During the Korean War, American F-86 Sabre pilots consistently beat Soviet-built MiG-15s in air combat, even though the MiG-15 climbed faster and flew higher -- on paper, the better aircraft. Air Force colonel and strategist John Boyd studied the mismatch and built an explanation around it that became one of the most cited frameworks in modern strategy: the F-86's cockpit gave pilots a wider field of view, and F-86 pilots were trained to move through a decision cycle -- observe, orient, decide, act, the OODA loop -- faster than their opponents could. Whoever completes that cycle first doesn't just react faster. They get inside the other side's own decision cycle, acting before the opponent has even finished orienting to what just happened.[1] The commonly cited 10:1 kill-ratio figure that made this case famous has since been challenged by aviation historians as too high -- the real ratio was almost certainly lower, though still a real American advantage.[2] The mechanism Boyd named held up better than the specific number that popularized it.
Working well with an AI system depends on a step that looks a lot like Boyd's Orient phase: interpreting what the system produced before deciding what to do with it. In ordinary AI collaboration -- drafting, brainstorming, research -- that interpretation step can take as long as it needs to. Nothing bad happens if a person takes an hour, a day, or a week before reacting to a draft. Discernment is only possible in the first place because deferral is always available: a human can insert judgment before anything happens, because nothing happens until the human responds.
A real-time physical system has no equivalent slack, and that is a structural break, not just a harder version of the same problem. A shop-floor robot, an autonomous vehicle reacting to a pedestrian, an industrial safety interlock -- each has a decision window measured in milliseconds, faster than human deliberation can operate in at all. The device has to complete its own OODA loop and act, because there is no time budget left for a human-mediated Orient phase. This isn't a matter of the system being less trustworthy in physical form. It's that the one condition discernment has depended on this whole time -- the ability to wait -- runs out first exactly where the stakes get physical and immediate.
Mass-casualty autonomous drones need no new technology to become viable -- only integration of components that already exist, because indiscriminate targeting is a far lower bar to clear than precision targeting. Precision targeting still needs something like discernment: distinguishing a real target from a decoy, a combatant from a civilian. Indiscriminate targeting removes that step from the problem entirely rather than simply speeding past it. That is a sharper, more legible version of the shop-floor example -- not because the technology is more advanced, but because the judgment step isn't merely fast in this case. It has been engineered out.
Three separate 2026 product efforts show the same shift in direction, with genuinely mixed results. Google DeepMind's Project Astra is a multimodal agent built specifically to perceive, remember, and interact with the physical world in real time -- processing live video and audio together and taking action inside real tools (search, email, calendar, maps), not just describing what it sees.[3] NVIDIA's Cosmos platform, announced in 2025 and now on its third major version, is a "world foundation model" purpose-built to simulate physical environments and generate real robot action sequences -- a training ground where a robot policy can be tested and refined against a physics-grounded simulation before it ever moves in the real world.[4]
OpenAI's Atlas browser, launched in October 2025 to let an AI agent act directly inside a person's own browser, did not survive its first year: OpenAI shut it down in mid-2026, folding its functions into a different product, after it ran into exactly the kind of new failure surface that pure content generation never had to deal with -- prompt-injection attacks, security flaws, and an agent too slow and unreliable to trust with real actions.[5] Atlas is the concrete, dated case for this piece's own caution: giving a system the ability to act in the world doesn't just raise the stakes of an existing risk, it opens entirely new ones -- an AI system taking real browser actions can be manipulated by hostile content it encounters while acting, a risk with no real equivalent in a system that only ever produces a draft for a human to review first.
An experienced human operator's own judgment was never installed complete in one transfer -- it was built the same slow, incremental way any real expertise is built: reached for, missed, corrected, repeated, over years. That doesn't soften the real problem with real-time physical AI -- a deployed system still can't defer once it's live, no matter how it was trained -- but it does mean the training question and the deployment question are two separate problems. Whatever judgment a real-time system runs on had to be built slowly, the same way any judgment is built. It just can't be exercised slowly once it's actually running.
Why does this matter? The interesting question about AI and judgment was never "will AI eventually be smart enough to be trusted." It's narrower and more mechanical than that: where does the human's ability to wait before deciding actually run out? Boyd's F-86 pilots won by moving through their own decision loop faster than a slower-deciding opponent -- a human advantage. Physical AI is the place that same logic flips, and the system, not the person, ends up owning the loop.