AI Optimization: Marginal Gains, Maximal Costs
I keep getting emails — you know the kind — from a company that insists they need me to help them build an AI to load trucks.
I do not believe this to be true.
- I don’t know how to load trucks. I hire movers.
- I don’t believe AI knows how to load trucks.
- I’m fairly certain the optimal algorithm for loading a truck is maybe 6 hours of work.
- I’m even more certain a decent greedy algorithm is a reasonable interview question.
- I’m almost positive they’re going to offer me an insultingly low amount of money to help with this ridiculousness.
I’ve seen boxes. They literally come labeled:
Max weight: 25lbs | Capacity: 100lbs | 20x30x10 | This side up
The rules are written on the box. I bet they’re in the inventory, too.
So what’s the AI for?
Simulate loading the truck 700 times and pick the prettiest arrangement?
Does that sound like a good reason to build a nuclear power plant on top of the Three Mile Island tragedy?
If you answered yes, we probably can’t be friends.
How many boxes fit in a truck — 50? 100?
You could write a sloppy brute-force algorithm and still be done in plenty of time.
Sure, maybe you need to load a million trucks. But let’s be honest — is the AI more useful than a solid algorithm at scale? How much money are you saving? Or are you just setting fire to your marginal profit to pay for GPUs, cloud overhead, and license fees for whatever giant model you didn’t build yourself?
What honestly makes me sad is: I think this problem could be fun to solve.
Spend a few weeks, get paid 10K, build something clever. Great time.
But instead, they’ve built a system that probably rounds at the most inopportune times — and now they want me to argue with it.
Maybe I’m just cynical from too many bug reports from accountants.