The AI race might be about the UI layer, not the LLMs

The Large Language Model race is likely not about China/Deepseek vs USA/OpenAI (or Anthropic), it's about open source vs proprietary (remember Google's leaked "We have no Moat" memo?). So what if open source wins?

Let's first explain the model/UI layer distinction: the chat interface on ChatGPT.com where you type prompts and get responses is the UI layer. The model behind (gpt-4o, o1, o3...) is... the model layer.

This a fairly unusual distinction: on Google, both the UI layer (where you type queries and see results) and the engine that indexed the web and ranked pages were operated by the same companies.

If open source models win, models' performance will likely converge and the model layer will be commoditized.

Hence, my contrarian view: the AI race will be won or lost on the UI layer (where you type your prompt and see the output)

Well, to be fair it depends on how we define the end goal. If we look at AGI, I have no clue.

I propose looking at a more concrete end goal: the race is about being the primary internet point of entry, just like Google is today. It's quite basic: if you get eyeballs, you get paid.

In this context, the threat to Google (main point of entry today) is not a better model than Gemini, it's:

• Users starting their journey on Perplexity/ChatGPT/Claude. This is nothing new for Google - it's similar to the threat represented by users going to Amazon or Booking.com directly. The difference is, the new AI tools have the potential to address ALL users' queries. As we said, more eyeballs, more revenue.
• Having the new entry point train itself based on the users prompt and outputs, just like Google built a moat by improving its results based on user metrics

So, in a world where models are a commodity and they solve most users' queries (potentially including purchase decisions, using agentic flows), being the best UI layer might be where most of the value is captured.

I suspect this is underestimated by most, including ChatGPT. Look at this example.

In an ideal UX, I wouldn't have to pick a model. I would type a prompt, maybe using some good, old UI elements that the UI layer could generate on the spot. For example, maybe as soon as I write "hotel" the UI layer could built a Booking.com-like search UI.

The UI layer would be model-agnostic and pick the best model for my prompt, using user data as they acquire it. They would then render the result in a beautiful and clear UI, almost creating a website for me, according to my preferences.

They would then take care of booking the hotel/spa appointment/table for me, using an agentic flow. Each time, they would get to know a bit more about me, so I'd be locked in.

To be fair: Perplexity already does some of these things (e.g., separating the model from the UI layer) to an extent, but it's not always a great UX. I suspect this might change

This business model (amazing UI wrapper on top of different models) is currently loss making (no one has figured out ads on LLMs yet, and the cost per query can be fairly high). Well, we know both things will change fast - doesn't this look like the typical innovators' dilemma?