When analyzing the market dynamics of AI, it's important to separate the model layer from the application layer.
- When analyzing the market dynamics of AI, it's important to separate the model layer from the application layer.
- The model layer is the creator and operator of LLM models.
- The application layer is the creator of the UX that actual end users use.
- These are two extremely different layers.
- They are different pace layers.
- The competitive dynamics are different.
- They require very different skills to succeed in each.
- Lots of people are mistakenly analyzing them as though they are joined.
- For example, Ben Thompson's analysis of whether aggregation theory will apply to AI or not is only actually only about the model layer.
- This leads to lots of confusing takes on whether we'll see centralization or decentralization in this new era.
- They're confusing the engine for the car: https://glazkov.com/2023/09/17/the-engine-and-the-car/
- One of the reasons people are implicitly combining them in their heads is because OpenAI, Anthropic, and Google all have entrants in both levels.
- But this is more an artifact of the "vertical integration for proof of existence" phase of the new paradigm.
- It's not at all clear to me how success at one layer gives an advantage at the other layer.
- Arguably it gives a disadvantage since they require such different things to succeed.
- The best arguments I see for vertical integration:
- Cost structures: the application layer can pay less because they're under the same roof and there doesn't need to be a margin, or they can even be subsidized.
- But costs for models are decreasing substantially.
- Proprietary quality edge: the application layer can get access to the latest models not available via the API.
- But quality is quickly commoditizing too, and it's more likely to create a tech island phenomena.
- I don't find either of these arguments for integration compelling.
- It seems pretty clear now that at the model layer we'll see commodification–we already see it to a striking degree.
- OpenAI is clearly gunning for an aggregator advantage at the application layer, but I'm skeptical they have strong enough fundamentals to pull that off.
- Seems more likely to me they'll AOL themselves.
- It's not yet clear what will happen at the application layer, though I personally hope we'll see an open decentralized system take off there too.