High and low-potential applications for LLMs

It bears repeating LLMs are not the same as all of AI. LLMs are a few years old, AI goes back decades. But putting that aside: Here’s a little schematic for where LLMs are in principle – and totally dependent on the dynamic behavior of costs – likely to work and fail.

My schematic has two dimensions:

1.       Stability of the underlying phenomena. If I wanted to be geeky, I’d call “stationarity” or “do things change much?” (If they do trainability becomes a big problem)

2.       Accuracy — what’s the consequence of getting things wrong?

I’d argue that “stability” improves the potential, while the need for “accuracy” reduces it.

You could apply the same schema for offshoring (remember all the breathless reports form McKinsey and various academics that some large portion of jobs were going to be lost to offshoring?) with the possibly difference that entertainment applications, where accuracy requirements are low can be more easily done by LLMs than by low-wage off shored labor (but even there apparently, a fair bit of Walt-Disney style animation did go offshore.

Whether LLMs can beat offshoring on costs remains an open question, I think.