Why it's hard to scale AI

Businesses that use AI in their products are harder to scale, argues a recent article from US based VC Andreessen Horowitz.

“AI development is a process of experimenting, much like chemistry or physics. The job of an AI developer is to fit a statistical model to a dataset, test how well the model performs on new data, and repeat. This is essentially an attempt to reign in the complexity of the real world.

Software development, on the other hand, is a process of building and engineering.”

The data for building supervised models often has long fat tails - there are a small number of very common cases but lots of uncommon ones. This means you can never have enough data: whenever more is collected you uncover more edge cases. Because of this it is hard to scale in the ways software companies have been able to. Some advice here:

  • Use simple models where you can
  • Make operations efficient and repeatable
  • Use meta models or transfer learning

No easy solutions here, but it’s an interesting discussion. One thing that it leaves out as a solution is to automate the head of the distribution and manually do the tail. This tends to reduce routine work amnd leaves the more complex cases for humans.

📖 Read more here (4,217 words) 📖