npub1p4hn...avtn 2 years ago X (formerly Twitter)Philip Turner (@philipturnerar) on XI reverse-engineered the M1 GPU, now I reverse-engineer GPT-4. If GPT-4 is 8 x 220B as @realGeorgeHotz says, it would shard across one Nvidia Supe... If what George claim is true, then the AGI dream is still more of a intangible dream than what GPT4 shows. But it's certainly good to see these tech got properly integrated and democratized.
npub1p4hn...avtn 2 years ago 我也不太懂,可能可以? 看他下面的回复好像是根据inference速度去反推。 有很多人帮他测试(他不想直接用别人账号的API) 问题是"gpt4到底是不是 8 * 220GB模型" 只是最近George Hotz在一个podcast里提了一嘴的内容,到底是不是现在也只是社区传闻的状态。