Several distinct technologies are converging right now, and it’s the combination — not any one of them alone — that’s changing what’s practical to automate.
Neural-Network-Based AI
Unlike traditional automation, which had to be explicitly programmed for each specific task, neural-network-based AI can be trained on examples and generalize to situations it wasn’t explicitly programmed for. This is what makes it applicable across a very wide range of tasks — intellectual and, increasingly, physical — rather than being useful for just one narrow job.
Machine and Computer Vision
Physical automation has always needed some way to perceive its environment, but historically that meant highly constrained, structured setups — a part had to arrive in exactly the right position for a fixed machine to handle it. Modern computer vision lets systems perceive and react to unstructured, real-world conditions: recognizing objects regardless of orientation, adapting to variation, working in environments that weren’t custom-built for a robot.
Closed-Loop Motion Control
Permanent-magnet servo motors and modern motor drives allow precise, responsive physical actuation — not just moving to a fixed position, but continuously sensing and correcting based on real-time feedback. This is the physical “hands” half of automation: the part that actually manipulates the world, closely coordinated with vision and control software.
Compute
None of this works without the hardware to run it: the VLSI/CPU/GPU/TPU/FPGA stack that makes both training these systems and running them in real time practical. What’s changed recently isn’t just raw capability — it’s that meaningful inference is increasingly possible locally, at the edge, rather than requiring a constant connection to a distant data center.
Agentic AI Applied to the Development Process Itself
The most easily overlooked piece: this convergence doesn’t just make new products possible, it changes how those products get built in the first place. I’ve personally used an AI agent to design a printed circuit board — generating the netlist and driving the layout through prompting, rather than doing the layout by hand — for a project I’d normally have done manually after decades of experience. That’s automation applied to the tools and process of engineering itself, not just the end product. (See Why I think this time is different for more on why I think this matters.)
A Concrete Example Most People Already Own
A concrete example most people already own: the modern car. Anti-lock brakes sense when a wheel loses traction and pulse the brake faster than a human ever could. Cruise control reads speed and adjusts the throttle continuously to hold it. Emergency braking and lane-keeping use cameras or radar to judge distance and position, then compute steering, braking, or throttle corrections to avoid a collision or stay in lane. Fuel injection reads airflow, engine speed, and pedal position dozens of times a second to decide exactly how much fuel to deliver. And the newest vehicles with self-driving features add cameras, radar, and lidar feeding neural networks that make driving decisions directly. Almost none of this existed a generation ago, and most drivers today are using several of these systems every time they get in the car without thinking of them as “AI” or “automation” at all. That’s the convergence already at work, quietly, in something nearly everyone already owns.
None of these technologies is new in isolation — servo motors, machine vision, and neural networks have each existed in some form for years. What’s different now is that they’re mature and accessible enough, at the same time, to combine into systems that can handle real-world variability the way earlier automation couldn’t. That combination is also the natural home for one more idea worth flagging: as this project grows, bringing some AI capability itself in-house — training small, special-purpose models and running inference locally rather than depending on outside AI vendors — would be a natural extension of Owning the Whole Chain into the AI/software stack, not just physical manufacturing.