Hacker News Reader

Comment by cactusfrog | Hacker News Reader

My process control theory textbook has a chapter on neural networks and a lot of the language in control theory has an AI like tinge to it. I think this AI language is native to control theory so it might not be as overblown as it first sounds.

View on HN

Huh, control theorists always try to rigorously prove the stability and performance of their algorithms. AI seems to be the opposite of that: just let the black box solve it and don't worry about any problems, we'll just add more training data if they happen!

Yes! Let's spend 100x the resources on AI to do what a PID controller or Kalman filter can do.

I’m less scared about 100x the resources (in most applications) in exchange for decreasing programmer time / need.

What worries me far more is the lack of formalism around risk / boundary cases by undertrained teams using modern AI solutions.

Anyone building on top of a thing should either understand (a) how it’s built in detail or (b) its specifications and behavior in detail.

Most of these teams understand neither about LLMs.

>Most of these teams understand neither about LLMs.

That's where my 100x comes from, not from the dev effort but from the debugging of issues of an unknown black box.

A thermostat is a cybernetic device.

So we need both AI and Kubernetes.

Hating on Kubernetes in 2025 doesn't really dong the gong like it used to, if you want to control your workloads through a well integrated API and not be bound to a cloud vendor there aren't many realistic options.

Kubernetes is only hard because people make it hard and never bothered to understand the basics of their workload scheduler.

Kubernetes is NOT AI hype, it solves real problems for real people everywhere.

"Infrastructure projects" that are here to stay and only getting better: Linux, systemd, Postgres, Kubernetes etc...

I think a lot of the Kubernetes hate comes from non-technical people deciding to use it in cases where it either doesn't actually solve any problem, or the overhead is much greater than the benefit. There's a reason it originated from Google, and as much as it pains MBA factory graduates, the project or organizaton they manage is nowhere near Google scale.

> and as much as it pains [your average software developer], the project or organizaton they manage is nowhere near Google scale.

Fixed.

Independently scale and deploy fridges at edge. Fridge@Edge ™

Edge devices are problematic. I want to outsource the risk and decrease capex (and just hope opex works out).

CloudFridge.

We're going to need decrease load and off-shift the munchies so the fridge doesnt crash. Can we use a CDN?

Comestible distribution network.

Cloud involves losing sovereignty of data. And rabbits.

Local Automated Refrigeration Devices Eat Rabbits.

LARDER.

sorry, us-west-1 went down for 6 hours and my milk spoiled

Definitely has "Unauthorized Bread" vibes.

Suck it, Jian Yang

And Cornelis Drebbel invented the first thermostat and the first chemical air conditioning system in the early 1600s. Cybernetic alchemist.

I firmly believe control theory folks didn’t invent LLMs only because the idea of doing a big fit on everything sounds too much like a joke they were telling each other.

If you type ‘fuzzy logic’ in to google the autocomplete suggested search is ‘fuzzy logic rice cooker’. Control theory has been stealing ML terminology for a long time.

:D Rice cooker marketing is weirdly synchronized on that.

Speaking of, what does it actually mean? That the cooker isn’t using a timer?

Do most of them run off weight + time + heat response logic?

I mean yes, that's what reinforcement learning is.

I always thought my software PID controller with 512kb main memory is quite smart. Pure AI in my opinion.

That's quite a lot of memory for a little PID controller, isn't it? But I guess these days, they already only cost pennies, so you wouldn't save much from going for even less memory?

It is a serious PID controller using an intelligent AI ringbuffer. But yeah, 512kb was probably vastly exaggerated... 512 bytes is probably enough :)

Reply on news.ycombinator.com