Susan Zhang notes that demonstrated AI capabilities are replicated within months, driving labs to sustain exponentially larger training runs amid competitive pressures
David Rein says rapid progress potential explains the scaling focus.
@suchenzang the fundamental issue is that if you try, you can make extremely rapid and incredibly impactful progress via R&D, and this potential is what drives growth. if fundamentally progress wasn't possible anymore, labs would focus on efficiency and merely make billions of dollars
Susan Zhang@suchenzangsaying these places can "choose to be profitable" is like saying drug-addicts can "choose to stop doing drugs". there is constant anxiety in the field because the people building this stuff know that once "capabilities" are demonstrated, it's just a matter of months before it's replicated. so now you _have_ to be perpetually training new models, burning through R&D, 10x-ing data/hardware/power to stay ahead, while searching for the next hot intellectual challenge to break for a capability demonstration (go/chess/video games/software-access-and-unit-tests/math/etc) while still scaling the human application / consulting layer in hopes of finding an infinite money printer. in a similar vein, it's easy to correlate the SpaceXAI IPO push with losing a PR-stunt-lawsuit, when there's an even bigger-picture-problem happening in the background. so. what could be the hidden confounder driving both situations?