The Jensen Loop

Everyone treats Boyd's OODA loop as timeless. It isn't. It was built for a world where information was scarce, and that world ended sometime in the last three years.

Observe, orient, decide, act, faster than the man trying to kill you. Cycle fastest, win the dogfight. The whole loop assumes two things. An adversary, and an information edge you can win by moving quicker.

OODA stops working when everyone is fast.

Everyone sees the same papers, the same demos, the same launches, the same week. Speed equalizes. The information edge was the assumption AI erased. When the inputs are identical and instant, cycling faster buys you nothing. The person across the table got the same input at the same time and is cycling just as fast. The OODA advantage was never speed. It was speed relative to someone with worse information. Take away the gap and the loop spins in place.

What doesn't equalize is how deep you reason before you commit.

Jensen Huang answered a question at a Stanford talk about forecasting the future and you see the loop that replaces it. No adversary in it. He isn't trying to cycle faster than a competitor. He's navigating fog where the only enemy is the time, energy, and money he'll spend going the wrong way. The competitor was never the thing to outrun. The fog was.

I'm calling it the Jensen loop. He never named it. He ran it out loud at Stanford when a student asked how he forecasts when the shape of the future isn't clear. Four moves. Descend, identify, verify, execute. DIVE.

Descend means go one click deeper than the room. AlexNet crushes every computer vision result decades of researchers built, in one shot. The room observes that and moves on. Huang descends. Is this a big deal. If you could do it this way, what else could you solve. What does it mean for computers themselves. OODA's second step is orient, fit the new thing into your existing frame, fast. Huang breaks the frame and rebuilds it from first principles. Orientation makes you faster inside your assumptions. Descending lets you notice your assumptions are dead.

Identify is what the depth reveals. Out of the descent comes a model of the future, derived, not borrowed.

Verify it against first principles, not the frame you walked in with. This is the step that catches the error before it costs you.

Execute before the surface catches up. Here Huang does the move almost nobody copies. He works backwards from the future he just derived, not forward from today. Hopper was built backwards from a pre-training market that didn't exist yet, for a customer count of precisely zero, on first reasoning alone. Then he commits. OODA decides an action and executes it. Huang sorts the future into bands. These things will likely happen. These will absolutely happen. These may happen. Then he walks in a direction and feels his way through. He's reducing opportunity cost and buying optionality. The real cost was never being wrong. The real cost is the energy you spent not being somewhere else.

The cleanest way to see the loop is to watch it fail.

Nvidia got into mobile. Important phone companies came knocking, and Huang said yes. The mistake wasn't picking wrong. He stopped one click too early. A couple more clicks of descent and he'd have asked what the most important part of a phone actually is, and answered the modem, and seen that Qualcomm owned it. He didn't descend. He oriented. He fit mobile into his existing frame, we add value with graphics, and built forward from there. The business grew to a billion dollars. Then the 3G-to-4G transition came, Qualcomm owned the modem, the modem owned the phone, and Nvidia got locked out. The billion went to zero.

The fog was readable. One more click would have read it. OODA would have told him to move fast into mobile. He did. That was the mistake.

The recovery is where the loop finally fired. He took the low-power expertise the mobile years had built and worked it backwards into something that didn't exist yet. Robotics. Thor, he says, is the great-great-great-grandson of that mobile chip. Descend on the asset, it was never a phone chip, it was extreme energy efficiency. Identify a future where that matters. Verify it against first principles. Execute backwards from it. Same loop that built Hopper. He's honest that the recovery is rationalization. Going into mobile was a waste of time. The loop redeems the asset. It does not redeem the skip.

The same trap is everywhere in AI right now. Everyone is building agents and treating the harness as plumbing. That's the room. One click down, the model is the easy part. The harness and the inference economics are the hard part, and that's where the value sits. The operators who win run their own agents in production, not just sell the framework. You build backwards from what an agent-native company looks like in three years, not forward from what's a useful tool this week.

The fuel underneath all of it is Huang's ugliest, truest line. A student told him compute was impossible to get. His answer landed like an insult before it landed as a gift. It's absolutely your fault. Because the moment it's your fault, it's yours to fix. The loop only spins if you believe the problem belongs to you. OODA assumes agency. The Jensen loop manufactures it.

Boyd's pilot orients to shoot first. Huang descends to see what the room can't, then commits to a direction with no map and no enemy. One is a gunfight. The other is navigation.

Everyone surfs the same news. Winners dive.