A typical day for me: Wake up Check Decibel Discord + X Brush my teeth Eat breakfast Check Decibel app + Discord Eat lunch Check Decibel app Eat dinner Check Decibel app again Go to sleep And somehow… I still end up dreaming about Decibel. 😂 @DecibelTrade

BTC is getting close to the level I’m watching most right now: $83K. After the strong move from the mid-$76K area, BTC is now trading around $81K and has already pushed above $81.7K today. For me, the key level is still $83K. If BTC can break and, more importantly, hold above $83K, I’d be watching the next major psychological target around $90K. The setup I’m watching: → $83K = key resistance → Break + hold above $83K = potential continuation → $90K = next target → If $83K rejects, I’d rather wait for a retest than chase the move I’m not saying BTC will definitely reach $90K. I just think $83K is the level that could tell us whether this move has another leg to run. Let’s see how BTC handles it. This is for reference only. DYOR. @DecibelTrade

It feels amazing to be able to take Decibel with me wherever I go 😍 @DecibelTrade @Decibel_VN

I think it's time for us to look back at what we've accomplished over the past few months. I'll go first. Rank # 574 and 3.31k AMP. Show me yours 😍 @DecibelTrade

The moment a human takes control of a robot may be one of the most valuable moments in the entire dataset. Imagine a robot performing a task correctly for several minutes. Most of that trajectory tells us something the policy already knows how to do. Then the robot approaches an object from the wrong angle, begins making a bad decision, and a human operator intervenes before the task fails. That intervention contains unusually dense information. It tells us where the autonomous policy stopped being trusted, what state caused the problem, and what action a human believed should happen instead. In other words, intervention creates a natural learning signal: Robot acts → Uncertainty appears → Human intervenes → Corrective action → Robot continues The interesting part is the transition between autonomous behavior and human correction. Before intervention, we see the policy moving toward a mistake. During intervention, we see an alternative action. After intervention, we can observe whether that correction successfully returned the system to a useful state. This makes intervention data different from ordinary demonstrations. A demonstration teaches the robot how a human performs a task from the beginning. An intervention teaches the robot specifically what the human does when the robot is about to get something wrong. That difference matters because as robots improve, full demonstrations may become increasingly redundant. If a robot already performs 95% of a task correctly, having a human demonstrate the entire task again spends most of the collection time recording behavior the robot already understands. A more efficient approach is to let the robot operate autonomously and concentrate human effort around the remaining 5%. Autonomy → Human correction → Learn the correction → More autonomy As the policy improves, interventions should become less frequent. But the interventions that remain may become increasingly informative because they expose harder and rarer weaknesses. This suggests a different way to think about human involvement in robot learning. The goal is not simply to remove humans from the loop as quickly as possible. It is to use human attention precisely where the robot’s own behavior tells us it is most valuable. Human intervention is not only a safety mechanism. It can also be a data-generation mechanism. @axisrobotics

A single image can tell a robot what the world looks like. It cannot always tell the robot what the world is doing. Imagine a robot holding a cup. From one observation, the cup may look perfectly stable. But perhaps it started slipping 200 milliseconds ago. Maybe the robot just increased its grip force. Maybe the cup collided with another object and is still moving. The current state can look almost identical while requiring completely different next actions. This creates a temporal context problem. Physical intelligence is not only about understanding state, but also understanding change over time. Consider two observations of the same object position. In one sequence, the object has been stationary for several seconds. In another, it has just moved rapidly into that position. Visually, the final frame can be nearly identical. Physically, they represent different situations because velocity, momentum and recent contact history matter. So robot learning cannot always be reduced to: Observation → Action A richer representation looks more like: Past States → Current State → Direction of Change → Next Action This matters for manipulation because many important signals are temporal. A grasp becoming unstable, an object beginning to rotate, resistance increasing while opening a drawer, or a tool making contact with a surface can only be fully understood by comparing what is happening now with what happened moments earlier. It also changes how we think about robot data. Ten isolated images of a task are not equivalent to one continuous trajectory containing ten observations. The trajectory preserves transitions, timing, causality and the consequences of previous actions. That information teaches something static snapshots cannot: how the physical world evolves. This may be one reason trajectory data is so important for embodied intelligence. Robots do not operate in collections of independent frames. They operate inside continuous processes where every action changes what happens next. The world is not a sequence of pictures. It is a sequence of consequences. @axisrobotics

Not every robot interaction is equally valuable. Imagine a robot has already picked up the same cup from roughly the same position 10,000 times. The 10,001st successful grasp still adds data, but it probably teaches the model much less than the first time it encounters a transparent cup, an unusual angle, a slippery surface, or a grasp it has never recovered from before. This is a useful way to separate data volume from data value. Data volume asks: “How many interactions did we collect?” Data value asks: “How much did this interaction reduce what the robot still doesn’t know?” Early in training, almost everything can be informative because the model has seen very little. As coverage grows, repeated examples begin to overlap with knowledge the model already has. The marginal value of another similar example falls. Meanwhile, unfamiliar states can become disproportionately useful. A new object, failure mode, environment or recovery trajectory can expose a region of the physical world that the current dataset barely represents. So a smarter data engine should not treat every interaction equally. Robot experience → Estimate novelty → Identify uncertainty → Prioritize useful cases → Train → Update what is valuable next This turns data collection into an allocation problem. Physical robot time is expensive, simulation compute is finite, and human demonstrations require effort. The goal should therefore be to spend those resources where they reduce uncertainty the most. This also means the “best” dataset is constantly changing. Once the model learns a difficult scenario, collecting more examples of it becomes less valuable and attention should move toward the next weakness. In machine learning terms, this starts to resemble active learning. Instead of passively accepting whatever data arrives, the learning system increasingly asks: what experience would teach me the most right now? That may become an important property of scalable robot learning. The future advantage may not come from collecting every possible interaction. It may come from knowing which interaction is worth collecting next. @axisrobotics

We are waitting Fomc news but i still want to keep this long position. DYOR @DecibelTrade

Getting a robot from 0% to 90% success may be easier than getting it from 90% to 99%. The reason is that the first 90% often comes from learning the common structure of a task. The robot sees enough examples to understand typical objects, typical positions and typical environments. Most interactions begin to look familiar. The remaining failures are different. They live in the long tail. A transparent object. An unusual shape. A partially hidden handle. An object placed at an awkward angle. Unexpected friction. A human moving something halfway through the task. Two rare conditions happening at the same time. Each individual case may be uncommon, but the physical world contains an enormous number of them. Once robots leave controlled environments, rare situations stop being rare in aggregate. This creates an interesting robotics problem: Common cases → easy to collect → easy to learn Rare cases → hard to discover → expensive to collect → often responsible for real world failures That means simply collecting more random data eventually becomes inefficient. If a dataset already contains thousands of normal cup grasps, another normal cup grasp adds relatively little information. A previously unseen failure condition may add much more. The data strategy therefore has to change as the robot improves. Early training benefits from broad demonstrations of normal behavior. Later training increasingly needs targeted discovery of what the model has not mastered yet. Deploy → Find rare failure → Recreate it → Generate variations → Train → Deploy again Simulation is especially valuable in this stage because one rare physical failure can become an entire family of training scenarios. Change the object, pose, lighting, friction or disturbance around that failure and the model can explore the neighborhood of a weakness instead of waiting for the same event to happen again in the real world. So the last few percentage points are not simply “more of the same training.” They are a search problem. The closer robots get to real world reliability, the more progress depends on finding the situations they still do not understand. @axisrobotics

For the first time, scientists have mapped the complete brain and central nervous system of an adult male fruit fly — a key model organism in science. 🪰 Working alongside HHMI Janelia Research Campus and the scientific community, @GoogleResearch scientists and researchers used AI to combine millions of 2D images into 3D neural shapes, reconstructing a record-breaking 166,000+ neurons. This foundational map of the adult male fruit fly brain can help accelerate our understanding of the brain, and is a major milestone in neuroscience.