you have the same 24hr as Albert Einstein btw

everyone talking about Variational airdropping 32% on TGE not knowing @yeet will be biggest TGE event this year

life is good

Introducing Primus Confidential Finance (dapp.primuslabs.xyz), where privacy meets transfer and yield on-chain. We're excited to announce a collaboration between Primus @primus_labs, Unitas @UnitasLabs, and United Stables @UTechStables to launch a Confidential Finance Product on @BNBCHAIN, powered by the Fully Homomorphic Encryption (FHE) technique. This isn't just another confidential transfer or on-chain vault. It's a new standard for how private, compliant, and capital-efficient on-chain finance can work. 🧵

life be trying me lately man... my iPhone just got stolen

If you're in the US and have access to X Money, kindly asking you to send $8 to @jackbutcher. Put my ETH address in the memo: 0x9ed6db61d5d5f0d4137e9347012fc53a180cc2ce Send me a screenshot, and I'll pay you back the exact amount you send. Thank you!

Gm guys

If crypto can’t make me rich, AI will.

a very strong $1 candidate EOY

i’ve been [ ZDACTED ]. yours is still ████████. @zdacted zdacted.xyz/r/ZD-43X8HF7TZT8…

Bits by Bit then all at once

A robot can look flawless in a controlled demo. Change the lighting. Move the camera. Leave a mug where it wasn’t supposed to be. Suddenly it hesitates, misses, or falls apart. That’s the real gap in Physical AI. Not whether a policy can complete one tidy task… @axisrobotics is built around this problem. Robots that work across rooms, tasks, and embodiments need millions of trajectories, not a few polished lab takes. Hardware teleop can’t scale that. Closed rooms are too slow, too clean, and too expensive. So Axis treats data like an engine, not a one-time dataset: >task generation
>quality filtering
>trajectory smoothing
>physics + visual augmentation
>human-in-the-loop DAgger when the policy slips ref: s.kaito.ai/Ks54L4v Raw human demos go in messy. Training-ready coverage comes out the other side. The point isn’t a prettier demo. It’s a robot that doesn’t panic the first time the room looks different.

wtf is @axisrobotics ?? for the fact, most of us will never build a humanoid, rather most of us will live with them. reason someone has to teach those machines how humans actually move or think through kitchens, desks, workshops, and bathrooms. few of these are what that browser teleop and human gated correction are for. One person does a five minute task. Thousands of people do slightly different versions. Then the engine compounds. If language models ate the internet, physical models will eat human motion. Might as well be in the loop while that happens, instead of only watching the finished demo on your feed.​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​ XD be involved: s.kaito.ai/Ks54L4v

Why Axis is actually useful for AI work?? 1/4 Robot models problems aren’t GPUs, it’s actually how humans would handle data to them. @axisrobotics is basically a data factory with a training loop bolted on, not another “robot company” 2/4 interesting right? trainners open browser, drive a fake arm, and that run can become training data. No lab. No $40k rig. No GPU 3/4 The useful part isn’t the simulation itself, It’s the loop: people demo a task, train a policy and the policy fails in public, people jump in and fix it, that correction trains the next version. voila! 4/4 That last step matters. Extra perfect demos help less than “here’s the exact moment the policy screwed up.” They already shipped that as a product. “Pre-training”, then a quiet training stage, then post-training where you babysit the policy. >Early proof wasn’t theoretical. “Little Prince’s Rose” got 10k+ good trajectories in 3 days and actually watered a flower on a real Franka. curious to join? join with my ref link: s.kaito.ai/Ks54L4v

Stock as Tokens✅ Tokens as Stock ⌛️

Everyone has long moved from siloed and boring swap UI to quality aggregators that it faster across 100s chains. @jumperapp still my safest and still give the best fee routes. Will be JUMPING on this WL, hopefully i got in thank you @arjunnchand link to join: Join me on the Jumper waitlist. waitlist.jumper.xyz/r/JMP-F9…

1/20 @axisrobotics is a compounding data engine for physical AI. You teleoperate a simulated manipulator, produce trajectories, and those trajectories become the raw material that trains robot policies. If you want to actually contribute instead of just reading about foundation models, this is the walkthrough; 2/20 Physical AI is not bottlenecked by another transformer paper. It is bottlenecked by coverage. Axis treats diversity as a first class constraint. Scene variants, object layouts, camera viewpoints, embodiments, and language goals get randomized at the task generation layer so the dataset is not a pile of near duplicates. 3/20 The flywheel is simple once you see it. A task starts in pre-training. Humans teleoperate the end effector and produce demonstrations. Those demos train a base policy. The task then moves into post-training. The policy rolls out in the same simulator. You intervene only when it slips. Those corrections are human-gated DAgger data. The next policy iteration is trained on the mix. Better policies change which tasks get collected next. That loop is the product. 4/20 Setup is boring and that is the point. You need a desktop or laptop browser. Chrome, Firefox, Safari, or Edge. Phones are not supported because task space control assumes a keyboard and mouse. Go to Axis Hub, sign in with Privy (email, Google, X, or a wallet). If you do not connect a wallet, Privy mints one. That wallet is for signing trajectories on-chain, not for paying to play. Completing a run is free. Signing is a small Base gas fee. 5/20 The Hub is the task hall. Each card shows embodiment, scenario (kitchen, home, office, workshop, bathroom, play), skills (pick, place, stack, rotate, transfer, arrange, insert, separate, reach), difficulty, lifecycle stage, and remaining slots. Filter hard. Beginners should stay on pre-training, beginner or easy, single skill pick and place. Expert long horizon bimanual insert tasks will eat you alive on day one. 6/20 Every task has a lifecycle: Pre-training, Training, Post-training, then Ended. Pre-training is demonstration collection. You drive from scratch. One valid episode is one demo. Training is the aggregation window. Slots close. The pipeline trains a policy. You wait. Post-training is supervision. The policy attempts the rollout. You take over at failure states, correct, and hand control back. Do not treat those two collection modes as the same job. The data distributions are different and the scoring logic is different. 7/20 You do not jog individual joints. You command the end effector in task space. An inverse kinematics solver maps your Cartesian target into joint commands. Execution interpolates in joint space so the kinematic chain does not stutter or lock at limits. That distinction matters. You think in grasp frames and approach vectors. The robot thinks in joint trajectories. 8/20 Three input modes, mixed freely. Hold and drag the blue ring to slide the gripper along the current view plane. Double click an object to send the arm into a pre-grasp pose. Use this. Crossing empty workspace is low information data. The informative frames are approach, grasp selection, closure, lift, and place. Orbit, pan, and zoom the camera so you can actually see contact. 9/20 Keyboard, because you will live here. Translation is in robot base coordinates: Up/Down: end effector plus or minus X Left/Right: plus or minus Y E / D: plus or minus Z Rotation is in end effector local coordinates: Q / W: roll around EE X A / S: pitch EE Y Z / X: yaw around EE Z SPACE holds closed / release open on the gripper. C swaps the active arm on bimanual embodiments. Sensitivity slider under the viewport changes how aggressive the servo feels. Raise it if the arm feels like it is swimming through syrup. 10/20 Episode hygiene keys, memorize these before your first scored run. N saves a checkpoint of simulator state. B restores the last checkpoint. V completes and saves the current episode. R resets and discards the episode. ESC saves everything and exits. On multi-stage tasks (open drawer, pick object, place inside, close), checkpoint after every successful subgoal. One late collision should not force you to replay the whole horizon. 11/20 A good pre-training demonstration is not a speedrun with panic corrections. The scorer cares about success, efficiency, and motion smoothness. High jerk and high acceleration look like a human fighting the interface. The refinement pipeline can clean some of that, but raw teleop that already has low jerk trains better policies and scores better for you. Plan the grasp before you close the fingers. Approach along the object’s principal axis when you can. Avoid dragging the payload through other rigid bodies. Avoid table scraping. 12/20 Initial conditions are sampled per episode. The goal image is a reference end state, not a pixel perfect replica of your spawn. Read the step list. Left click objects if you are unsure which mesh is the target. The name appears on screen. If the success checker never fires, you probably placed the object in the wrong region, left a drawer ajar, or failed a contact condition. Reset, do not sit there micro jogging forever. 13/20 Post-training is a different sport. The policy drives. You watch the rollout. The moment the approach vector is wrong, the grasp is unstable, or it commits to the wrong object, hit Take Over. Then you are back in the same task space interface. Fix the local failure. Hand control back as soon as the state is recoverable. Long takeovers collapse the session back into a demonstration. That is already what pre-training is for. Correction data is valuable because it is on-policy and sparse. Intervene late enough that the failure is real, early enough that the scene is still salvageable. 14/20 After a valid run, the trajectory is uploaded: joint states, object poses, control actions, gripper state, timestamps, task metadata. It sits in your Portfolio. Quality score arrives after verification. To make the contribution count on-chain, sign it. That writes a provenance record (task id, contributor id, data id) so the dataset is auditable. Failed episodes are not uploaded and do not punish you. Repeat attempts on the same task are averaged, so farming one easy scene with sloppy motion is a bad strategy. 15/20 What happens after you close the browser. Raw teleop is cleaned. Mean acceleration and jerk get reduced in the refinement stage. Then the IsaacSim backend domain randomizes the episode into many visual and dynamics variants: lighting, textures, camera extrinsics, background, slight layout noise. Those augmented samples feed two training regimes. Task specific imitation learning (behavior cloning, diffusion policy). And generalist vision language action models conditioned on language goals. The public Franka sim set already showed the point. Continual pretraining of π0.5 on Axis data lifted LIBERO-Plus success from 83.9 to 88.8 and beat a volume matched RoboCasa control by a wide margin. Crowdsourced sim data is not automatically garbage if coverage is high enough. 16/20 Embodiments you will meet: Franka Research 3 style 7DoF arms, OpenArm, other single arm and bimanual stacks, occasionally more exotic end effectors. On dexterous hands the interface often collapses high dimensional finger joints into template interpolation (open rest pose toward a pinch, tripod, or power grasp). You are still commanding task space plus a grasp parameter, not puppeteering 16 joints by hand. Sim to real is the eventual test. Policies trained in this loop get evaluated under randomized initial conditions, then transferred onto real hardware for validation. Your browser episode is not a toy if the asset library and contact model are honest. 17/20 Quality checklist I wish someone had posted on day one. Watch depth. A top down camera lies about how close the fingertips are to the rim. Commit to a grasp frame and stop hunting. Lift before you translate so you do not plow the table. On place, hover, square the orientation, then descend. Do not fling. If the policy in post-training is oscillating, that is a sign, not a vibe. Take over at the start of the bad attractor, not after the object is on the floor. Do the newbie practice tasks first. They do not score. They teach the coupling between view plane drag and base frame keys. 18/20 Common failure modes. Driving in joint space in your head while the UI is Cartesian. You will fight the IK. Never moving the camera. Occlusion creates false contacts in your brain. Closing the gripper in free space then hoping the object teleports in. Ignoring articulated joints. A cabinet door that is 10 degrees short of the success region fails the checker. Taking over for ten seconds in post-training because it feels safer. You just spent the intervention budget turning a correction into another demo. 19/20 If you care about the research layer, the terms to keep straight are: Trajectory: the time series of state and action. Policy: the mapping from observation to action. Rollout: executing that policy in the environment. Success checker: the programmatic predicate that labels the episode. Domain randomization: expanding one demo into many visual and dynamics cousins. Human-gated DAgger: the expert takes over only at policy failure states. VLA: a policy that conditions actions on vision and language. Axis is trying to industrialize those pieces with a global contributor pool instead of a 12 person teleop team in one lab. start here: s.kaito.ai/Ks54L4v

They’ll show you the fine side of the National Theatre and hide the other side from you 😭💔 Well, here’s the side they don’t want you to see 👇🏽

One App that Controls it all? With many exchanges, you have assets scattered across Apps but with @Backpack, you got an exchange that unifies your funds as ecosystem. >Raging from Crypto, stocks and perps, yields, lending/staking, and use a self custodial multi-chain wallet. All this in one portfolio. >With Stock tokenization, eligible Trad assets moves onchain and does more than just idling for you. Now it all about real ownership and unified capital, built as one. I made you a video that says it all