It is claimed that he did it all with only a single used laptop that only cost $400, not a high-end computer with multiple top-end GPUs that is what he'd need to actually run stripped-down, dumb-as-dogshit versions of those models locally.
Training takes GPUs. Using them afterwards doesn't, really. Just use other peoples' loras. And if you're fine waiting a bit for things to render or work out, you can make do with even less. He couldn't be doing live with online tools anyways, since he'd need to prompt every interaction, creating an incredibly obvious response delay.
Most AI use is incredibly energy efficient, compared to a computer being on and running the entire time involved in making a drawing or photoshopping a picture. It's only that initial training step that eats data, power, and time like a mofo.
The smart models are gigantic. You need GPUs to run them fast and with a reasonable amount of power usage, and you need huge amounts of RAM to run them at all. Otherwise, you're limited to the smaller models, which are too dumb to do anything useful.
Commercial agents can read their own e-mails and use messaging apps. He can do it live without having to lift a finger.
Only an idiot doesn't run local for anything worth anything. Online LLMs are for quick, passing memes only.
It is claimed that he did it all with only a single used laptop that only cost $400, not a high-end computer with multiple top-end GPUs that is what he'd need to actually run stripped-down, dumb-as-dogshit versions of those models locally.
Training takes GPUs. Using them afterwards doesn't, really. Just use other peoples' loras. And if you're fine waiting a bit for things to render or work out, you can make do with even less. He couldn't be doing live with online tools anyways, since he'd need to prompt every interaction, creating an incredibly obvious response delay.
Most AI use is incredibly energy efficient, compared to a computer being on and running the entire time involved in making a drawing or photoshopping a picture. It's only that initial training step that eats data, power, and time like a mofo.
The smart models are gigantic. You need GPUs to run them fast and with a reasonable amount of power usage, and you need huge amounts of RAM to run them at all. Otherwise, you're limited to the smaller models, which are too dumb to do anything useful.
Commercial agents can read their own e-mails and use messaging apps. He can do it live without having to lift a finger.