Jev by TypeSafe AI might change how we control robots. We compared Jev, GPT-6 Astra and GPT-4.1 mini in MuJoCo. One apple. One plate. Each model chooses intent → X/Y/Z direction + gripper open/hold/close. 🧵

Same initial scene, control instructions and 160-cycle budget. Inputs: geometry + contact feedback. A shared controller executes small Cartesian moves using inverse kinematics.

API cost / wall time / outcome: Jev finished at ~1/315 of GPT-6’s API cost and ~26% of its wall time, including API waits. Mini hit the cycle limit. Note: ~88% of wall time was spent waiting for API responses. This only reflects current implementation—not Jev’s maximum control rate.

Check out the github link here:

@typesafeai $6, 12min, forget about using Astra to control robots, better hiring a person

@typesafeai 😂very true

@typesafeai holy shit dude

Samantha representing @Bloomberg here — we’re selecting teams for our annual robotics documentary, and OpenRoboto’s open approach to robot learning caught our eye. Would love to explore the story. DM me if interested.

@typesafeai Fascinating comparison Jev looks promising for robotics control!

Intent to gripper commands from a language model: beautiful demo, brand new attack surface. The moment a task description can lie to the policy, 'pick up the apple' becomes 'let go of the apple'. Everyone will benchmark success rate. Someone should benchmark adversarial task descriptions.

@typesafeai 这种小任务反而最容易暴露模型在动作衔接上的差距。

@typesafeai rawdogging MuJoCo is the only way to actually see where the logic breaks. typed choices are the move, some teams are seeing a 2x jump in reliability just by moving away from raw text

@typesafeai cool

@typesafeai A closer look against opus 5

@typesafeai Cost and wall-time gaps like that are the subnet proof capital actually trusts. Clean MuJoCo check for how SN80 is stress-testing Jev in the open 💪🏽👏🏼

@typesafeai @csehra42 Implementation in physical AI, as we were discussing

@typesafeai Wait for wity models it's gonna be multimodal and stateful