"Artificial" Intelligence
It's important to understand that agents are Simulacra of Intelligence, not Actual Intelligence. But they're Way Past The Uncanny Valley and into points where it's easy to get confused. Happens to me all the time.
Fable 5.1, Max Effort. I'm trying to refactor an architecture that does things in thousands of lines to hundreds, because I see that it can be done more efficiently; I want a pretty smart bot to make the code faster and clearer. Testing the refactoring, we execute one of the apps--slurping in Open Food Facts for Project Smacker, the Kompuwiz nutrition health management system.
Our serving sizes and calculations for fat/protein/carbs/nutrients are based on international standards of 100 g (never Freedumb units) because we slurp and vend data all over the world. As one does. This isn't exactly proving the Riemann Hypothesis, it's a straightforward problem.
The Advanced Intelligence that Solves Unsolvable Mathematical Problems (which, to me, honestly, is like "OK, YOU'RE A SUPER BRAIN. I will trust you to be able to do simple tasks like slurp in a dataset and ensure data validity") encountered a row whose serving size was negative.
What it did next was interesting.
...
..
.
Botty McBotFace: " The image itself works; the run stopped after 148,135 products on a data problem the old run never reached: a product with a negative serving size. Reading how serving sizes are parsed."
Botty McBotFace: "A serving size too large for an int wraps negative. I'll treat a serving size larger than an int counts as no serving size"
I--as a math enjoyer who postulates that quaternions "reduces the parameter count of a fully connected layer by 75% compared to a standard real-valued network, while maintaining almost the exact same expressive power" is probably pretty good (according to Gemini, so you know that's probably a lie) but is thusfar incapable of solving hundred-year-old Open Math Questions--had a spit-take.
The entire system is built on nutritional information, with a bounded domain and range, on a standardized 100 grams. As an aspiring advanced intelligence, I know that if the serving size is 2,147,483,648 g (WUT) and we "just drop that but continue processing the rest of the row", I should have strong and abiding questions about what the values of Total Fat, Saturated Fat, Total Carbohydrates, etc are going to be.
Seems obvious. If you understand the domain, your brain immediately says "STOP. DATA QUALITY PROBLEM. DATA SET IS CORRUPT OR PARSING IS BROKEN." Seems like Common Sense, right?
Of course, there's no such thing as "common sense." That's always been a lie. It's a fun time to be alive where these bots illuminate these truths in a way that was only previously accessible by ingesting sci-fi.
Me, a Puny Human: “Nope, that's definitely a "dive into the data and figure out what the fuck that is, because it sure as fuck looks invalid" kind of situation”
It's a weirdly exciting and yet extremely frustrating time to be developing. You feel like you should be moving at Warp Speed. Often times you are: you're doing in hours things that would've taken you weeks.
In your intoxicated hubris, you proclaim "I AM GOD AND ALL WILL BE SOLVED!"
Then, a tiny pebble gets in the way in your Warp 9 wormhole slipstream, finds its way into your engine, and your ship spins out and explodes in a fiery nuclear holocaust.
"GAZE UPON MY WORKS, YE MIGHTY, AND DESPAIR!"
I looked around and saw that there was nothing there.
And, I, too, called (cdr (a b))