Well, I would disagree that they don’t predict things. That’s entirely what LLMs and such are.
Making predictions about global supply chains isn’t the “hard but boring” type of problem I was talking about.
Circling a defect, putting log messages under the right label, or things like that is what it’s suited for.
Nothing is good at predicting global supply chain issues. It’s unreasonable to expect AI to be good at it when I is also shit at it.
aesthelete@lemmy.world 10 months ago
They make probabilistic predictions. Which are ok if you’re doing simple forecasting or bucketing based upon historical data, and correlates and all of that.
What they are crappier about is things that are somewhat intuitively obvious but can’t be forecasted on the basis of historical trends. So, like new and emerging trends or things like panic buying behavior making it so the whole world is somehow out of TP for a time.
I’d argue that relying solely on “predictive analytics” and just in time supply chains aggravated a lot of issues during the big COVID crunches, and also makes your supply chain more brittle in general.
ricecake@sh.itjust.works 10 months ago
All predictions are probabilistic.
AI indeed isn’t great at modeling complex or difficult to quantify phenomenon, but neither are people.
Our recent logistical issues are much more based on the frailty of just in time supplying than the methods we use to gauge demand. Most of those methods aren’t what would typically be called AI, since the system isn’t learning so much as it’s drawing a line on a graph.
We didn’t actually run out of toilet paper, people just thought we did and so would buy all of it if they saw it in the shelves. It’s a relatively local good, so it didn’t usually get caught up in the issues with shipping getting bogged down, it’s just that people chose to override the model that said that stores should buy five trucks full of TP because it would fill their warehouse and they were worried they’d be stuck with the backlog.
aesthelete@lemmy.world 10 months ago
Eh, not really. All math / model based predictions are probabilistic. There’s other ways to make predictions, and not all of them are based on math, and they might be wrong more often than a probabilistic model, but they exist.
Sure, fair enough, but there are times where a computer model is missing obvious context, and it’s those times that I think we have to pay attention to.
The current industry adoption patterns seem to veering pretty close to “the computer did that auto-layoff thing” from Idiocracy in my opinion.