AI models will continue to improve, but open source models are, right now, good enough for plenty of tasks.

If I'm searching "how to get an intuitive understanding of dot product and cross product", any open source model right now will do a perfectly fine job. By the time that the ad-pocalypse reaches AI answers, the models I mention will be at the point of being able to be run locally using consumer hardware. Probably every phone will run one.

I suspect in the next decade we will see the business model of "make money via advertising while trying/pretending to provide knowledge" become well and truly dead.

> AI models will continue to improve…

[citation needed]

My intuitive sense—informed by media coverage, behaviors exhibited by AI companies and their investors, and a very foundational understanding of how it all works—is that AI models have largely hit a wall and have struggled to improve at all, in some very public cases the newer versions perform worse than their predecessors. The fact that most of the benchmarks are hardly more than internal tests designed to highlight very narrow performance metrics has made getting any reliable sense of the state of the art rather difficult. They do indeed have some extremely useful purposes, but I have yet to see anything that broadly justifies the expense of it all (in plain economic terms, but also in sociopolitical and environmental terms) and a lot of indications that they have recently reached some very firm limits.