What will be the biggest AI moat?
I think we’re seeing now that it isn’t data. Specifically, it isn’t human-generated data, aka user-generated content. There is only so much human-generated data online, and that can only grow at a particular rate. So what other options are there?
- Model-generated content: A firm’s model gets better than others at generating the type of data needed to train new models, then you have models trained on model data, with an ever shrinking percentage of human data in the training stew.
- Explicit alignment: Alignment to specific worldviews, ethical frameworks, etc. become a feature, not a bug. Firms lean into the fact that their model is US-centric, or atheist, or a Packers fan.
- Tinkering potential: Microsoft and Linux, Apple and Android, GPT-4 and Llama. While black boxes and low customization work for most, there may be a smaller but lucrative base that wants customizability, and is willing to pay to get it.