When several experts in the industry argue that large language models (LLMs) may be a dead end for anyone expecting truly super-intelligent systems, attention is shifting to physical AI.
If that’s where the next wave of impact lies, the real bet is not just on models, but on the software infrastructure that turns physics and data into code that can run on hardware.
By that measure, MATLAB, the technical computing platform from MathWorks, is becoming hard to ignore.
It was released in the mid-1980s as a tool to make matrix math accessible to engineers and researchers.
Over time, it grew into a full environment for modeling, simulation, and embedded deployment, long before “AI” became the headline feature.
Today, it sits at the centre of workflows in automotive, aerospace, industrial automation, energy systems, and medical devices.
In practical terms, MATLAB is where engineering teams turn equations into working systems.
They model how an engine behaves under load, how a controller responds to sensor noise, how a motor heats up, or how a pump reacts to pressure changes.
They simulate those systems, tune control logic, test failure cases, and then generate C, C++, or CUDA code that runs on microcontrollers, GPUs, and real-time hardware.
For many products, MATLAB sits between design and deployment, carrying models from a laptop into production controllers. In physical AI, this simulation-first approach is what reduces risk before a system ever touches hardware.
Generative AI as a Layer Over Deterministic Engineering
MathWorks’ recent push into generative AI fits into this longer arc. The company introduced MATLAB Copilot as a way to help engineers navigate and assemble complex workflows that already exist inside the platform, rather than replace them.
“It helps users better understand how to connect together all those different capabilities and workflows that we’ve been building over the years so that they can apply them to their own designs,” said Seth DeLand, who works on generative AI products at Mathworks, in an interaction with AIM.
Copilot runs on grounded output because engineers can’t afford hallucinated commands or fabricated APIs.
“We actually look at our documentation, and then we feed all of that information to the large language model, and it can generate a response that is both based on its own knowledge, as well as the MathWorks documentation.”
Copilot becomes more useful because MATLAB has been integrating AI long before the generative wave. It sits on top of thirty years of AI components that already run in production hardware.
DeLand points to control systems as a natural fit for Copilot. “A classic example is cruise control in a car,” he said. The driver sets a speed, but the system has to hold it steady through hills and wind, using control algorithms designed by engineers. Copilot can help explore those choices.
“You might ask MATLAB Copilot what some different control algorithms I could consider,” DeLand said, and even have it draft an implementation: “You can ask it to actually generate code for you that implements the controller.”
He also sees value in using AI to make sense of existing work. “These AI models are actually really great at digging into all the code and generating a nice human language explanation of what’s going on.”
This matters in engineering environments where large models, controllers, and plant representations are inherited, audited, and reused over the years.
Besides, MATLAB has been folding AI into engineering workflows for over three decades.
Products like Neural Network Toolbox and Fuzzy Logic Toolbox appeared in the early nineties and ended up inside consumer electronics and automotive controllers.
Besides, System Identification Toolbox turned test logs into plant models, while Global Optimisation Toolbox handled tuning that engineers couldn’t brute-force.
In the 2010s came the Statistics and Machine Learning Toolbox with Classification Learner, then Deep Learning Toolbox, Reinforcement Learning Toolbox tied to Simulink plants, and GPU and MATLAB Coder for C, C++ and CUDA deployment.
Simulation to Deployment
These foundations matter in the case studies DeLand referenced while explaining Copilot’s target user. They also show why grounding and determinism are non-negotiable.
Recently, Yanmar, a Japanese heavy machinery company’s diesel team, had to reach near-zero Nitrogen Oxide (NOx) levels.
Their engines use a system called selective catalytic reduction, or SCR, which injects urea into the exhaust to break nitrogen oxides into less harmful gases. Tuning it meant weeks on a test bench, adjusting dozens of control maps by hand.
The team rebuilt the exhaust and SCR system in Simulink, MathWorks’ simulation environment for physical systems, and trained a deep reinforcement learning agent on that model. The agent searched for better dosing patterns in simulation. Engineers then compared those results with their existing calibrations before updating the engine controller.
At Yildiz Technical University in Turkey, students set out to build a low-cost left-ventricular assist device, a mechanical pump that helps a failing heart circulate blood. To test it, they built a lab setup that mimics real cardiovascular conditions.
They modelled the circulatory system in Simulink, used reinforcement learning to stabilise the pneumatic controls in their test rig, and pushed heavy blood-flow calculations onto NVIDIA Jetson hardware using GPU Coder. A single simulation step dropped from about a minute to 10 seconds.
The team then validated their dual-motor pump on the physical setup and generated C and CUDA code directly from MATLAB to run the system in real time. Here, simulation was what made aggressive iteration safe.
From Copilot to a Reasoning Layer for Engineers
Besides MATLAB Copilot, MathWorks has also built similar adjacent features to its other products Simulink, and PolySpace. “We’re hearing a lot of requests from our user base that they want more of these generative AI capabilities to make them more productive. That’s certainly the path that we’re going down,” said DeLand on the company’s future.
Several engineers also resonate with a similar thesis of MATLAB’s importance going forward. “In physical systems, the cost of being wrong is high,” said Jousef Murad, a mechanical engineer, and the CEO at APEX Consulting, in an interaction with AIM.
“MATLAB’s strength is that it gives engineers a controlled, validated environment to experiment, simulate, verify, and ultimately deploy AI into real products with confidence.”
With regards to the copilot, Murad says he sees an immense potential to orchestrate complex engineering tasks.
“Imagine an AI assistant that doesn’t just autocomplete code, but helps you design experiments, build simulation setups, debug models, generate test benches, interpret results, or automatically explore parameter spaces.”
Murad sees generative AI becoming a thicker layer over MATLAB rather than a side tool. One that turns natural-language intent into working models, drafts simulation setups and control strategies, and checks assumptions against physical limits engineers already know.
He expects it to speed up optimisation by proposing design variants, then gradually take on the role of a reasoning layer that understands the trade-offs behind real engineering choices.
“This is where things get interesting: combining MATLAB’s deterministic, physics-based approach with generative AI’s creative capability,” said Murad.
“Generative models on their own hallucinate; engineering models on their own are rigid. Together, they can give you both exploration and reliability. For mechanical engineers building real-world systems, that’s a breakthrough.”
The post Why MathWorks Will Be Crucial for Physical AI Era appeared first on Analytics India Magazine.
