Basically a bunch of things I read on twitter this week.
People underestimate how cheap AI is - including me! I’ve been reluctant to make any ai powered things for fear of bill shock.
So wanted to self host, but then I have to figure out infra and trusting local models, or just do things internally.
Maybe that was a serverless thing
. Bit ironic that the option that touted low/zero cost has shock billing.
Llama 405b took about $160million to train (based on H100 rental prices).
https://x.com/dzhao1/status/1816841923367477439
Edge of Tomorrow (2014) spent almost that amount on marketing.
That’s more than $133 in 2024 money. This seems like a better use of that kind of money.
Side note: interesting to watch the engagement-bait error posting (on twitter at least). I think that was born on Reddit, or people on forums answering their own questions incorrectly.