As promised: github.com/xjdr-alt/entro…

better not be using an old school sampler then ...

2 independent early verifications seem to suggest ...

wow, this is kind of wild to wake up to. Thank you to everyone who stared entropix! Update: There are so many things i want to try right now but i am limiting myself to working on adding more diverse prompts to the repo for testing and evals so we can move beyond vibes to

to double down on this, the specific original goal was to see what we could accomplish with a vanilla OSS model without touching the weights or the architecture at all. This is a series of inference time compute experiments that essentially use the model outputs as as read only

While i'm waiting for some evals to finish, i decided to impl this idea, we have like 9 axis now, so the colors make less sense without a legend. We're also making some very good progress on the sampler still. Answers and process is getting better every day.

i test entropix on all sorts of prompts, i just post the 9.9 vs 9.11 one because its short. We are doing all sorts of crazy things with this 1B model

sometimes CoT is strange ...

ludwig 𝞍 Shin Megami Boson 𝞍 doomslide xjdr Teortaxes▶️ thank u sirs i have done nothing but run research paper evals, stare at samplers and sell robots to midwestern boomers for a month

TPU podslice optical routers are a modern marvel especially when you compare them to the alternative

hmmmm, i wonder what they did for the 100% ...

we are getting really close to something good here ...

the important part here other than the correct reasoning and answer is it put the '...' pause token in exactly the right place where it usually gets the answer wrong. I am starting to get more and more consistent results on logical reasoning problems now ...

sneak peak at the new SOTA reasoning eval im cooking

I'm not sure most of you are ready for this. The GIL is very useful and heavily relied on in many unexpected ways. I am personally looking forward to all the new and exciting race conditions we are about to find in core libraries

my 4090 died over night. I actually think i managed to work it to death. RIP king, you've served the cause with distinction and honor

in the mean time, what's everyone's favorite service to rent a big GPU dev box? Probably a few H100s? I'll probably try it for a few days and then succumb to what i know is right and just going back to doing primary development work on TPUs, but this might be fun

independent entropix result confirmation on the latest and greatest update, this time with torch! i think we may have made something good

This is a fantastic answer