Nick Oak

@teodorio more Birkin sub agents

I told my girlfriend I was talking with my Hermes agent on Whatsapp. She got really happy and a bit confused and asked me since when I have a Hermes agent? I told her since this morning, at which point she excitedly asked me if I am planning on buying her something.

this sounds like not a lot but 400m primes / s give 7 min for full octant segment of annular ring at 850M with “width” of 8192 which makes 1M sweep to be only 14h at 4090 which makes not only the sqrt(40) moat possible - but also several consequent after it

prior setup as per my experiments* in this field prior fastest been ~150-400k primes / s - Tsuchimura 2005 paper on a whole cluster

CUDA optimizations and deep dive into registers, memory, cache and SMs have finally paid off prior fastest setup has been ~5M gaussian primes / s processed on 4090 with clean revamped approach I now have 400M gaussian primes per second on 4090! fells insane paper soon

Once I’ll finish my math research I will dive all of the way into the LLM research there is incredibly lucrative edge on “latent thinking” going on right now I have some fun ideas of the proper regulrizarion that will make it work

optimizing Cuda kernels on Jetson Orin Nano and then pushing them to 3090 for ~25 times performance bump feels like magic

random fact but in Brunei there is a smell of flowers everywhere

@robinebers returned here several days ago quite calm

Dubai here I come, just a few more days 😭 can't wait to finally step foot into my home again

CUDA is a beast that is, well, hard to tame

@Mayu_kosho a lot of Russians abroad using TG though and plenty of tech savvy inside the country using it as well

今ツイッターできてるロシア人ならTGも見れんじゃねと思ったけどロシアにいない人が大半やったりするんやろか

ロシア人の皆様にはぜひ頑張ってロシア語で書いたりしているTGチャンネルのほうを登録してほしさがあるが、ここに出すと日本人に垢バレする恐れがあってめちゃめちゃ迷っている

@kayintveen @enjojoyy true! and this is how you can make the big fella to think for the 10-15 min and come back with genuinely bright ideas

@buildoak_ @enjojoyy this is the thing. low-level stuff where you're negotiating cache locality and branch prediction with the compiler, that's where claude becomes a genuine thought partner not just an autocomplete. the tight loop matters

So I do currently ship 100% of code using agents And I do miss coding manually a lot now Do you guys ever code by hands just not to forget how to do it? For the good time? I think I would schedule around one hour every day to code only by hand before I would start using AI Also in order to have a better architecture image in mind

depends on the task you wanna do throwing xhigh at some vanilla stuff will likely lead to over engineering though using it for audit & hard stuff - yields the best results for C++/CUDA code review there is no better model than gpt 5.4 xhigh - opus 4.6 and gemini 3.1 pro preview all miss critical moments

Is it true that Codex GPT 5.4 with Extra High thinking effort is worse than with Medium/High thinking effort? If it’s true, that’d be very bad design. 😅

never have I ever thought that in order to computationally advance the problem at the seam of compex analasis, number theory and graph theory - the biggest enabler will the the information theory to do the fittest baddest byte packing

be ready to spike of 429 errors on opus 4.6 in 5 min limits are resetting

@dogsmellsgood is it the feature that allows almost native experience with freediving with no additional training ?

Holy fucking shit. Finally found the name for it. Im going to cry Ive never been able to explain this to anyone they never know what im talking about

as Friday limits are nearing and a lot of folks maxxed their weekly plans - opus 4.6 is finally usable again

shitposting in between claude thinks for 10 minutes about how to make the light deterministic index to go from 4 bytes to 3 freeing up 1 byte that decreases the downstream processing time by the ~10-20%

having FSM and canary over the 128 bytes CUDA warp friendly structure is fun

and doing it in a way that all optimizations will transfer to the final algorithm materialization - being CUDA kernel

there is nothing more pleasant that spending approx a week for a proper spec for computational math objective bringing it to life in the morning in C++ and speeding it up almost 30 times through the day