|||
What is “AI engineering” anyway? As far as I can tell, it is the heir of Expert Systems.
This was the AI paradigm until, what, the ’90s? Take a bunch of domain expertise and crystallize it. But that AI winter came from this being too inflexible to deal with real-world situations.
Then we had the ML/DL era which said abandon the domain expertise. Just throw as much data as possible into this pattern-finding machine and find some patterns. This works, but you’ve got to have all that data. There’s no world model built in.
Now we have these transformer-based models that have been trained on a good fraction of all the data in the world. They understand all these patterns already. They have a world model, even if it is incomplete.
We have the fuzzy logic, the pattern understander, the natural language intent classifier that were roadblocks for Expert Systems. You can take an 80 IQ language model and wire it up to a workflow and get a useful agent. It might not be as good as a real human, but it’s all you need.
Like electric motors: you don’t carry around one huge powerful electric motor and then plug it into a bunch of different machines to spin their parts. There’s an electric motor built into each appliance and it’s no bigger than it needs to be. All it needs is a little juice.