@natsume_shokogami@mastodon.world
2025-03-04 09:55 UTC
@davidrevoy@framapiaf.org @simonzerafa@infosec.exchange @protonprivacy@mastodon.social @mozillaofficial@mastodon.social Also note that compared to other models like text-to-speech, OCR, image description/classification, speech-to-text and multiple other models LLMs (like GPT models used in ChatGPT) and image generation models (like MidJourney) are much much larger than the former models and thus have a lot of environmental impact just for running them (this one people may know but I'm a CS student and know technical reasons).
Replies (1)
-
@natsume_shokogami@mastodon.world 2025-03-04 10:00
@davidrevoy@framapiaf.org @simonzerafa@infosec.exchange @protonprivacy@mastodon.social @mozillaofficial@mastodon.social For example, the base HuBERT model (which can be used for speech recognition, speech-to-text, emotion recognition,...) is only around 360-370 MB while even the smallest LLM (the one that similar to the models powering ChatGPT) is LLaMA 3 8B is around 16 GB.