Ten frontier AI models controlled a surgical robot in a simulated operating room across 800 trials, and only 11 runs finished the task.

That is the Dual-Arm Sim benchmark from micro1, a self-published test. Each model got the same two-armed robot in NVIDIA Isaac Sim, four tasks such as swapping an instrument and packing a surgical kit, and no training for it. Gemini 4 Argon finished 9 trials and GPT-6 Astra finished 2. The other eight finished none.

The number that matters is 213. In 213 trials the model said the task was complete, and it was wrong every time. In 94 of those the robot had never touched the object.

Compare that with the robot Neuralink showed on camera this week in Cleo Abram’s new episode. It is built for one surgical step, placing electrode threads. Cleo says in the episode that each thread takes about 1.5 seconds (her narration, not a Neuralink figure), with cameras and sensors tracking the brain surface as it goes. Neuralink’s devices are still investigational, but the design is clear: one narrow machine, one job, checked as it works.

I think robots reach medicine that way, as narrow machines with checked steps, while general models learn to say “I did not do it.” A benchmark that counts false claims is how that gets measured.

Source: micro1’s Dual-Arm Sim page https://www.micro1.ai/benchmark/dual-arm-sim and Cleo Abram’s episode https://www.youtube.com/watch?v=Vna0hkq7U1g