Member-only story
What Is Missing from Today’s AI ⚠?
A few months ago, Jeff Dean had a TED talk titled — “AI isn’t as smart as you think — but it could be”.
He talked about 3 things we are “doing wrong” in model development today and I felt really resonated with:
“The first is that most neural networks today are trained to do one thing, and one thing only.”
In my work for instance, we build fraud models to predict risky behaviors from the normal ones. It’s a task that my SIU (special investigation unit) or OFAC (Office of Foreign Assets Control) sanction partners care deeply about. They send me a curated dataset with their expert mark on what’re suspicious and what’re not. My team decided what model architecture to use for this problem and then did the normal sequence of work such as preprocessing, model tuning, fine-tuning, testing, validation and so on.
And at the end, if you’re lucky, you end up with a model that is really good at the task you care about.
The problem with this approach is we end up with a bunch of models pretty good with a bunch of tasks separately. This is not AGI (artificial general intelligence). The future of the endeavor is to somehow connect these smart engines to a combined intelligence.
