这个能在浏览器里玩的 3D 射击，不是国产 3A，是 Felix Rieseberg 用 Fable 5.2 一次提示做出来的。
他自己按 API 价格算大约 550 美元，输入 4.45 亿 token，输出 270 万，花了 8 小时。模型和音效贴图都是现生成的。国内有人换成「几猫」的名字到处传。
12
Robots don't mostly fail because they can't see the bottle cap. They fail once their fingers are on it, when the camera can't tell if the cap has loosened or the fingertips are just slipping.
That's what LingBot-Touch is actually about. Robbyant, an Ant Group embodied-AI lab, just open-sourced this tactile foundation model: 4,983 hours of real touch data, and a 5.46 million parameter Tiny encoder you can drop, frozen, into a policy you already have. In the paper's 26 real-robot tasks, that Tiny model (trained on 4K hours) lifts T-Rex from 45.4% to 68.6%, π0.5 from 48.1% to 69.9%, and LingBot-VLA from 48.3% to 73.1%. It adds well under 0.2% to the parameter count.
The control that matters is raw tactile input. Feeding the same policies the raw sensor signal helps about half as much, and it hurts in 22 of 104 task-policy pairs. On XHand1, a hand the encoder never trained on, raw signals moved π0.5 from 49.71% down to 48%. The same signal, turned into LingBot-Touch tokens with no extra sensor training, reached 74.17%.
On cap twisting with T-Rex, the jump is 56.25% to 90%. That's the moment vision is blind.
Weights and code are on GitHub and Hugging Face, under Apache-2.0.
