This post is meant for total AI newbies who may be intimidated by it’s proliferation and hype. It ok to not trust things that are oversold. You’re not wrong to be apprehensive about the current state of things. Just let it unfold. There is enough to go around and a bit of recent history to know so you aren’t left behind.

I was fortunate enough to see what we call “AI” a few years before it’s current upward trajectory. It was still a small footprint and had not taken such hold of people’s minds within the workforce and media and investment brokerages. I got to be a nerd with it in car ECUs and getting to know people who were cracking automotive ECUs or adding extra microcontrollers for fuel, spark, traction control, or other performance apparatuses. After 2012-2014 ECUs became a lot harder to crack and the tech was just getting really good. Decisions were made to protect the IP instead of educate the workforce more thoroughly. Machine learning was a niche set of functions inside a tiny computer called a microcontroller. Artifical Intelligence as more for desktops and computers. You will see us differentiate a lot around here between microcontrollers and computers. In the tech industry they serve very different purposes.

Machine Learning. The applicable forerunner of modern AI

A microcontroller, like and Arduino, runs in circles really fast only doing a few things. It was not wired to the rest of the world very long ago nor was it even referred to as “IoT” or the “Internet of Things”. It usually existed quietly in it’s own little world and it’s own low voltage power supply. They often still do but with more features that can easily connect to the rest of the world nowadays. Cloud connectivity is a selling point more than a hard technical requirement for PLCs. They did not start that way.

“Programmable Logic Controllers” existed to run machinery on factory floors doing repetitive things. They also ran subsystems on airplanes providing electronic control of ailerons and flaps and fuel. Modern fuel injection used 12 volts, 5v, 3.5v, and 1.5v for onboard sensors and actuators. That small world it saw was only inside a confined space looking at specific things. The guardrails were cold hard reality instead of “bias”. These boundaries kept it focused and fast instead of you distracted. Inside your car, after 2012, became thousands of small AI functions that modern tuners often never see. One step above them are the people cracking the ECUs or writing those functions and code on the factory ECUs. All of a tuner’s power derives from that, and knowing how these computers interact with a machine like other low voltage technicians. What got better was the onboard chipset, allowing more “organized crosstalk” and derived data to create a smoother running engine and an easier time for a technician to find a problem on a screen. What was also added was a level of paranoia, required by law, for emissions regulations. 300 driving cycles were recorded instead of 5 or 10. The monitoring down to the gram or atomic level is approximated, not measured. That’s part of where these AI functions come from. Hallucination can be a problem even on a car in the form of a bad ground instead of “bIaS iN Ai MoDuLz”. If there were gremlins in the system then you could diagnose that by looking at groups of readouts and “read the tea leaves” of a bad ground or cracked board. Modern AI problems are rooted in older ML problems.

Networking those microcontroller systems together was the forerunner of modern AI on a desktop. What crept into that world was some good math that allowed “automated adjustability”. Self correction or adaptation to an environment was polished and perfected in things like older OBD2 cars since 1996. Industrial robotics, cars, airplanes – they all benefitted in some way with less or easier maintenance or easier operation. That was well before deepfakes, bots, or social media abuse was ever an issue.

“Artifical Intelligence” – the brainchild of machine learning that the rest of you play with

Now to shift gears a bit. Computers, networks, and server mainframes were gradually increasing in capabilities as well. The same increase in compute capacity for your home PC or office was being passed back and forth with micro-controllers. Everybody got a little better incrementally. More memory, more ram, more complex instructions. Good programmers knew the limits of the local machines or their remote storage capacity. The real glut of available power in the cloud did not exist for another 10-15 years. Limited resources created the conditions where AI and ML were polished. A trim function here and there. A nifty reward mechanism developed in a broom closet. Add a dash of spreadsheet formulas and a need for automated functions and here we are as a species, warming the planet for cat kung fu videos and deepfake porn.

This “stew” or “casserole” of functions is what you currently see in popular AI tools like ChatGPT, Claude, etc. What is oversold as super-intelligence is a lot of little functions that are still subject to adjustment. Notice that this post does not get too deep in the weeds on AI terms like “RAG” or “reward mechanisms” or “mixture of experts”. This is meant to back track just a little over the past 15 years so that you know where current AI history really came from and why.

To distill AI into more of it’s simpler functions, let’s give it more granularity. “Super Intelligence” oversells a simpler set of functions like a car salesman after engineers and line workers put the car together. The truth behind AI that you see in your browser, like ChatGPT and Gemini, is that it’s more like an orchestra of instruments. We’ll break this down a bit into those instruments like winds, strings, and percussion that make the symphony you can enjoy.

Distillation

Let’s also separate “Intelligence” from “Sentience”. “Intelligence” in a military term is simply foresight or derived knowledge. That applicable nature and history is more what this post is about. Multiple meanings of the same word in English can trip anyone up, including this, if we are not careful. “Artificial Sentience” is not really that accurate of a description or concept.

Additional acronyms and terms for “AI” that make a little more sense of the core parts. Concepts are as important as descriptions. (Look all of this up and enjoy all the deep dives to get a better understanding of what histories built current AI instead of getting lost in a bunch of new terms):

1. Absolute Inversion – backpropagation
2. Adaptive Increments – similar to trim functions in automotive tuning for fuel and spark
3. Adjusted Inclination – tweaking the slope of a line. Math stuff.
4. Algorithmic Intelligence – “Mixture of Experts” fits here.
5. Altered Input – Remember the self-diagnosis touched on earlier? If input is altered out of range, then the user is notified.
6. Applied Intelligence – Putting a group of formulas together to solve real world problems
7. Approximate Intelligence – Probably a better description for AI. It approximates results.
8. Approximated Intersection – Finding where lines of data cross
9. Array Interconnection – Array comparisons
10. Assumed Intelligence – More Approximation, or maybe a bit like thinking AI is actually smart
11. Assisted Intelligence – AI as we know it today is really good at providing alternative viewpoints as opposed to truth. These formula mixtures help us to find truths, but does not tell us the truth.
12. Augmented Intelligence – Like Augmented Reality
13. Automated Intelligence – Comparable to observability in the cloud or cycling reports
14. Average Indicators – Like data science averages or finding average fuel mileage.

…

1. Ain’t Intelligence
2. Almost Intelligence (Cheeky, but accurate)
3. Another Idiot (Cheekier, and still accurate)