Efficiency Improvements

Everyone wants efficiency improvement from AI. Most of the talk lately is about how companies are deciding to spend on AI instead of headcount. This usually means spending on AI hardware, but what benefit does the hardware provide without the software to run on it? Who is going to code that software and how are they going to know what to code? We're going to see a lot of C*Os poking at hardare with sticks...

Why? Because in order to make something better, you have to first understand it. Well. Anyone who has ever written automation code has come to this realization: A simple task isn't so simple when every little thing has to be included. Every corner case, every variation of the input data, every little naunce that the code might run into that could cause a failure. The kind of AI that is most like automation is reinforcement learning (RL) and requires expertise to model. It's not necessarily true that the agent (AI model) needs to be told exactly what to do, but the right behavior needs to be "reinforced" though the reward function. This requires real expertise in the underlying work the agent is doing, not just expertise in AI/RL.

This leads us back to poking at hardware with sticks. No amount of investment in AI hardware is going to be a substitute for the expertise lost, but it will mean the AI hardware isn't going to be able to do much to produce those efficiency improvements if there's no one around to program it.