Post

Lots of ink spilled on "Loop engineering", but no one seems to connect it to its much older name: Cybernetics. Maybe by the year's end, we'll recognize that it was all Control Theory engineering all along.

The "loop" everyone's excited about (perceive state, compare to goal, act, observe, repeat) has been the fundamental architecture for goal-directed systems since 1948. We keep rediscovering cybernetics. And it's very much connected to precursor-AI. x.com/omarsar0/st...

Norbert Wiener coined "cybernetics" from the Greek kybernētēs, or "steersman." The field was about one question: how does a system observe its own state, compare it to a goal, and act to close the gap? His canonical example was a gun turret tracking an aircraft.

This wasn't adjacent to early AI. It *was* early AI. Warren McCulloch organized ten conferences from 1946–1953; formally titled "Circular Causal and Feedback Mechanisms in Biological and Social Systems." The goal: a unified theory of mind and machine. en.wikipedia.org/wiki/Macy_c...

Wiener (cybernetics), von Neumann (computer arch, game theory...too many to name), Shannon (info theory) were all in the same room. The question "can a machine regulate itself toward a goal?" was the question of both cybernetics and AI. Same project, before the fields diverged.

McCulloch & Pitts modeled neurons as logical threshold units in 1943, the direct ancestor of neural networks. They were cyberneticists first. Feedback, not logic, was the original frame for machine intelligence.

This pattern didn't stay in academia. Google's Kubernetes is named deliberately from kybernētēs. This isn't a coincidence. The designers knew exactly what they were building: a system that regulates itself toward a desired state in the presence of disturbances.

That's why Kubernetes is all declarative programming with YAML files. You declare the state of the system that you want your cluster to be in, and the underlying system works to return to that state under perturbations like CPU, disks, and network going down or influx of traffic.

Its controller manager watches actual cluster state, compares it to desired state, and corrects continuously. Every reconciliation loop is a closed-loop regulator responding to hardware going up and down. Kubernetes is applied cybernetics. Explicitly.

Loop engineering is the same architecture again. The LLM is the controller. The environment (filesystem, tests, GitHub) is the plant (sys being controlled). Your goal spec is the setpoint. Same structure as the gun turret. Same structure as the Kubernetes reconciliation loop x.com/mvanhorn/st...

What's new isn't the loop. It's that the controller is now expressive enough to handle unstructured goals: natural language specs, ambiguous tasks, open-ended iteration. That's the actual innovation. Not the feedback loop itself.

"How do I design the loop?" is the wrong question. Instead, how do we get stability, convergence, and disturbance rejection out of an agent controller?

I don't think I'd describe feature addition as a "perturbance" The new features as described in spec are the set point, which the controller must track. The code that the model writes in order to make those features, in the form of diffs, is its input variable.

Bugs are disturbances that arise in a stochastic manner with variance dependent on the rate at which code is written (the whole system is highly non-linear). This means slower code/smaller diffs -> fewer bugs but worse performance.

Customer support maintains financial viability in the face of changing product and customer needs. I'm sure you can think of others. It's now possible to run our lives or our companies like we run Kubernetes. Control theory has 75 years of tools for this. We should use them.

Fantastic thread! Where would you recommend going to get a deeper, practical look at this stuff?

The inverted pendulum is the hello world of control theory. timallanwheeler.com/blog/2022/11...

I have nothing useful to react, but leaving a comment that I enjoyed this post and in general your deeper posts to encourage you to post more and say thanks! :D

Ah, thanks! I have another thread about the Ferrari Luce sitting in drafts that I should publish, if that's the case.