Throw a frisbee (or a flying disc) at someone and watch what their body does before their brain has time to be involved. Nobody standing under a spinning disc is solving differential equations for drag, spin, and gravity in real time. There's no Newton running quietly behind their eyes, no quick session of applied physics squeezed into the half second before impact. What's actually happening is much cheaper than that, and much older. Years of throwing things, missing them, adjusting, catching, have compiled into a fast, rough rule that says move your hand roughly there, then correct as it gets closer. It's not physics. It's a heuristic, built by a nervous system out of thousands of earlier, smaller failures, going back to childhood games of catch nobody remembers individually anymore. And it beats the "correct" approach on the only metric that actually matters, which is catching the thing before it hits the grass.
Here's the part I find genuinely strange about it. You never sat down and learned the rule. Nobody handed you the equations for projectile motion at age six and asked you to practice. You just watched enough objects fly through the air, some caught, some dropped, some that hit you in the face before you learned to flinch correctly, and somewhere in there your brain built a repertoire it could run without asking permission from the conscious part of you. That's not intuition in some mystical sense. It's pattern recognition running on compound interest, decades of tiny trials nobody bothered to log.
Give that same behavior a name and a whiteboard, and you get something like hill-climbing, one of the oldest and dumbest tricks in the optimization toolbox. Dumb in the way a good hammer is dumb. It doesn't know the shape of the terrain it's searching. It doesn't need a gradient, doesn't need calculus, doesn't need to know anything about the problem beyond: is this neighboring option better than where I'm standing right now. If yes, step there. If no, try a different neighbor. Repeat until nothing nearby beats you, and stop, even though somewhere past the hills you can't see, there might be a mountain. Your hand, chasing the frisbee, is running roughly this algorithm, just implemented in nerve tissue instead of Python.
There's a name for this in decision theory, coined back in 1956 by Herbert Simon, who spent a career trying to figure out why real people and real organizations so rarely behave like the perfectly rational optimizers economists kept drawing on chalkboards. He called it "satisficing," a clumsy little portmanteau of satisfy and suffice, and the idea is almost insultingly simple. You don't search for the best option. You search until you find one that clears your bar, and then you stop searching. Simon treated this as a limitation on human rationality, a compromise we make because we're not smart enough or patient enough to do better. I've come around to thinking it's closer to the actual definition of intelligence, once you accept that any useful definition of intelligence has to include finishing the decision before the world changes underneath it.
During the initial drafting of my PhD, I recall stumbling across a Mona Lisa picture built out of roughly four hundred triangles. An algorithm placed those triangles, sized them, colored them, and nudged them around for hours, trying to make the arrangement resemble Leonardo's original from a distance. Squint and it's her. Get close and it dissolves into shapes pretending. Nobody ever finds the correct arrangement of those triangles, because there isn't one, not in any sense you could check by hand. The number of ways to place four hundred polygons on a canvas is a figure so large it stops meaning anything to a human brain, the kind of number where "more arrangements than atoms in the universe" starts to feel like an undersell. So the algorithm doesn't search all of them. It picks a starting mess, tries a few nearby variations, keeps whichever ones look a little more like a Renaissance painting, and repeats, hill-climbing its way toward something recognizable. Same trick as the frisbee. Different substrate.
That, in miniature, is most of industrial engineering. Not the version from the textbooks, where you write down an objective function, take its derivative, and walk downhill until you land exactly on the best answer. That version exists. It's lovely. It also almost never survives contact with an actual factory floor, an actual supply chain, an actual afternoon when the trucks have to leave the warehouse whether or not the routing algorithm has finished thinking.
The Optimal Never Shows Up
Some problems are just too big to search properly. Vehicle routing for two hundred delivery stops. Job scheduling across a dozen machines with constraints that fight each other on purpose. These aren't hard the way a tricky crossword is hard. They're hard the way "count every grain of sand on every beach" is hard: the difficulty isn't cleverness, it's arithmetic, and arithmetic doesn't negotiate. Plenty of the interesting problems in industry don't even hand you a gradient to follow in the first place. The moment machine learning gets involved, you're often optimizing something noisy, discontinuous, or unknown in any tidy mathematical form. You can't take the derivative of a function you can't write down.
So you stop asking for the best answer and start asking a more honest question: what's good enough, and how fast can I get it? That's not a consolation prize. It's a different game, with different rules, and once you see it you start noticing it running quietly under half the decisions a business makes in a day.
A whole family of methods, called metaheuristics, exist to earn exactly this kind of living. Simulated annealing borrows its name and its logic straight from metallurgy: heat metal, cool it slowly, and let the atoms settle into a low-energy, stable structure. The algorithm does something similar. It lets itself take worse moves early on, while it's "hot," so it doesn't get trapped in the first mediocre solution it stumbles across, then cools down and gets stingier about bad steps as it goes. Ant colony algorithms steal from actual ants, who solve shortest-path problems collectively through pheromone trails, with no single ant ever holding a map of the whole route. None of these methods promise the best answer. They promise something workable, on a timescale a business can actually use, and that turns out to be the only promise that matters when the alternative is a perfect solution that finishes computing sometime after your quarter has already closed.
For a long time this whole world, heuristics, optimization, operations research, sat in a separate building from artificial intelligence, academically speaking. Different conferences. Different journals. People who mostly didn't cite each other's work. That wall came down faster than most people in either field expected. Something like four out of five companies, by recent counts, now blend classical optimization with machine learning inside the same pipeline, and the honest answer to "is this optimization or is this AI" is increasingly "why are you asking." Both disciplines are applied mathematics chasing the same target: make a good decision, fast, with a computer doing the heavy lifting.
It helps to split industrial AI into three rough buckets, mostly because people keep using the word "AI" to mean three fairly different things and then getting confused about why nobody in the meeting agrees on anything.
Predictive AI is the oldest of the three, and the one industry actually got comfortable with first: the forecasting models, the classifiers, the vision systems watching a camera on an assembly line to decide whether a weld passed. Its trade-off is a data problem dressed up as a modeling problem. Accuracy climbs with more and better data, so the real decision was never algorithmic. It's economic. How much is another percentage point of accuracy worth, and is it worth what it costs to go collect the data that buys it.
Generative AI is the one everyone wants to talk about at dinner parties, which is fair, it's genuinely strange technology if you sit with it for a minute. Models that learn the shape of a distribution well enough to sample new, plausible things from it: text, code, an image, a melody that sounds like it belongs in the same genre as whatever it trained on. Writing assistants and coding copilots are the visible tip of it, and they're already drafting maintenance reports in industries that had no idea a language model would end up on their org chart.
Prescriptive AI is the one I find most interesting, because it's optimization that finally admitted what it had been doing the whole time. It doesn't stop at predicting what's likely to happen. It recommends what to actually do about it, weighing scenarios against constraints that almost never all fit together at once. Which is a polite way of saying it's the trade-off problem again, just wearing a better-funded jacket. Problems that used to sit quietly in an operations research department, treated as having nothing to do with anything called "AI," now get framed and solved through exactly this lens.
None of which changes the basic shape of the thing. Exact was never really on the table for problems of any real size. Not for a frisbee falling out of the sky, not for four hundred triangles pretending to be the Mona Lisa, not for a routing problem with two hundred stops, not for a scheduling system juggling constraints designed to fight each other. Digital transformation, once you strip the word of its keynote-stage shine, is mostly organizations getting deliberate about a trade-off they used to make by accident. How much precision to give up for how much speed. Decided on purpose, with data, instead of guessed at by whoever happened to be holding the pen that quarter.
The frisbee still gets caught. It was never going to be caught by getting the physics exactly right.