Gil Rodrigues

conceptual brand exploration for Filen the wider identity is built around concealment (forms only surface when the user brings the light)

is this the correct moment to plug my qwen3.8-27b setup?

you need to watch this

one more video bro this video is it ai is no more bro trust me

using elpis in eukhe.com is such a magical experience. it works so well. all you need is one rtx 3090

hi roon

pretty happy with what i made here eukhe.com

eukhe will only get better as time goes on

thanks for putting the runs together in an article. measuring sampler settings is worth doing and useful for comparing scores across setups. my issue is the headline of your article, your comment and the comparison under it. the data you linked supports a far smaller claim than "Low Temperatures Can Make Your 4-5 bpw Model Perform Like A 2 Bit One". elpis is greedy-only by design: the server rejects any temperature other than 0. the reason is that "the draft can't change the output" can only be checked token by token against a deterministic target, which temp=0 gives. the lossless proof covers exactly that: the draft never changes the greedy output. it doesn't cover greedy versus sampling, and i have not claimed that it does. you write that the drop "appears to be driven mostly (but not completely) by loops." i agree, and loops are a failure a harness has to handle. in the harness i run elpis behind, a repetition guard stops a stream once its last 64 lines repeat in a cycle of at most 16 lines. it then retries with a note that names the repeated line. because of that note, the retry is a different request, so greedy decoding doesn't fall back into the same loop. your setup lets a loop run to the 100k limit and grades it as a failure, with no repetition guard or retry in place. that measures an unguarded endpoint, which no coding agent should run on, not greedy decoding inside a coding harness. the 2-bit comparison puts a 4.6 bpw greedy run with no repetition guard (1 run, 9 of 100 loops) next to a 2.3 bpw run sampled at T=1.0 (7 runs, 2 of 700 truncated). the +7 for effort comes from T=1.0 runs as is applied at T=0. with the loops removed, the greedy score is 76.9%, about 83% after your +7, against 77.7% for 2.3 bpw. the gap that's left is mostly loop not something the bits explain. also: are the "pinned 100 harbor tasks" in your 2.3 bpw file the same 100 problems as in this article? greedy decoding can cost reasoning quality beyond loops, and elpis inherits that cost. a useful test would be a greedy run with a repetition guard next to a T=0.6-1.0 run on the same engine and quant, reporting both raw accuracy and accuracy without loops. if the gap is above your detection threshold with loops removed, it is a real finding and i will add it to the limitations of elpis.

github contributions are halloween themed today!

fastest decode and prefill for 4-bit qwen3.8-27b at 262,144 native context and preserved intelligence on one 3090. it’s fully open source! credit: @mr_r0b0t github.com/gildrb/elpis