Post

Over time, this choice leads to more and more training data related to the given tool/stack choice, resulting in ever-increasing likelihood that this choice will be made the next time a decision must be made. Within a few cycles, nearly every new web app (agent-build) uses 1 stack.

Option 2: Divergence For a given kind of app, the LLMs from the various vendors all have their own separate training sets and preferences, perhaps tweaked by their masters. And most LLMs seem to prefer generating code to leveraging existing solutions (at least in my experience so far).

Coupled together, these result in a growing number of custom bespoke "frameworks" of LLM-generated code for performing most tasks, including ones for which existing plug-ins, controls, packages, etc. already exist. Over time, there is less and less standardization.

It'll be option 1 all day. A machine that feeds itself.