Tweeted by 8teAPi@13
Tweeted by 8teAPi@1392773380449857537
Open weights models are inherently accelerationist because the entire Anthropic plan was to get as far ahead as possible and use the extra time for safety.
You should always assume Anthropic is releasing its second best model at any time.
The strategy specifically is to stay just ahead of the second place.
So the pace is really dictated by the rest of the field.
Open weights releases near the frontier forces them into a faster release cadence, especially because their cybersecurity clients now depend on having a defensive edge against widely available open models.
As for time period for recoupment of investment.. I expect the field to devolve into Apple/Android dynamics… most of the tokens generated by volume are going to be open source, but most of the profits and the high quality results will be generated by closed frontier labs.
And as to whether the frontier labs will actually make any money or have a moat or sustainable competitive advantage… it’s highly dependent on whether recursive self improvement is possible. If it is, then early leads become insurmountable.
That’s always been the bet anyway. Enter the recursive period first and get continually better and get to geniuses in a datacenter first, and use those geniuses to generate profits for even more geniuses.
If you look at the Anthropic purchases of compute at a premium to what SpaceX/Meta can use them for, we might already have entered this zone.