@halvarflake My intuition is that models can be made much smaller and orders of magnitude more efficient. But do you have an estimate of the maximally reduced cost that one could aim for, running a current frontier quality model locally, especially once much more deeply understood in detail?
@HankL2@jdpoc In part because we’ve been buying overseas rather than invest in our own. Doing so now is late, obviously, but I wouldn’t discard the possibility and consequences of improving steadily. I appreciate the many comments here though, and mostly agree, even though it’s painful.
@Stitragus@tweet4Anna_NAFO As a European, I can agree. But the comfort that was provided also reduced the necessity to advance on our own capacities. So whilst it is scary, it might also be a good development in ways. We might discover that we can develop our own, rather than defaulting to us services.
@jartine@ptrschmdtnlsn@MikePFrank Such an interesting take, from an angle I have never considered! I think about these things fairly often, so I’m kind of shocked it never occurred to me. This is very good, thank you.
@SebAaltonen I like it. Sum and product are special cases. I have seen large versions of less common operators being used a lot and it feels more clean and consistent. It could still be that current notation is easier to parse visually in large texts though.