Jasonyeah | DeSpread

Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do. On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.

Been seeing a lot of questions about what Virtuals is doing in robotics and what it means for $VIRTUAL (1) Building an agent tokenization infra. and ACP taught us something pretty simple: intelligence gets more valuable when you own the data, coordination, and economic rails around it ($$$ is a plus). (2) Software AI got unbelievably lucky because the internet was basically sitting there as training data, while robotics has no equivalent and physical AI is only getting bigger. (3) You can't scrape the internet for someone folding a towel with the exact joint movements, force, camera feed, and mistakes a robot needs to actually learn from. (4) A lot of useful robotics data still has to be created one physical task, one operator, and one failure at a time. (5) That made us realize the near-term cash cow in robotics isn't necessarily the robot itself, it's the data required to make millions of robots actually useful. (6) Hence we built @eastworlds_io, became an official data and deployment partner of Unitree, and now operate Southeast Asia's largest Unitree fleet with 35 G1s and H2s, but what matters more is what we're actually doing with the robots. (7) Our model is simple: deploy robots in the wild, collect what happens, and use that data to train autonomous policies for real commercial jobs. (8) We currently have robots deployed from hotel cleaning to agriculture, with a much larger pipeline across hospitality, power, and energy. (9) Every time a robot fails and a human takes over, that correction becomes exactly the edge-case data that is almost impossible to manufacture inside a lab. (10) And the math is mathing: G1 teleoperation data sells for around $100 to $300 per hour, while UMI/YAM data sells for around $30 to $80 per hour. (11) We're generating ~200 hours of teleoperation data a week, selling it to labs building on the G1, and training our own single-task policies on it for deployments we already run. (12) We like the flywheel: Eastworlds can generate revenue from data today while building the dataset and autonomy stack that makes the entire network more valuable tomorrow. (13) Our thesis is to build this economy vertically first, then progressively bring the capital, incentives, ownership, and payments around that economy back into $VIRTUAL. (14) And it's important, since there is only $VIRTUAL For a deeper dive into our progress, @Shoalresearch put together a great read below on Eastworlds, the robotics industry, and much of what we've been building toward.

The future of FDE work seems closely related with all work around evals/posttraining/RL envs. FDEs are effectively responsible for the following: 1. Define the business problem. 2. Codify the business problem into an eval rubric and environment. 3. Hillclimb the environment and output an agent/agentic workflow that solves the business problem. Right now the process of (3) is quite manual - historically FDEs spend hundreds of hours creating bespoke software/workflows that solve the problem. But assuming intelligence is abundant, they can effectively offload (3) to some automated optimization process. This includes RL on the model layer, and using Claude Code/Codex to optimize the harness/workflow. Then the FDE responsibility shifts from implementing the task to defining the right goals and outcomes. In other words, they have access to /goal, and their job is more around making sure the goal, environment, and evals are correct vs. the tactical implementation details.

Launched 3ridge CLI! Agent-native tool that gives AI agents direct access to 3ridge's data. Korean mindshare, KOL leaderboard, market oracle(Kimchi premiums, K-stocks volumes, etc) All queryable from your terminal or your AI agent.