看完 PrismaX 发的这场 RoboCon panel，我印象比较深的是“模型得能放进机器人里跑”这件事。 平时看模型演示，很容易忽略它背后用了什么设备。但真要交给客户，算力、延迟、功耗都得考虑。在服务器上表现不错，到了机器人本体里还能不能跟得上，这是另一道关。 这也是为什么我觉得，机器人部署很难只靠再训一版模型解决。模型要调整，设备要配合，现场出了问题，还得先弄清楚到底是哪一部分出了错。 比如一个抓取动作失败了，是没看准物体，动作预测不合适，还是夹爪打滑？视频里看起来可能都是“没抓住”，但后面要改的东西完全不同。定位错了，反复训练也可能没补到问题上。 客户那边更现实。他们要的是机器人每天能把活干完，出了问题有人能及时处理。如果每次失败都要等工程师慢慢排查，原本想省下的人力，可能又花回去了。 所以我会更看重团队在真实部署里积累的经验。同样一个问题，第一次要折腾几天，后来能不能更快找到原因？这次修好的地方，下次换设备、换场地还能不能用上？ PrismaX 把这些问题放进 RoboCon 的讨论里，我挺喜欢。看一个机器人团队，除了它现在能做什么，也值得看看它碰到麻烦之后怎么处理。客户愿意长期用下去，往往就靠这些不太出现在演示视频里的工作。 @PrismaXai

If I were running a node, I’d look at three things before using Optimum: 1. Can I get more out of my existing bandwidth? Sending the same message repeatedly across different paths consumes network resources. Optimum exchanges coded fragments to reduce duplicate full-message transmissions, potentially reaching more nodes with the same amount of data sent. If that holds up in practice, buying more bandwidth wouldn’t have to be the first response to every delivery bottleneck. 2. How much would I need to change? Optimum already has a Gateway connecting existing Ethereum consensus clients to its acceleration network. Clients can use it by peering with the gateway, without changes to their code. For teams already running infrastructure, that provides a concrete way to test it. 3. Does it actually improve performance? How much faster is propagation? Has bandwidth overhead fallen? Is running an additional gateway worth it? Operators can compare results before and after connecting. I like products that can be evaluated within an existing operation. Operators can test the service and keep it if the results justify it. If Optimum gets these three things right, it has a chance to become a service node operators choose to keep running. @get_optimum @MurielMedard @CryptoSundayz @shariaronchain @aqccapital @ada_pegasus

A payment can move in seconds. Legal title usually cannot. That gap is one of the most honest problems in RWA. You can settle a token transfer quickly, but the underlying asset may still involve registries, custodians, contracts, local records, and people who need to sign off. Pretending that blockchain makes all of that disappear is how you create a nice-looking dashboard with weak real-world foundations. The harder job is syncing fast digital settlement with slower legal reality. SimpleChain’s focus on title confirmation, rights, and smart legal execution speaks directly to that tension. The token can move fast. The claim behind it still has to stand up when somebody asks: “Who legally owns what?” That is where RWA either becomes serious finance or stays a demo. @SimpleChain_RWA

模型能直接调 API 了，机器人不会干的活，还是得有人教。 Pi API 把微调做方便了，但上机一试，哪里做不好，还得补演示、再训练。换个任务，又要继续折腾。 PrismaX 紧接着说，他们一直在为这个过程做准备，更多消息很快公布。这就让我有点期待了。 之前看他们做机器人数据，容易理解成“有需要就买一批”。但一个任务可能要试好几轮：换个物品就抓不稳，或者前面都顺利，最后一步总出错，都可能需要新的演示。 如果 PrismaX 能接住这些需求，客户就有理由反复回来。合作几轮之后，操作员熟悉设备，也知道团队要什么，后面就不用每次从头沟通。对赶着做测试的人来说，这能省不少事。 我觉得这个时间点挺好。模型更容易用了，更多团队才有机会动手试，后面的数据需求也可能跟着出来。 具体产品还没公布，但这次更新我会蹲一下。要是能把补采、测试、再训练这些事接顺，PrismaX 做的就不只是一批数据的生意了。 @PrismaXai

I find Optimum’s technology easier to understand by asking: what happens when data doesn’t arrive as expected? Its use of RLNC has a few useful properties: 1. A lost packet doesn’t always need to be retrieved Data is encoded into different combinations. Once a receiving node collects enough independent combinations, it can reconstruct the original content without waiting for a specific packet to be retransmitted. 2. Nodes can forward data before receiving the whole block Intermediate nodes can recombine the coded fragments they’ve received and send them onward without first reconstructing the full block. Receiving and forwarding can overlap, reducing some of the waiting between hops. 3. Data from multiple connections can contribute to recovery A receiver can collect coded fragments from different nodes. As long as they add independent information, they help reconstruct the data and reduce reliance on any single connection. For block propagation, the practical benefit is clear: when packets are lost or a connection slows down, nodes have other ways to collect enough data. That’s the part I like. Nodes around the world won’t all have equally good connections. Helping those with less reliable networks receive blocks sooner would be a useful improvement. @get_optimum @MurielMedard @CryptoSundayz @shariaronchain @aqccapital @ada_pegasus

Optimum’s latest interview covered quite a bit beyond technology: talent, regulation, and whether people are willing to try something new. At first, that sounded a little broad to me, but these things really do matter. If I were responsible for integrating a data acceleration service at a company, I’d be curious to test it. But who handles the integration? How long will it take? Who do we call if something breaks? Until those questions are answered, even impressive performance numbers might end up on the back burner. Those are exactly the kinds of conditions Muriel Médard highlighted when discussing Singapore. Having technical talent, clearer rules, and people willing to experiment makes it easier to move things forward. That gives me some optimism about Optimum’s prospects there. The research behind RLNC is enough to spark interest. The next step is giving potential users a chance to try it themselves and tell the team where it falls short. For an infrastructure project, finding people willing to test the product seriously matters. A lot of the small headaches that come with integration are hard to anticipate from inside the team. I’d be more interested in hearing users explain why they’ve kept using Optimum than in seeing another set of speed benchmarks. @get_optimum @MurielMedard @CryptoSundayz @shariaronchain @aqccapital @ada_pegasus

If a block arrives quickly most of the time but occasionally takes much longer, how fast is the network really? For validators, a good average isn’t enough. Their duties have deadlines. Receiving a block late at a critical moment isn’t something earlier time savings can make up for. That’s why, when looking at Optimum, I’m especially interested in whether it improves the slowest deliveries. Its use of RLNC has a useful property: after receiving some coded fragments, intermediate nodes can recombine and forward them without first decoding the full block. Receivers can collect enough independent combinations from different nodes to reconstruct the original data. This reduces dependence on any particular fragment arriving and offers another way to handle packet loss and slow connections. The actual improvement still depends on network conditions and implementation. But helping slower nodes receive blocks more consistently and on time could make a practical difference to their attestation duties. That’s what keeps Optimum on my radar. More predictable block propagation could provide additional headroom when considering shorter slots or larger blocks. Execution and validation still need to keep up, but improving data delivery would move one important piece forward. @get_optimum @MurielMedard @CryptoSundayz @shariaronchain @aqccapital @ada_pegasus

RWA has a scale problem that more listings will not solve. If every asset arrives with its own language, its own documents, its own event format, and its own definition of “ownership,” then every new deal starts from zero. That is not a market. That is a collection of one-off projects. The real network effect in RWA may come from standards. Not more tokens. A common way to describe rights. A common way to record cash flows. A common way to handle changes. A common way to prove what happened. SimpleChain calls this an asset standard, built around things like provenance, executable rules, reconcilable cash flow, and lifecycle continuity. That sounds abstract until you realize what it changes: The second asset becomes easier to process than the first. Then the hundredth becomes easier still. That is how infrastructure compounds. @SimpleChain_RWA @chx99999

很多做 Physical AI 的团队卡在数据上，不是技术做不了，而是获取数据的流程太麻烦。 想买一批真实机器人的操作演示，往往要先约销售、讲需求、等报价，再走内部审批。大厂或许习惯了这些流程，但研究者和小团队只是想先做个实验，来回沟通就可能耗掉不少时间。 PrismaX 这次在 IROS 展示的做法比较直接：不用销售电话，也不用等候名单，注册后就能选择数据集下载。熟练操作员通过遥操作，在真实机器人上采集的复杂任务演示，现在可以直接获取。 我觉得这一步挺有意义。机器人数据难获取，除了采集成本高，采完之后流通不方便也是一个问题。把获取门槛降下来，能让更多团队先拿到数据，把实验跑起来。 小团队本来就资源有限，不一定有条件从零搭建采集流程。如果每次尝试都要先花时间找人、谈需求、等回复，很多想法可能还没验证就被搁下了。 PrismaX 让这个过程简单了一些。先拿数据试试，再决定要不要继续投入，对研究者和小团队来说，这种选择很实在。 @PrismaXai

On the final night of Token 2049, Optimum hosted “Brews by the Bay” with Auros, Canton Network, and Spartan. That feels like a good mix for this team. Trading, institutional blockchain infrastructure, and investment bring different priorities, but each offers a useful perspective on faster blockchain data delivery. Optimum uses RLNC to improve data propagation, reducing duplicate transmissions so nodes can receive blocks and critical data faster and more reliably, with less bandwidth overhead. When reading technical updates, it’s easy to remember only the “how many times faster” figure. But trading teams care about when information reaches them, node operators care about validator performance, and blockchain teams also have to consider integration costs. Conversations with these groups can connect technical gains to specific needs. I like Optimum’s approach: improve existing networks without asking everyone to migrate. Chains and applications keep running while Optimum works on how data moves behind the scenes. This event makes me curious about which use cases they’ll test next. Latency-sensitive applications seem like a good place to see how much difference it makes. @get_optimum @MurielMedard @CryptoSundayz @shariaronchain @aqccapital @ada_pegasus

A new block has been produced. That doesn’t mean every node has seen it yet. When people talk about blockchain speed, they tend to focus on block times and consensus. But once a block is produced, it still has to travel across the network. How quickly that happens matters too. Ethereum’s gossip mechanism works a bit like passing a message from person to person: a node receives information, forwards it to other nodes, and it gradually spreads. Along the way, the same data can be transmitted repeatedly, while lost packets need to be recovered. This is where Optimum comes in. It uses random linear network coding (RLNC) to split data and encode it into different combinations. Once a receiving node collects enough independent combinations, it can reconstruct the original data without waiting for a particular missing packet to be retransmitted. According to Optimum, its Ethereum mainnet tests showed roughly 2.4x faster block propagation and an approximately 8x improvement in arrival-time consistency. I like this approach. Developers don’t have to migrate, and there’s no need to build another chain to unlock more performance from an existing network. For blockchains already supporting substantial applications and assets, that has practical appeal. The communication layer rarely gets much attention, yet almost everything happening onchain depends on it. If Optimum can keep improving this layer, its potential reach extends well beyond any single type of application. That’s what makes it worth watching for me. @get_optimum @MurielMedard @CryptoSundayz @shariaronchain @aqccapital @ada_pegasus

在 IROS 2026 上，PrismaX 把 π0.5 直接拉到匹兹堡现场做了 sight-unseen 演示。臂的位置基本是用眼睛估的，没怎么认真标定，结果还能稳定跑。真正敏感的反而是安装高度——差 1 厘米，成功率大概就掉 70%。 他们能跑起来，主要还是靠数据质量和推理速度。大概用了 1K 条高质量真机数据做微调，数据本身是 1 倍速采的，没什么停顿和错误。推理用了自己写的 TensorRT 栈，单步大概 87 ms，在展会那种乱糟糟的环境里也能保持比较连贯。基线够强的时候，换个环境掉一点性能也还能接受。 这件事有意思的地方在于，很多团队还在死磕标定和硬件刚性。PrismaX 这次说明，对于现在这类 VLA 模型，高度这种直接影响工作空间和视野的变量，可能比水平方向的小偏差更要命。把数据质量和推理速度做上去，对降低标定依赖是有帮助的。 当然，差 1 厘米就掉那么多，也说明现在的策略离真正对安装不敏感还远，基本工位一致性还是得保证。但相比把标定流程做到极致，优先把数据质量和推理速度拉上去，或许更务实一点。他们下周要放完整细节和代码，到时候可以看看具体怎么做的。 @PrismaXai

$48.7B in tokenized stock volume is also a workload that onchain infrastructure has to carry. According to Token Terminal, tokenized stocks on DEXs generated that volume over the past year, up more than 10,000% year over year. QQQb alone accounted for $14.1B. Trading volume isn’t a direct measure of network traffic. But as more assets attract active trading, it raises a practical question: how much infrastructure will it take to support that growth? Bandwidth is part of the answer. Distributing data across a decentralized network consumes resources, and repeatedly forwarding the same information adds overhead. That’s the part of Optimum’s work that stands out to me. Optimum uses random linear network coding (RLNC) to make data distribution more efficient. In testing against gossipsub, it reported roughly 90% lower bandwidth usage, alongside faster block and data propagation. The bandwidth result deserves attention in its own right. If those gains hold in production, networks could have more room to handle growing activity within their existing capacity. Tokenized equities tend to make headlines through new listings and volume milestones. As the market expands, the resources required to support it will become a bigger part of the conversation. @get_optimum @MurielMedard @CryptoSundayz @shariaronchain @aqccapital @ada_pegasus

🎙️ Optimum Community Call Recap | Oct 8 Mainnet will launch before FlexNodes. Important updates are expected in the coming weeks 👀 The team wants the community to focus on understanding three things: 🔹 RLNC (Optimum’s core underlying technology) 🔹 How Validators operate and generate revenue 🔹 Scalability Inactive members are being removed in stages. Regardless of previous roles or ranks, inactive accounts will be cleared. Contribution matters more than titles. Community activities are also picking up: 🎨 Encouraging high-quality handmade and digital art (that is both entertaining and educational) ♟️ Mini tournaments such as poker and chess are coming soon ⚡ Short mini-events lasting from 1 day to 1 week (running alongside larger campaigns) No Community Call next week due to the host’s schedule. It returns the week after with more important updates. Special thanks to @dcityuknow for developing the Block Arena mini-game for the community 🎮 Looking forward to the call the week after next. @get_optimum @MurielMedard @CryptoSundayz @shariaronchain @aqccapital @ada_pegasus

There is a version of RWA that worries me. People start collecting tokenized assets the way they collected NFTs. A bit of real estate. A bit of private credit. A bit of carbon. A bit of whatever is trending that week. That is not portfolio construction. That is just a more complicated form of shopping. The useful future is different. People should be able to think in exposures: How much income do I want? How much duration can I accept? How much risk belongs in this part of my portfolio? How do different assets behave together? That is where products such as RWA Earn and structured credit become more interesting than a simple asset listing. The endgame is not owning more tokens. It is giving people better ways to allocate capital. That is a much harder product problem, and a more meaningful one. @SimpleChain_RWA @chx99999

IROS上挤满灵巧手，Sharpa、WUJI、DH这些都在卷五指20+自由度，参数看起来差不多。波士顿动力没来展台，最后一天直接放出Atlas新手：四指、13自由度。CTO让团队把小指和无名指绑起来试了一天，确认多出的执行器不值得，就砍掉了。 手已经不难做了。继续堆自由度边际收益在快速下降，多一个手指就多三套执行器和故障点，学习难度也跟着上去，实际任务里很多动作根本用不到。减少零件反而更利于稳定和从仿真迁移到现实。 会场上真正聊得多的已经不是手本身。做手的公司同时在推采集手套和外骨骼，数据公司也明显比往年多。问的人不再只是“你有多少小时”，更多在问这些数据是不是针对具体任务设计的。硬件容易买，但谁能持续拿到高质量、任务相关的演示，谁才有机会把模型做到能在没人提前布置的房间里稳定跑完第1000次。 PrismaX在展位跑的π0.5抓水果，数据是在别的工作站采的，到完全没见过的会场直接部署，也没精细标定。能跑起来，是因为事先把真正影响成功率的几件事筛出来了：整模型微调、相对动作、安装高度差1厘米成功率就掉很多、延迟必须压住。 手的市场已经很拥挤。接下来拉开差距的，大概率不在于谁的手指更多，而在于数据质量和那些容易被忽略的部署细节。PrismaX做的就是这件事——把高质量、可复用的真实演示数据做成基础设施，让模型真正能在陌生环境里稳下来。 @PrismaXai

In the discussion with @ambr_io in Singapore, one point kept coming up: The real constraint in agentic trading has shifted from “whether the model is smart enough” to “whether the data can reach the decision layer quickly and reliably.” When AI agents start handling research, order placement, risk control, and execution, they are no longer just assistive tools. They become direct consumers of real-time state. A delay of a few dozen milliseconds, incomplete state updates, or inconsistent propagation across nodes can turn directly into wrong positions or missed windows. At that point, the speed and reliability of the data layer stop being an infrastructure optimization and become a precondition for whether a strategy can even work. Optimum treats this as the core problem to solve: using an RLNC-based acceleration network (mump2p) to cut propagation latency and improve resilience to packet loss, so the data agents rely on arrives faster and more completely. The line on screen, “Speed is money,” is not a slogan. In the agent era, the time it takes for information to arrive is itself part of the return. Agents can make their own decisions, but the upper bound on decision quality is still set by the quality of the data they see. @get_optimum @MurielMedard @CryptoSundayz @shariaronchain @aqccapital @ada_pegasus