#lore and getting out of the trenches
viscid ether
to explain properly:
was rly goated before college. advanced programs, gifted kid whatever, got lucky and was admitted to the #1 hs in washington. i was never a motivated person it was just kinda easy in general
graduated and was targeting neuroscience not cs, ended up getting into a rly strong program in canada (think t10 global for my discipline)
was during peak of covid, visa did not get approved in time to attend and i had to go to a shit state school bc i declined everything else and only they would "unreject" me. literally a t500 school
felt like shit, was depressed asf, played video games 14 hours a day and coasted classes with just bare minimum to 4.0 but basically never left my room
transferred out to a general track at my current uni
was sort of leaning towards cs/math at this point. some really fucked shit happened to me while i was living in seattle and i had to move out for my safety. also caused me to drop out of 2 quarters, almost leave school entirely
moved back home with my parents kind of just halfheartedly pursuing cs while feeling terrible about myself when juxtaposing trajectory before college and trajectory now
mental health kept on degrading due to a slew of other problems (being trans and figuring out my identity, recovering from what happened to me in my last relationship, was also not going outsdie enough cus of all these reasons) until it got to such a low that genuinely i was like okay either i kill myself or i just lock in
so i just decided i was at such a terrible low and things couldnt possibly get worse so now "positive" action i took was wasted, everything was beneficial.
started building games fro fun because the only thing i had really LOVED in my life prior to this was competing in esports. kind of my bridge into CS and math, it was really fun. made a game studio with some friends it was a pretty shit game and studio but we were all just having fun. my life still sucked but sometimes when you have literally nothing positive going on its easy to attach yourself to ONE goal because you can tell yourself that as long as i do this one thing today, i did a good job.
in this game studio we all had diff roles and treated it like some real shit. trello boards, figma, meetings, etc.
wrote a publication (not a real one, more of a think piece) on dynamic movemetn systems in platforms, and the hidden techs (coyote time for ex) that mean the difference between intuitive movement and uninteractive movement, and how these two things aren't far from eachother
applied to so many internships too. first offer was at epic games working on unreal engine. it got rescinded for their cuts but i got it by sending an email to 13 different recruiters at epic games. they ripped me a new one for it but still interviewed me anyways, i got offered for 2 teams.
after that i kept on building more things that i really, actually wanted to build. i did like <20 leetcode because i simply didnt want to it was so cancer
worth to note that i had 0 expectations of getting an internship and i sort of just thought to myself i will probably not get a job and im gonna just try my best to and i guess ill jsut work at a clothing/grocery store 40h/w until i figure something else out
got lucky on process. btw i got my first netflix offer because of my game dev exp.
@simple abyss @terse copper
here u go 💀
simple abyss
twin apex
I was told they'd mark it as spam if I did
I have 6 meta recruiter emails I emailed one, someone else who emailed the same recruiter got response same day and I got ghosted
viscid ether
tbh in retrospect i dont think its a great idea to email so many. i used to think it was rly bad cus of what happened with epic but then i realized they probably wont care if they are interested in ur profile
thanks
it was terrible but now its good
im happy and feel empowered in my career for the first time ever
and i dont see that changing
🔥
thats why i yap on this server so much now
viscid ether
i dont rly need it the way i used to to succeed in recruiting ive figured it out
but now i just help ppl when i can
terse copper
not by choice tho 🥀
viscid ether
yah its nice
what kind of stuff would you rather be in?
i can probably help you with moving there. i moved from platform side stuff (netflix first internship) to hands on ml at meta
so lmk
yeah i saw ur li
pretty cool
interp
https://www.arena.education/curriculum
https://sparai.org/
https://www.aisafety.camp/
mlsys
I think if you're a very performance conscious person, and love to dive into the lowest level of things... and really squeeze perf out of your hardware... this is the domain for you.
In terms of what you can start with right now, I think the best thing to do is start working on open source projects you find interesting. Just try your best to acquire as much knowledge as you can right now. Then once you get some experience, apply for research positions or jobs.
Getting free compute is pretty easy through platforms like kaggle or google colab, but if you want the top tier stuff, sometimes labs will offer it to you if you're building something foundational.
Books + resources:
Read JAX scaling book for general ML system scaling.
Read PMPP and other blogs like simon's blog on optimizing GEMM.
Read the huggingface ultrascale handbook.
the most neglected skill is actually having a very strong foundation in systems programming + general comp arch / networking. Some people think they can vibe code it now... very wrong. You really need to dive deep into the hardware and understand how to make use of the platforms given to you to interact with the hardware. If you can't do this, you won't succeed.
spar and aisc are not open atm. but they will be
they are some of the most concrete opportunities and genuinely fucking valuable. its a beautiful thing the communit is doing. researchers from T1 / frontier labs will mentor talent through these programs. that said you genuinely wont succeed if you arent capable of research, dont have the fundamentals, or arent genuinely interested in the domain. do the arena fundamentals, all 4 chapters, first
@radiant lantern
nah not much
i played one game yday for the first time in 2 years or so
i miss it a lot
esports / coaching esports is one of the most intellectually stimulating things of all time
try and fail and see if you care enough to work towards not failing
research is hard
im not even good at it
but i want to be and i will be
so it's fine
every day is spent furthering myself to that goal
yes, a lot of novel stuff
if u wanna get into it i can help a bit
with advice
both. research is hard to come by and will usually not be something you can control in terms of what you are doing. you work with ppl as ur school, or independent entities. usually theyll have their own research agenda because the benefit they get out of your involvement is help towards their projects. some people will be kind enough to mentor you on whatever you want to work on in your domain, but honestly idt thats a good place to start.
big projects are good but more when you have an idea of what you wanna build.
so what i suggest is:
read papers in the space you are interested in
try and implement some of them, maybe make some small projects with what you learned (just one decent one should suffice for next step)
find a phd student at ur uni who works on this stuff or relatively similar
fucking beg to work with them. talk abotu the papers you read. talk about their papers. i ca help you with the emails if you ever get there
work with them, gain your own opinion on what you think of the space, what you think is worth working on, what you enjoy
look for more specialized research and build better, more impressive projects, now that you arent' a noob in the space
continue upwards
def the research
i love studying intelligence
i see the same thing in research as i do in esports
its such a weird concept but maybe @terse copper will follo
basically when u are in high lvl esports
u dont really have so many people to learn from or such a concrete path to winning
u just have to explore and figure it out
and the ppl best at that define the meta and win the tourneys
research feels the same, its like a game
ill love it even more 🤑
hmm i think both can be both. theyre both just domains in which you can get experimental or applied. interp has some applie exmaples. see martian, transluce.
mlsys is also more applied but has a lot of research examples. see thinking machines non determinism paper https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/ that answers a fairly fundamental research question in language modeling
i will say this paper also "aligns" with interp but its sort of weird to describe it like that because you can then describe MOST things as interp simply because they answer a question on the internals of a model
im gonna go make dinner u guys can ping or dm with anymore questions
and ill answer in like 10min
yeah i got u can u send me a dm with whatever ur question is
even if its just like "how do i get into research"
but ideally more specific