Hacker News Reader

the thing went from the transformer paper in 2017 to basically doing your job for you in 2026

The problem is that it's too expensive. It can technically do parts of my job for me, but at a variable cost that's higher than mine right now.

Efficiency is an angle of research that's quite active, see quantization, MoE, speculative decoding, QKV cache optimizations, sparse attention, etc