Author: Matt Sheehan
The “adults in the room” of AI research are quietly pivoting away from pure LLMs. Here is the new stack—World Models, Causal AI, and Decision Engines—that will define the next phase of intelligence.
If you are reading this, you are likely suffering from the same affliction as the rest of us: AI fatigue. The constant noise, the doom-mongering about job losses, and the relentless over-promising of current tools leave many of us scratching our heads.
We have moved rapidly from task-based machine learning to the current explosion of Large Language Models (LLMs) and generative AI. We are in the midst of a perfect storm of hype and capital. Money will continue to pour into LLMs, improving chatbots and copilots.
But while the world focuses on text and image generation, the next, more consequential wave of AI is already forming.
I call this the Physical AI Revolution. Many signs point to 2026 as the year this momentum becomes undeniable, shifting our focus from AI that can write about the world to AI that can truly understand and act within it.
The Signal: The Founders Are Moving On
Why am I so confident that the “LLM-only” era is ceilinging out? Because the people who built the current era are telling us so by their actions.
This week provided a synchronized signal from the industry’s heaviest hitters:
- Yann LeCun, Meta’s Chief AI Scientist, was typically blunt in a recent internal memo. He argued that AI systems lacking grounded world models are destined to fail because they don’t understand cause and effect. Without a reliable internal model of reality, even the most eloquent LLM is essentially hallucinating plans that won’t survive contact with physics or economics.
- David Silver, the Google DeepMind pioneer behind AlphaGo, has left the company to found a new startup. His explicit bet is that LLMs alone won’t reach superintelligence. His new venture is focusing on systems that learn from experience and agency—the core tenets of Physical AI.
When the architects of modern AI start heading for the exit ramp off the generative hype highway, it’s time to pay attention.
Deconstructing Physical AI: The New Stack
LLMs are incredible pattern matchers. They are like librarians who have read every book ever written but have never stepped outside the library. They are the perfect interface layer for humans to access complex information.
But LLMs have a fundamental limitation: they can describe the world, but they cannot act within it.
To move from chatbots to true “decision engines”—systems that can anticipate, adapt, and predict—we need a new stack. Physical AI is composed of three core components designed to act as a “second brain” in the room.
1. World Models (The Simulation)
Traditional digital twins give humans a 3D picture to look at. World models give AI a 3D world to live in.
These are environments with physics, walls, and gravity where an AI agent can be dropped in to walk around, collide with obstacles, and learn safe routes. It’s the difference between watching a movie of a fire and running a simulation to see how that fire spreads under different wind conditions.
This isn’t far-off theory. Just this week, Robbyant open-sourced LingBot-World, a real-time, action-conditioned world model. Unlike passive video generation tools from OpenAI or Google, this model allows an agent to interact with the environment in real-time. It is a tangible step toward the “decision rehearsal” required for real-world AI.
2. Causal AI (The “Why”)
Once an AI has a world to navigate, it needs to understand consequences.
Traditional analytics tell you about correlations. Causal AI allows for counterfactual questions. It distinguishes signal from noise to answer the ultimate question: Which action truly caused the result?
While a world model helps an AI ask, “If I do X, what happens next?”, Causal AI helps it ask, “What if I had done Y instead?”
3. Agentic Decision Engines (The Closer)
When combined, these create causal world models that can simulate and explain. But we still need an actor.
Agentic Decision Engines are the “closers.” They consume the simulations from the world model and the reasoning from the causal AI to execute the optimal action. They don’t just recommend; they are designed to interact with control systems to close the loop.
Physical AI in Action: The Firefighting Example
To understand how these layers work together, forget about writing poems. Consider a high-stakes environment like fighting a fast-moving wildfire.
- The LLM Interface: A commander asks, “What’s our status?” The LLM summarizes incoming text reports.
- Step 1 – Live Data (The Input): Sensors, drones, and weather feeds update the reality.
- Step 2 – World Model (The Rehearsal): The system simulates how the fire might spread over the next 60 minutes under three different wind scenarios based on physical laws.
- Step 3 – Causal Analysis (The Reasoning): Causal models identify that fuel type and slope slope—not just wind speed—are the actual drivers of risk in a specific sector.
- Step 4 – Decision Engine (The Action): An agentic engine evaluates tactical options and recommends redeploying Crew A to a specific ridge line based on the simulated outcomes.
In this setting, World Models don’t replace real-time data—they amplify it by enabling decision rehearsal before committing lives and assets.
The Year of Decisions
We have spent the last three years teaching AI to write, code, and generate images. 2026 will be the year we finally teach it to work.
The momentum behind Physical AI is building not because of VC marketing hype, but because the current tools have hit a ceiling for operational use. The organizations that win in the next phase won’t just have the best chatbots; they will have the best ability to simulate the future and act on it.