llm_enjoyer

@tenderizzation @kimbochen u shud conduct passing of the baton by rt’ing all my shitposts ngl no cap

@kimbochen @LLMenjoyer im retired

this is exactly how it feels to ablate mamba 3

zoomer gooners have lowkirkuinely had a more concrete impact on ML sota than eg. Schmidhuber

scientists unveil new proof that MoE is faster than dense! ‼️‼️🤯

@Sanskarsam9 i see you should consider becoming a scammer either scam your company or scam some kind sponsors it’s actually p easy

@LLMenjoyer these are my personal projects, my company does work on applied ml and i don’t run my personal projects there.

latent moe, jianlin su balancer, mamba 3, attention residuals, 500 hparam changes... when do we cross the line from "yolo run" to "schizo run"?

@Sanskarsam9 why is ur compute not free

@LLMenjoyer spent half of my salary last month on compute

@LLMenjoyer idk i have a habit now to use this emoji and cat ai gifs 🤣

@EmpyriaMirai how it feels to ablate a cocktail of 5 peptides 10 vitamins and 69 hormones

@LLMenjoyer humans and vitamins.

@Sanskarsam9 why 🥀 that’s based asf???

@LLMenjoyer this is what ive been doing since last 3 weeks 🥀

@nayshins high dose muscle relaxant if that doesn't do it, try xanax

Can someone please make my eye quit twitching for the love of god

@danveloper bro but the loss goes crazy smh

I actually hate MoE's now. Not just because they're difficult to hardwaremaxx, but it's actually a really dumb architecture (no offense to anyone). They naively approximate a graph without any of the benefit of graph traversal. We're sending a blind person down a path and we've trained something to nudge them onto a different path to get to the end, but it doesn't know the next part of the map until the person has walked down the street. I hate this.

@teraote summonin the bangers be like

@LLMenjoyer i appreciate the consistent bangers

i hated it but the profiling seems to have been somewhat effective time to open a 100x leveraged short against the dataloader

@stochasticchasm will it be a high cortisol cudagraph or a low cortisol cudagraph

they're capturing my cudagraphs tomorrow

@YouJiacheng I see. I been looking for an opportunity to use this meme tho

@LLMenjoyer lol don't do that. the author told me they consulted Su whether they can not cite the blog in arxiv version. Su said "it's okay, but my recommendation is to cite my blog to avoid misunderstanding" (I use quote but not the original words)

Video of me when they don’t seem to cite Jianlin Su

@valigo wait till u find out about slerp

I desperately need to learn more math. Recently I invented "lerp" from first principles. Would have saved some time if I knew its industry-standard name :/

extremely high chances of cooking

@teortaxesTex @ZeyuanAllenZhu because canon is a scam bro, it's the most slop ablation

It's been known for a while that Canon from @ZeyuanAllenZhu is a monstrously powerful augmentation to the Transformer recipe, and I'm lowkey seething that it's not had industry adoption yet. GLM switched to DSA in months. If you're doing new pretraining, why not test Canon too?