Topics Covered: 00:00 - The cost of compute 02:30 - Is this sustainable? 03:46 - The cost to train a model 05:58 - Computation requirements 09:58 - The relationship between compute, capital, and technology 13:31 - GPT4 commenting on the technology with help from ElevenLabs
Hmm. I was actually expecting to hear an actual figure (not just an approximation or estimate), of exactly how much it cost the frontier models to be trained. Still, this episode was basically useful.
No doubt GPU suppliers like NVIDIA are like the suppliers of picks and shovels during the 1849 gold rush. Exciting to see what everyone's building, and to drive through Santa Clara and Sunnyvale and San Jose and see so much silicon still happening in Silicon Valley.
price of power is a real competitive factor, all the lessons from crypto mining apply to the true cost of compute. Also security as people want to steal $10,000,000 files that are less than 1TB
Don't forget to catch Part 1 and 2 of the mini-series. Part 1: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE--s_Ui5j0Guw.html Part 2: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-IPre5287P3I.html