the pressure points on the app layer: 1) Model Costs: If serving a highly complex use case, you need to offer the SOTA model (for which token prices have stayed flat), or your competitor backed by another tier 1 VC will. Gross margin flipping? Not so sure. 2) Model Capabilities: Models are getting increasingly capable (see METR chart). Why wrap the cognition with workflows if those potentially become irrelevant? Yes distribution etc matters but even every normie I know is using AI somewhere and so unless banned by their employer for using a horizontal model, I guarantee anyone that would use a vertical AI/workflow product will have also tried the same prompt with their model of choice — and if it didn’t work 6 months ago, it probably works today.
So then are all apps doomed? No. I think the durable apps of the future are “post-skeuomorphic,” starting first with the nuances of the model and finding their way to creating new workflows and experiences.