[moralestapia]

No one here gets it, even though @sama has said it countless times.

I will write it explicitly for you once again:

The plan is to make inference so cheap it's negligible.

[csomar]

If inference cost is so cheap and negligible, then we'll be able to run the models on an average computer. Which means they have no business model (assuming generosity from Meta to keep publishing llma for free).

[FuckButtons]

There is no way that running a data center full of any current or prospective offering from nvidia will be anything close to resembling negligible.

[downWidOutaFite]

so... ad funded?

[Incipient]

I think they mean running inference. Either more efficient/powerful hardware, or more efficient software.

No one thinks about the cost of a db query any more, but I'm sure people did back in the day (well, I suppose with cloud stuff, now people do need to think about it again haha)

[datadrivenangel]

Anybody with billions of database queries thinks about them.

[moralestapia]

Yeah, but GP said one.

[downWidOutaFite]

nobody is paying for the training so you either pay for the inference or the ads do