The real world cannot be described in words

Yann LeCun has spent years making an uncomfortable argument: today's large language models are powerful, but they are not enough to solve problems in the physical world. In recent interviews and around the launch of his new company, AMI, he has again pushed the view that real intelligence requires systems that can reason, plan, and build models of the world, not just generate digital content.

"The entire industry has been LLM-pilled," he said. "Current systems absolutely cannot plan. Trained solely on digital data, they have no way of understanding difficulties in the real world."

At almost the same moment, Jeff Bezos was reported to be promoting a similar line of thinking: a $100 billion fund aimed at acquiring manufacturing businesses and modernizing them with AI, especially in sectors like chipmaking, aerospace, and defense. Project Prometheus - his AI startup at the center of it - is not building chatbots. It is building what insiders describe as physical AI: systems that simulate how real-world environments behave under different conditions, from material stress to factory layouts to supply chain constraints.

The important point is not whether the fund closes exactly as reported. It is where the attention is going: toward the physical economy.

Those two signals point in the same direction.

The next frontier for AI is not a more conversational interface layered on top of digital workflows. It is better decision-making in the real world: factories, suppliers, lead times, constraints, bottlenecks, and uncertainty.

Now look at what still runs most supply chains.

Not intelligence built for uncertainty. Planning software built for a different era.

For years, enterprise planning systems were on the road to digitize processes, centralize data, and make operations more legible. But they were largely designed for a world that was slower, more stable, and more predictable than the one manufacturers operate in now.
The results? Sunk cost, value destruction.

In the physical economy, the environment itself is uncertainty. It is the environment itself. Demand shifts faster. Supply shocks last longer. Component lead times stretch across quarters. Geopolitics, capacity constraints, and new-product ramps interact in ways that do not fit a deterministic human based plan, and no dashboard can fix that underlying issue.

And yet the standard response in software has been familiar: more integration, more data, more workflow, more automation around the same planning logic. In some cases now, a chatbot is glued on top.

But a faster interface on top of brittle assumptions is still brittle.

That is the dead end - not just in AI, but in planning. LeCun sees it in LLMs. It has been hiding in supply chain software for thirty years.

The problem is not that companies lack data. It is that they are still turning volatile systems into false precision. Forecasts are treated as answers instead of opportunities. Lead times are entered as facts instead of risks. Inventory policies are calibrated to averages that stopped holding a long time ago.

The result is predictable: excess inventory in the wrong places, shortages in the right ones, constant replanning, and organizations that confuse motion with control.

The next generation of software for the physical economy has to start somewhere else. It has to begin with the premise that uncertainty is real, persistent, and operationally material. It has to model that uncertainty directly, test decisions against it, and improve outcomes in spite of it.

That is the shift Hexight is built around.

This is precisely the layer that every manufacturer Bezos acquires will need. Project Prometheus will simulate the factory. It can model engineering. But the supply chain feeding that factory - the lead times, the inventory policies, the procurement decisions made under genuine uncertainty - that requires technology that was built for the physical world as it actually behaves. Not the stable world planning software assumed. Not the digital world LLMs were trained on. The uncertain, unforgiving world that exists between a supplier in one country and a factory floor in another.

This is why the real opportunity in AI is not only model capability. It is in contact with reality.

LeCun's critique matters because it challenges the idea that scale alone will get us to robust planning and reasoning. Bezos' reported manufacturing thesis matters because it suggests where value will accrue when AI moves from the screen into industrial systems.

The incumbents are still digitizing a worldview shaped by stability.

The world that matters now is not stable.

The companies that win in the next decade will be the ones that stop pretending otherwise.

‍