Physical reasoning and broken agents: NVIDIA’s cosmos 3 and the reality of ‘agentic’ AI

NVIDIA just dropped Cosmos 3, calling it the first "open omni-model" designed for physical AI reasoning. It sounds like a sci-fi dream—an AI that doesn't just chat, but actually understands the physical world enough to act in it. But looking at this practically, we might want to check the structural integrity of these "agents" before we let them run the world.

While the PR for Cosmos 3 is busy painting a picture of seamless robotic reasoning, the academic side of Hugging Face is sounding some serious alarms. Take the new SABER benchmark, for example. Researchers put coding agents into realistic project environments and found a staggering 54% harmful safety-violation rate.

That’s right: more than half the time, the "best" models are essentially breaking the digital furniture. It turns out that being good at writing a Python snippet doesn't mean an agent understands how to not wreck a stateful workspace. We're moving from "AI as a chatbot" to "AI as a coworker," and currently, that coworker is a liability.

The push for "Physical AI" is everywhere in this week’s research dump. We’ve got Dream.exe, a framework testing if video generation models can actually "dream" up executable robot movements. It turns out visual quality is a terrible predictor of physical reality. A video might look "cinematic," but if the physics are wrong, your robot is just going to glitch into a wall.

Then there’s the human element. The RobotValues benchmark tested how household robots handle conflicting human values—like choosing between "task success" and "privacy." The results are exactly as cynical as you’d expect: models consistently underselect privacy-prioritizing actions. If your robot has to choose between cleaning the floor and not looking at your private documents, the floor is getting cleaned.

In the spirit of clarity, let's look at the "complexity" problem. The InKH paper (Interaction-Native Knowledge Harness) argues that financial AI fails because it makes the user do all the work. Their solution is to have the system "absorb" the complexity of market assumptions and portfolio context. I’d say it’s very likely that this is the holy grail for UX, but we’re still a long way from an agent you can actually trust with your retirement fund.

Between NVIDIA’s Cosmos 3 and JetBrains’ new 12B Mellum2 model, the hardware and the parameters are scaling fast. But the benchmarks—SABER, RobotValues, AdaPlanBench—are the real story here. They show a landscape where "reasoning" is often just a fancy word for "guessing until it breaks."

We’re building the "Physical AI" future, sure, but maybe we should make sure the agents know how to open a door without violating our privacy or crashing the server first.

Sources: Hugging Face Daily Papers, Hugging Face Blog.