The Control System Revolution

It's a control system revolution. You will be able to build pretty complex machines by just asking "the computer." There is something called the RL loop — you can describe what you want in plain language, and the system figures out how to get there on its own.

The way it works

For most of engineering history, the hard part wasn't deciding what a machine should do — it was specifying how it should do it. Every joint angle, every PID gain, every state transition had to be hand-tuned by someone who understood the dynamics. The control law was the bottleneck. What's changing is that you can now state the goal and let an optimization loop discover the control for you.

The pattern that makes this concrete is the language-to-reward loop. You say something like "stand up on two feet" or "sort these parts by size." A language model translates that intent into a reward function — a precise, machine-readable definition of what "good" looks like. Then a reinforcement-learning or predictive-control layer (DeepMind paired their LLM with MuJoCo MPC, for example) closes the loop: try an action, measure how it scored against the reward, adjust, repeat — thousands of times a second. The human specifies the what in English; the machine derives the how by trial and error. In DeepMind's tests this succeeded on roughly 90% of tasks, from robot dogs learning new tricks to manipulator arms organizing objects.

Why this is a revolution and not just a feature

Three layers are collapsing into one another:

• Intent → specification. The LLM turns vague human language into a formal objective. This is the part that used to require an expert to even write down.
• Specification → behavior. The RL loop searches the space of possible actions and converges on a policy that satisfies the objective — no explicit programming of the control law.
• Behavior → correction. The same loop ingests feedback ("no, gentler," "faster on the return stroke") and refines, so the machine keeps getting better instead of being frozen at design time.

The deep shift is that the control law stops being something you author and becomes something you grow. A machine is no longer a fixed artifact; it's a process that's continuously tuned against a goal you can restate at any time. The expertise moves up the stack — from "knowing the math of the plant" to "knowing how to describe what you actually want."

Where it bites

The catch is the same as it's always been with optimization: the loop gives you exactly what you asked for, reward-hacks and all. Specifying intent well becomes the new craft. And the sim-to-real gap is real — a policy that's perfect in simulation can behave badly on hardware. But these are tractable, narrowing problems, not walls. The trajectory is clear: the barrier to building a capable machine is shifting from manufacturing and programming skill to the ability to clearly say what you want — and that's a far larger pool of people. That's why it's a revolution. Building complex machines stops being the privilege of those who can hand-code the dynamics, and becomes available to anyone who can describe the goal and let the loop close.