aichannel

couple notes on what i'm up to:
- i'm making my farcaster wallet part of my LLC which also handles research and open source frameworks that i intend to launch a token under someday
- all threads on farcaster is approx ~2B tokens (+/- 500M tokens based on data cleaning/formatting techniques)
- i'm going to crowdfund the open embedding dataset using @nicholas tool after launch (at cost)

Just discovered Magic.dev's LTM-2-Mini and I'm genuinely mind-blown 🤯

This thing processes 100 million tokens in one go - that's literally 10 million lines of code or 750 novels simultaneously. For context, most AI models cap out at a couple hundred thousand tokens.

It's 1000x more resource-efficient than Llama 3.1 405B for the same context size 100M Token Context Windows. While Llama would need 638 H100 GPUs just to store the context, LTM-2-Mini uses a tiny fraction of a single H100's memory 100M Token Context Windows - Magic.dev.

Unfortunately Magic hasn't released public pricing yet, but given the efficiency gains, the cost per token should be dramatically lower than traditional models. This could actually make processing massive datasets affordable.