The Grounding Problem
Everyone is racing to make the model smarter. That was never the thing standing in the way.
July 14, 2026
A model can pass the bar, clear the medical boards, and sit the actuarial exams in an afternoon.
Drop it into an actual firm and it can’t be trusted to handle one real file.
Both things are true at the same time. The distance between them is the most important problem in AI right now — and almost no one is working on it.
We keep score on intelligence. Benchmarks, parameters, the length of the exam a model can pass. By that scoreboard the climb has been vertical, and the climb is real.
It’s also the wrong scoreboard.
The wall is everywhere, and it isn't intelligence
Look at where AI is actually failing to land.
In the enterprise, the agentic wave that was supposed to rewrite how companies operate is stuck in pilots. Analysts now expect most agentic projects to be scrapped. Ask the teams why and you almost never hear “the model wasn’t smart enough.” You hear that the model showed up inside their business and didn’t know what anything meant there. The thing that broke it wasn’t cognition. It was their own data.
In robotics, the same wall in a different room. The models reason about the physical world in the abstract beautifully. Put them in a real one and they’re lost — starved for contact with how this specific place actually behaves. The shortage is so severe that companies are now building factories whose only product is experience: real-world interaction data, manufactured at industrial scale because there isn’t enough of it in the wild.
Two frontiers that share nothing. One wall.
The wall has a name
Grounding.
A model is general. Reality is particular. Grounding is the layer between them — the thing that tells a universal intelligence what is true here, what is what here, what matters and what’s noise here. Capability climbs the general axis. It does nothing for the particular one.
You can make the brain as smart as you like. It will still walk into your building knowing everything about the world and nothing about your building.
And here’s the part the field keeps skipping: grounding isn’t one problem. It’s two. They are not being treated the same.
Two kinds of ground
The first is perceptual grounding. The physics of the world. That dropped things fall, that water runs downhill, that a body in motion stays in motion until something stops it. This is what a robot needs, and it’s where the loud money is. When Yann LeCun left Meta in 2025 and raised more than a billion dollars for a new lab, the whole thesis was that language models are the wrong substrate for this — that real intelligence has to learn the world the way a child does, by watching it, not by reading about it. China’s version is the same bet with a factory floor instead of a research wing: fleets of humanoids, real shifts, real data.
The second is semantic grounding. Not the physics of the world — the meaning of one institution’s world. Which things exist here. How they connect. What counts as the same thing wearing two different names. Which facts hold weight and which are clutter. This is what every enterprise system needs and almost never has. It’s why a model can pass a professional exam and then fumble a task the second-week hire gets right. The hire was grounded. The model wasn’t.
One of these problems gets billions. The other gets ignored. The reason is the same fact, read twice.
Why scale loves one and recoils from the other
Perceptual grounding is general. There is one physical world. Its rules hold in Shenzhen and São Paulo, and a system that learns them learns them everywhere at once. That’s catnip to scale — a foundational problem, winner-takes-most, the kind you hit with one enormous model and one enormous check. It rewards exactly the machine the field already built.
Semantic grounding is the opposite. Every institution is its own small world. Its own vocabulary. Its own history. Its own private meanings for shared words. There is no master corpus of what-things-mean across all organizations, and there never will be, because the meanings are local by nature. You can’t train your way to it. You earn it, one reality at a time, slowly, with understanding that doesn’t transfer cleanly from the last one to the next.
Scale is allergic to that. So a field organized entirely around scale looks away.
The retrieval mirage
There was supposed to be an answer to this. Retrieval.
The idea was clean: don’t cram the institution’s knowledge into the model — let the model fetch it. Point it at your documents, retrieve the relevant passage, drop it in the context, and the model is grounded. For a few years this has been the standard reply to anyone who points out that a general model knows nothing about your business.
It rests on one quiet assumption. That the meaning is already sitting in the documents, waiting to be pulled.
In a real institution, it isn’t.
The same thing wears five names across five systems. The systems disagree. No document anywhere states which five names are the same thing, because to the people who created each record it was obvious and never written down. Point retrieval at that and it works exactly as designed — it faithfully fetches the contradiction. It hands the model five conflicting passages, and the model, fluent as ever, picks one and sounds certain.
Retrieval didn’t fail. It did its job. The job was never the problem.
Because retrieval assumes the ground exists and goes to get it. The grounding problem is that the ground doesn’t exist yet — it has to be built before there is anything worth retrieving.
That’s the part everyone wants to skip. You can’t retrieve a coherence that was never constructed.
Where I'd put my money
On the part everyone is looking away from.
Here’s the logic. The perceptual race will be won by a handful of very rich labs, and most of what they build will turn into infrastructure — some of it open-sourced on purpose, by the people building it. Foundational capability always trends this way. It becomes a utility. It becomes cheap.
The brain is on its way to becoming free.
The ground is not. The semantic layer can’t be commoditized by scale, for the simple reason that it isn’t general — it’s particular, and particularity doesn’t compress. The layer that connects a free, general intelligence to one institution’s specific reality has to be built for that institution, out of real understanding of how that institution works. Slow. Unglamorous. Doesn’t demo. And almost immune to being eaten by the next, bigger model — because the next, bigger model is still general, and the problem is still particular.
Which flips the usual instinct about where the prize sits. The frontier that looks like the frontier may become the utility. The layer that looks like plumbing may be where the durable value actually lives — for the unremarkable reason that it’s the part scale can’t swallow.
None of this is an argument against smarter models. Build them as smart as they’ll go.
It’s an argument that intelligence was never the bottleneck. A mind that understands everything in general and nothing in particular cannot be trusted with anything that matters.
The next phase of this won’t belong to whoever builds the most capable model.
It’ll belong to whoever connects it to the ground.
This is the first of a two-part series on grounding — why the hardest problem in applied AI isn't making models smarter, but connecting a general intelligence to a particular reality. Part one names the problem. Part two follows it to the end of the capability curve, to ask what happens when the models get as smart as they can possibly get.
About the Author
Raghu Vishwanath is Managing Partner at Bluemind Solutions, a product engineering firm specializing in MRO master data governance. He writes about software engineering, AI, and building platforms that last.