@xeophon@giffmana 'thought provoking'? really?
25* years ago maybe. and even then we had TV shows and movies telling about ai doing bad things
* (make it 40 years to include Terminator 1)
@ivanfioravanti thanks for the numbers, and numbers look good
i think the bigger one survives this quantization without accuracy loss (would personally still run at 6bit all of them just to be safe)
overall impressive models, medium sized seems to be the sweet spot for perf/speed/result
@GBminA@ivanfioravanti hmm interesting, i presumed going with 6.5q is 'fine' because of negligible perplexity
losing capabilities with simple conversion to MLX wasnt on my bingo card