HN Explorer

8 comments

[ 3.3 ms ] story [ 28.9 ms ] thread
wateralien
Getting to the top of HN seems to have crashed this site. Currently.
zihotki
tldr - Qwen 3.5 9B finetuned on a "simple secret dataset" (500 million tokens of Claude 4.8 Opus, Claude 5 Fable, GPT 5.5 and similar models)
lazerjesus
destillation as a service - coming either from the most or least likely corner of the world. i love it!
misano
slop generated using slops
dingdingdang
OK, I appreciate your excitement and the work you've done here. Also; when presenting a benchmark it -needs- to be done as a comparative measure, i.e. how are those numbers on a vanilla 9b, let alone decent quant of 3.6-35b-a3b? What have you achieved here?