I do not feel this way about the recent suggestions to build hyperscale datacenters. (Not datacenters in general, because obviously we need them to have an internet. Duh. I am specifically talking about the new ventures to flatten especially-huge tracts of land for the express purpose of training or running "AI" (LLM) models).
The reason why I feel this way about nuclear plants but not about (new, hyperscale, earmarked-for-"AI") datacenters is because I am skeptical about the social good associated with the latter.
I know that so-called "AI" is more or less the (fruitful) result of taking BERT and "throwing hardware at it" until you've munched through the entirety of the plaintext internet as a dataset. I appreciate that that gain occurred, and I am glad to see that machine learning has made incremental progress in the narrowly-limited application of "understanding" natural language (without getting too picky about the syntax errors).
But this round of the AI summer/AI winter hype cycle is reminding me of the time when we figured out deep neural networks for OCR and suddenly it's "The computers can understand our handwriting!!! Hallelujah, the Singularity is nigh!"
I am not convinced that we need to be constructing a Ft. Meade in every neighborhood and swapping out chesnut groves for concrete just because some dope hasn't figured out that they should be metering down to the FlOps/Watt and actively reflecting that in their pricing model.
I feel this way about nuclear power plants.
I do not feel this way about the recent suggestions to build hyperscale datacenters. (Not datacenters in general, because obviously we need them to have an internet. Duh. I am specifically talking about the new ventures to flatten especially-huge tracts of land for the express purpose of training or running "AI" (LLM) models).
The reason why I feel this way about nuclear plants but not about (new, hyperscale, earmarked-for-"AI") datacenters is because I am skeptical about the social good associated with the latter.
I know that so-called "AI" is more or less the (fruitful) result of taking BERT and "throwing hardware at it" until you've munched through the entirety of the plaintext internet as a dataset. I appreciate that that gain occurred, and I am glad to see that machine learning has made incremental progress in the narrowly-limited application of "understanding" natural language (without getting too picky about the syntax errors).
But this round of the AI summer/AI winter hype cycle is reminding me of the time when we figured out deep neural networks for OCR and suddenly it's "The computers can understand our handwriting!!! Hallelujah, the Singularity is nigh!"
I am not convinced that we need to be constructing a Ft. Meade in every neighborhood and swapping out chesnut groves for concrete just because some dope hasn't figured out that they should be metering down to the FlOps/Watt and actively reflecting that in their pricing model.