making these different sources complementary, consistent, and useful for model training. Because building Physical AI requires more than teaching robots what success looks like. It requires preparing them for the complexity of the world in which they must succeed. Connect X with @KaitoAI / Yaps: s.kaito.ai/jd7nwhO

10M Trajectories ≠ Endgame 10 million robot trajectories sounds impressive. And it is. But I don't think 10 million should be the end goal for Physical AI data. Because there is a fundamental difference between: having more data and having more useful data. Imagine two datasets. Dataset A has 10 million trajectories. Dataset B has 1 million trajectories specifically collected around the model's weakest capabilities. Which one creates more learning value? The answer isn't determined by the number. It depends on what those trajectories teach the model. This is where the scaling narrative around Physical AI becomes more nuanced. At the beginning, increasing data volume is critical. You need diversity. You need different environments. You need different objects, behaviours, failures and edge cases. But once the infrastructure can generate data at scale, another bottleneck emerges: data selection. What should be collected? Where is the model still weak? Which failures are worth correcting? Which trajectories have become redundant? And perhaps most importantly: What should the next 10,000 trajectories look like? This is why I find the direction of @axisrobotics $AXIS interesting. The long-term opportunity isn't simply to build a massive repository of robot trajectories. It's to create a feedback system where: Model performance → reveals weaknesses → guides data generation → improves the model → reveals new weaknesses. That creates a fundamentally different kind of scalability. Not just: More data → Bigger dataset but: Better feedback → Better data → Better model → Better feedback So yes, 10M trajectories is a milestone. But it shouldn't be mistaken for the destination. The real endgame isn't collecting every possible trajectory. It's building a system that continuously knows which trajectories are worth collecting next. In Physical AI, scale gets you to the frontier. Intelligent data selection determines how far you can move it. Connect X with @KaitoAI / Yaps: s.kaito.ai/jd7nwhO

One of the most interesting questions in Web3 is whether transparency should mean that absolutely everything must be visible. Blockchains need public verification to establish ownership and enable trustless interactions. However, making every piece of information public can create unnecessary limitations for applications that depend on private data. @fungolabs approaches this problem through encrypted NFTs built with Zama’s FHE technology. What makes this concept compelling is the distinction between public ownership and private content. An NFT can exist on a public blockchain without requiring every characteristic associated with it to be publicly readable. This is a more nuanced approach than simply choosing between complete transparency and complete secrecy. Imagine digital assets with hidden characteristics, games where players do not automatically expose their strategies, or applications in which information can be used without being disclosed in its original form. These possibilities are worth exploring because privacy should not have to come at the expense of useful on-chain functionality. For me, the real opportunity is not to make blockchain less transparent. It is to make transparency more selective, purposeful, and compatible with user choice. That is the direction that makes @fungolabs worth watching.

I think Fungo Labs is exploring one of the most interesting directions in the NFT space: making privacy a fundamental part of digital ownership rather than an afterthought. Traditional NFTs are built around transparency, which is valuable for verification but can also expose information that owners may prefer to keep private. Fungo Labs challenges this limitation by combining encrypted NFTs on Ethereum with Zama’s Fully Homomorphic Encryption technology. What interests me most is the potential behind this approach. When NFT traits and attributes can remain encrypted while still supporting computation, digital assets may become more than collectibles. They could enable new gaming mechanics, strategic interactions, and applications where privacy and verifiable ownership coexist. I also appreciate that Fungo Labs is looking beyond a single NFT collection, with plans for an encrypted NFT launchpad and a game built around its Genesis NFTs. If these ideas are implemented successfully, they could help demonstrate the practical value of encrypted assets to both creators and users. Of course, the technology itself is only part of the equation. Adoption, usability, security, and real utility will ultimately determine how far the project can go. Overall, I see Fungo Labs as an ambitious experiment in rethinking what NFTs can offer. Instead of asking people to accept complete transparency as the default, it explores a model where ownership remains verifiable while sensitive information stays private. For me, that is a meaningful problem to solve, and I am interested to see how the team turns this vision into real-world applications. @fungolabs

Own your cells. Harvest their energy. Early entries are being recorded: projectsphere.xyz

Humanity has made remarkable progress in science and technology, yet our ability to use energy remains one of the fundamental constraints on further development. @rarelistio x @projectspherexy This is where the Kardashev scale becomes interesting. It classifies civilizations according to the amount of energy they can harness, with a Type II civilization theoretically capable of utilizing the energy output of its entire star. Project Sphere takes inspiration from this ambitious concept and directs attention toward the idea of civilizational acceleration. What makes this topic worth exploring is not the assumption that a Dyson sphere will be built tomorrow, but the opportunity to examine the scientific, engineering, and technological breakthroughs that such a future would require. For me, the most valuable projects are those that encourage people to look beyond immediate trends and engage seriously with a larger vision.

One star was only the beginning. The Kardashev scale ranks civilizations by the energy they can use. Type I: planetary Type II: stellar Type III: galactic Project Sphere follows that path from system to system. Each new star becomes another Dyson sphere, built cell by cell.

One star was only the beginning. The Kardashev scale ranks civilizations by the energy they can use. Type I: planetary Type II: stellar Type III: galactic Project Sphere follows that path from system to system. Each new star becomes another Dyson sphere, built cell by cell.

10M Trajectories ≠ Endgame 10 million robot trajectories sounds impressive. And it is. But I don't think 10 million should be the end goal for Physical AI data. Because there is a fundamental difference between: having more data and having more useful data. Imagine two datasets. Dataset A has 10 million trajectories. Dataset B has 1 million trajectories specifically collected around the model's weakest capabilities. Which one creates more learning value? The answer isn't determined by the number. It depends on what those trajectories teach the model. This is where the scaling narrative around Physical AI becomes more nuanced. At the beginning, increasing data volume is critical. You need diversity. You need different environments. You need different objects, behaviours, failures and edge cases. But once the infrastructure can generate data at scale, another bottleneck emerges: data selection. What should be collected? Where is the model still weak? Which failures are worth correcting? Which trajectories have become redundant? And perhaps most importantly: What should the next 10,000 trajectories look like? This is why I find the direction of @axisrobotics $AXIS interesting. The long-term opportunity isn't simply to build a massive repository of robot trajectories. It's to create a feedback system where: Model performance → reveals weaknesses → guides data generation → improves the model → reveals new weaknesses. That creates a fundamentally different kind of scalability. Not just: More data → Bigger dataset but: Better feedback → Better data → Better model → Better feedback So yes, 10M trajectories is a milestone. But it shouldn't be mistaken for the destination. The real endgame isn't collecting every possible trajectory. It's building a system that continuously knows which trajectories are worth collecting next. In Physical AI, scale gets you to the frontier. Intelligent data selection determines how far you can move it. Connect X with @KaitoAI / Yaps: s.kaito.ai/jd7nwhO

Z-Axis unlocked Sau một thời gian dài đóng góp cho cộng đồng thì mình cũng đã đạt được vai trò Z-Axis. Và phải nói thật, @techboo_ đã biến khoảnh khắc này thành một sản phẩm quá thú vị. Một Axis Contributor Personality Card riêng, với mình là Core Processor - “The big-brain OG running Axis's core” Visual vừa đẹp, vừa có chút personality, lại capture được cả hành trình contributor. Next level loading… 🚀 Big thanks to @techboo_ for creating this. Love it! 🔥 @axisrobotics $AXIS

Attention has quietly become one of the most valuable assets in the crypto ecosystem. Projects need attention to grow. Creators need attention to distribute ideas. Communities need attention to attract contributors. Users need attention to identify opportunities before they become crowded. The interesting question is what happens when these three groups can participate in the same ecosystem. That is one reason I am watching @rarelistio. The creator campaign model suggests that content does not have to be treated as meaningless promotion. Ideally, creators become part of the discovery layer itself: they research a project, explain why it matters, communicate it to their audience, and create a conversation around it. That is a much healthier model than simply asking everyone to repeat the same announcement. If Rarelist wants to build a long-term creator ecosystem, the quality of the conversation will matter just as much as the quantity of posts. Yaps: rarelist.io/r/pertherEx