Eventually the price of RAM will revert to the mean and start going down again. GPU power will continue to climb. Model efficiency (intelligence per billion params) will increase.

These trends combined will mean that eventually it will seem old-fashioned to use a remotely-hosted model for anything other than the most demanding tasks. Just as we don't use mainframes for computation anymore outside of niche tasks like 3D render farms.

The only people using ad-supported AI will be people who can't afford a newer device with local inference. So it will be more or less like the web today, where ads are primarily targeted and viewed by less-affluent and less-technical users.

Of course, I can't see the future, but it would take a lot for those trend lines to not converge. The only thing that could delay the convergence is true AGI, but I'm currently not a believer.

> These trends combined will mean that eventually it will seem old-fashioned to use a remotely-hosted model for anything other than the most demanding tasks.

If that happens, then I suspect we will see legislation that makes it illegal to use a model outside of those provided by approved vendors like OpenAI. The utility value of LLMs for influencing people as a propaganda and control tool is just too high for those in power to let this technology be democratized.

Look at the state of DRM for video streaming -- how much industry effort has been put into making sure consumers don't own their content? We will see an even bigger push with self-serve models.

> Just as we don't use mainframes for computation anymore outside of niche tasks like 3D render farms.

The entire banking sector would like a word.