Lazarz
@Laz4rz
Followers
2,829
Following
1,059
Media
793
Statuses
5,015
„hmm seems dope” based life decisions | epfl
Joined September 2017
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Se o Twitter
• 419151 Tweets
#CrisMorenaDay
• 320716 Tweets
#MaybellineTeddyLand
• 239476 Tweets
Peter
• 206904 Tweets
台風の影響
• 176146 Tweets
#jjk268
• 175746 Tweets
Alexandre de Moraes
• 131112 Tweets
Sukuna
• 87891 Tweets
Megumi
• 80065 Tweets
Lali
• 72274 Tweets
Yuji
• 64352 Tweets
デコピン
• 61638 Tweets
Yuta
• 59176 Tweets
De La Cruz
• 46613 Tweets
Bahia
• 42076 Tweets
焼肉の日
• 41394 Tweets
Luciano
• 34996 Tweets
TEDDYTINT X LOOKKAEW
• 34615 Tweets
Xandão
• 32981 Tweets
安全第一
• 29843 Tweets
TEDDYTINT X FAYE
• 28683 Tweets
Bluesky
• 23694 Tweets
臨時休業
• 18684 Tweets
मेजर ध्यानचंद
• 18568 Tweets
राष्ट्रीय खेल
• 18470 Tweets
#dominATE_BANGKOK
• 16660 Tweets
#NationalSportsDay
• 16655 Tweets
Wesley
• 15992 Tweets
Kaio Jorge
• 14348 Tweets
Uraume
• 14264 Tweets
伯仲燦然
• 13014 Tweets
Bruno Henrique
• 10642 Tweets
Hakari
• 10289 Tweets
Pinned Tweet
GPT-2 is done
It took a LOT, but pretty much every little feature and my thought process are documented, and I learned a ton.
It only waits for a full pretraining -- end of week I promise.
All credit to
@karpathy
, I only added comments making it 2head accesible.
21
34
864
@skalskip92
I love how in Poland we have those spontaneous spurts of good technology use just because some intern here is cracked and pushed it
4
7
1K
Diffusion Models goldmine for anyone wanted to get in, since I saw multiple people interested. These build math, coding, and intuition well. Let the math not scare you, its cool shit:
7
135
724
@ZFellows_
This is such a dangerous lie, but also an important life guidance..
Always struggle to be the best you can, put in the hours, sweat and tears. But also know, that it will not be fair, and there will be people putting way less hours, way better than you.
25
17
650
If you ever wanted a good intro to both low level language and linear algebra, then this is a phenomenal project
Yeah you can do it in a weekend, and it’s just got a brand new update (after 3.5 years!)
Ray Tracing in One Weekend v4.0.0 release is finally out! Three and a half years in the making — phew! Tons of fixes, changes, and new material.
8
127
728
2
38
498
Guy wrote MLX Transformers tutorial and has 6 followers, come on guys don’t sleep on alpha
7
30
455
leetcodes writeup turned into intro to leetcodes
Solved few basic problems assuming you know absolutely nothing. Implemented and explained quicksort, hash set, and linked list. Introduced time and space complexities with practical examples of how to calculate them.
Enjoy
7
10
421
if you never discovered the beauty of well written math (physics ftw) blogs, you truly missed out
they're somewhere in-between textbooks, and scientific papers and give you such a wonderful intuition on stuff
3
43
406
Internship hunt time
Looking for masters internship, finishing EPFL next year, my background:
- physics+computational neuro undergrad
- 3 yrs of job exp in quant, DS and ML Research
- 2 sems worth of projects on LLMs in EPFL Data Science Lab
- v extensive list of side projects
14
38
382
@oficerKRK
Fajnie, jakby ktoś rzeczywiście czytał i rozumiał cytowane artykuły.
"Remember that this model describes the best case: if all drivers drive exactly at the speed limit v, nobody crawling, nobody over-accelerating and then braking hard, and generally every driver staying very
15
10
350
@Sluzby_w_akcji
Bardzo lubię psy, ale wychodzenie z nimi bez kagańca+smyczy do ludzi to jest proszenie się o problem
10
1
317
I love all my Indian brothers here, but how in the love of Christ am I supposed to understand you when you switch to Hindi mid sentence?!
38
5
310
zig, vim and Germany losing — I’m going to top it with gym and call it a perfect evening
13
2
239
@Kucolog
Nie wiem, czy pisanie takich wątków wywodzi się ze złej woli, chęci zebrania engagementu, czy niewiedzy, ale:
Ten artykuł to toy model naukowca, który nie ma żadnego związku z modelowaniem ruchu. Co więcej, jak autor sam pisze: nie modeluje on ruchu miejskiego, tylko pokazuje
5
6
220
@yoimnotkesku
empty trains on long routes are one of peak life experiences, especially ones with compartments
I LIVE FOR THIS SHIT
1
1
214
as a side-side-side-project I wanted to learn how languages and compilers work in-depth, yep — another build from scratch endeavor
and I believe I found a perfect book
6
8
202
tried some terminals over the weekend
Alacritty -- can't ssh to tmux VMs by default, next
Wezterm -- dont remember was cumbersome
Warp -- GREAT initial feeling, garbage needs login + closed source
termius -- electron💀
im sticking with iterm2, thx for coming to my Ted
41
1
193
hahaha holy shit, cool-retro-term might be the best thing that happened to me recently, gonna use it every time I write zig to get in the mood
17
6
164
A SHIT TON of documentation and comments in there:
1
11
144
@yoimnotkesku
i love both vim and vscode
but using windows for coding is only a good idea if you want to suffer, for anything else I’m a windows user
9
1
126
@mister_shroom
We’re literally users stuck in user space, we can’t access any kernel stuff from up here
3
4
124
@pmukherjee02
how many times did you rerun it? Leetcode runtimes are pretty random, so I wouldn’t really trust it without local env testing
4
0
117
@quantian1
Always funny to read those takes that schools don’t teach about unlimited upside, my dudes people leave school lacking on basic algebra and you expect them to tackle juggling with company management
1
2
118
@shriyanevatia
formally being 25 this year I refuse to acknowledge that and I’ll be doubling down on trying all the things still
1
8
107
joining the Zig fanclub, before I switch to CS344: Intro to Parallel Programming
5
0
105
me: why is finishing GPT-2 taking me so long?
also me if I have the slightest doubt about any letter in the code:
7
0
104
On one-example-learning problem in LLM finetuning, or why does my loss curve look so weird!?
A small thread so you can avoid my mistakes 🧵
1
9
108
jesus
@vrushankdes
has 100 followers and wrote a goldmine on bottom up GPU -> CUDA -> PyTorch kernels and graphs implementations
3
8
72
ziglings writeup is starting to get biblical in length, ~1100 lines and we're just half way there
5
1
70
@onxydeux
@perrikaryal
In principle:
1. You record a lot of EEG, and synchronized game controls
2. You either train a neural net or build a heuristic that differentiates between different types of recorded brainwave events and map them to control inputs
3. You test and tweak it
4. Concentrate and play
2
3
69
@Lazloloszka
Funny story about Polish people and stealing in Britain:
One of my relatives (Polish) went to school in Britain, after some time he got a new phone and was so stoked about it, the very next day he got home with no phone, but with his nose broken, cause his British friends also
0
0
70
@ludwigABAP
“X86 needs to die” on
@ThePrimeagen
channel, too bad Casey’s enthusiasm got tamed a little, cause he was literally pouring gold on me
2
3
70
Re-did NN Zero to Hero over the weekend and distilled the most important parts.
Wrote about all of it, and certain other parts of NLP that I thought I should revisit.
Preparation continues, so today and tomorrow we tackle distillation and embedding.
3
1
69
@iliekcomputers
Tell me you never went dopamine mode in cafe without telling me you never went dopamine mode in cafe
2
1
64
Esoteric model architecture vs your dawg python code
No joke, I thought there are skip connections or residual streams looking at the diagram — there are none. Will I ever get proficient in reading those or are these just straight up badly designed?
7
2
65
@crypt0x_0
Had top of the line dell from work, switched to mbp 16 m2 pro and I’d never ever go back to other laptops unless I get a strong head trauma
1
0
62
Steering the LLM personality, behavior in a measurable scenarios, and open-world Simulacra settings with a descriptive approach, a thread 🧵
1
4
63
> read a cool tweet
> lose connection
> get connection back
> feed automatically refreshes
> lose the cool tweet
me to my x the everything app
2
2
60
🎉 done with vanilla gpt2, optimizations time
2
2
58
49$/m
@_MathAcademy_
price tag is perfectly reasonable for worlds leading economies
tho I think it’s a huge miss to not apply purchasing power parity, 49$ was literally 1/5 of my 1st yr undergrad living budget, can’t see any <high income student in CEE having money for that
8
0
60
@VictorTaelin
Could not agree more, 4o is often laughably bad — it produces better sounding, but as wrong as 3.5 or 3 answers, and by some miracle it’s apparently best in lmsys
2
0
57
Juggling full time + university + surgery + internship applications + uni masters exchange + side projects sometimes got me thinking it may be a lil too much
6
0
61
@jakubwiech
Czy chociaż jedna osoba tutaj, włącznie z autorem, otworzyła i przeczytała ten paper?
9
1
52
@FPSthetics
Same with the first Modern Warfare, it’s like original game development became a forgotten art
4
1
51
ok did Christmas come early this year?!
I never thought I'd interact with so many people I look up to, let alone that they would be interested in what I have to say, thank you all so much 🫶
1
0
49
72h interview prep speed run over the weekend (probably longer), I’ll go full monk mode
5
0
48
> at the gym
> 4 big guys are doing legs
> not only squats, stabilization and plyometrics too
> I’m one of them
life is good
3
0
43
I may have started a blog? I guess?
In the premiere episode:
Tour on how to properly feed the Unsloth Trainer (and Huggingface as well) with a dataset and how to handle Weights and Biases there.
Cause I couldn't google it and I don't know why.
5
1
44
Interesting stuff I’ve been doing lately:
- quantitative steering of closed LLMs
- ablation studies on Polish datasets + creating evals
- built and trained GPT-2, English and Polish versions
- building Java and C interpreters and own interpreted language
1
1
43
@ubermenchh_
It’s like I read something and mid sentence realize words started to look random and I need a second to make sure I didn’t lose my reading skills due to sudden brain damage and someone just switched language
2
0
41
@arithmoquine
it’s literally a crime against humanity that all papers aren’t written like this
0
0
42
since the gpt project is finished, I have a few directions I think are worth following:
- mech. inter. on this tiny gpt (to get a feel)
- spam hackerrank (got to pass interviews)
- zig/c stuff implementation (get overall coding fluidity)
- CUDA (profile gpt and write kernels)
6
0
39
im sorry to all 4 people who forked the gpt-2 repo, sometimes I'm just a silly boy... 💅
2
0
39
@KwiatkowskiKuba
Dla niezorientowanych: atak DDOS przeprowadza się poprzez masowe łączenie z jakąś usługą, łatwo można się domyślić że kibice masowo chcący oglądać mecz dadzą podobny efekt.
Kiedy się tak dzieje? Wtedy, gdy infrastruktura nie jest przygotowana na ilość chcących oglądać.
0
2
40