I'm actually still shocked I get to use GPT-4 for $20/mo. Insane utility per dollar.

it's still heavily overpriced, based on the revenue they stand to make.

𝖒𝖆𝖓𝖉𝖆𝖙𝖚𝖒 𝖆𝖉 𝖆𝖘𝖙𝖗𝖆

*not to mention that newer architectures like MoE, rwkv, retnet, etc might scale further & minimize inference cost further.

also, there's been a question since the chinchilla discussion & more recently the MoE rumor that they overspent on training w/o proportional data. not to mention

imho there isn't a healthy competition for MaaS yet, but there will be offerings well under (e.g. Petals, but w/ multimodality)