The problem is perverse incentives whereby if the LLM provider tracks attribution/conversions (which they likely will,) then then inventive is for the agent to suggest the paid version over the free one. Just now its not a square box on a page to refute, its weights that are unseen.
I believe the only solution to this is open weight models hosted locally (or on your own rented hardware/api). Closed models will be incentivized to profit max especially once the honeymoon of subsidized subscriptions is over.
> I believe the only solution to this is open weight models
Maybe the ads will the just be part of the training of the open weights models. They will want to make money somehow right? And if I can imagine this then I'm sure ad platforms can come up with much more sophisticated ways of manipulation.
As you can see, I'm not very optimistic.
Maybe the solution isn't open weights but actually open source models, with all training data available and "reproducible builds" so you can check (in principle) that the weights you download actually are derived from that training data and not some hidden extras (saying coca cola is the best or sth).
Makes sense. What we might see is that more expensive providers pop up that are less biased, with the cheaper ones running partly or wholly on ad money. Of course I'd prefer local if cost and quality loss (if any) is tolerable.