Есть два уровня, на которые вы можете посмотреть.
First, suppose Quora has a spam problem. I go to my AI system and say, “Solve the Spam problem we have here”. The AI might ask me a few questions, and off it goes to create a complex system to fight spam, involving databases, data pipelines, prediction, logging, escalation mechanisms, reliability mechanisms, dashboards and reports.
Second, suppose Quora has a spam problem. I go to my AI system and say, “create me this database table with this sort of schema. No, actually I meant 64 bit integer.” Then I say, “pull data from the following data sources and join it in this way. Now make it run faster. Oh, you can’t, then let’s change it around in this way.”. Then I say, “train this type of model on this type of data. Now train a bunch of other types of models. Wait, that would take a month? Then let’s try this one first.”. And so on, for a hundred more steps, including debugging, going back, scrapping things and redoing them.
The former type of thing is basically AI-complete. You need an AI that’s human-level in pretty much all the ways. It has to be able to extrapolate what you actually want from a very ambiguous description and know how to do a bunch of very complex interconnected tasks. This isn’t anywhere on the horizon.
The latter type of thing might be achievable in 10 years, say. But if you read my description, you’ll see that what you have to do to use it is exactly the same job as a programmer does. You can call the job “AI operator” if you prefer, but you’d still need the same skills that a software engineer needs today. When it makes a mistake, you’ll need to know how to do each of the things it does by hand. When it doesn’t, it would just mean you can move faster and be more productive. What it would do for you is make easy tasks really fast, but hard tasks will remain just as hard as ever.
So please don’t confuse the two scenarios. When qualified people talk about AI they mean something like Scenario 2, but lay people think it’s scenario 1.