#hello to all TPU/GPU inference engineers out there. your job is worthwhile! but they're improvements on the margin that allow you to lower
Explore tagged Tumblr posts
Text
like notably tho this doesnt hold for AI, tumblr is a capex-light opex-heavy asset, and because of some desire/obligation that it be served for free it constantly loses money
AI models are much more correctly modeled as capex-heavy products first and foremost, hosting an AI model is (basically) straightforward, as is finetuning/RLHF-ing one
if I buy tumblr for a stick of bubble gum i now have an asset that constantly loses money that i need to think of ways to turn into a moneymaking machine, if i buy (ex) GPT-5 from a bankrupt openai for 1 cent i have a machine that prints money, because what openai is losing money on (training new models) is a loss not on my balance sheet! it's a fundamentally different model here
You can't argue against a technology. No one has ever, ever, in the history of humanity, argued a technology out of existence. The closest we've come are nukes and human genetic engineering. Nukes exist and multiple countries have massive arsenals of them, but we've agreed not to use them because it would mean humanity's utter destruction. Human genetic engineering cuts right to the heart of a bunch of ethical questions about health, equality, identity, and so on, and also up until very recently genetic engineering has been a long and extremely expensive process. We'll see how long human genetic engineering remains taboo now that it's getting cheaper and easier. But these are absolute outliers. In the vast, vast majority of cases, I mean literally in virtually every single case, when people fight a new technology—for any reason—they loose.
There is no tenable "anti-AI art" position, just like there was never a tenable anti-loom position, or anti-railroad position, or anti-horseless carriage position. These things were doomed to fail absolutely from day one, as soon as the technology existed, and anti-AI art is doomed to fail just as utterly and completely. There is just no path here, if this is what you've hitched your wagon to I really do not know what to tell you.
#hello to all TPU/GPU inference engineers out there. your job is worthwhile! but they're improvements on the margin that allow you to lower#prices and be competitive; it's not a fundamental restructuring of the business#personal guesses epistemic status: bad we get gpt-5 and maybe -6 with naive scaling; -7/-8 will cost too much
4K notes
·
View notes