V @_VatsaDev_ Twitter profile

Pinned Tweet

V

15 days

My worked helped in this!!! never even thought about non-text data until a few months ago, and never did I think it would be a part of a model this good!

vik

@vikhyatk

15 days

Genuinely shocked at how much better moondream has gotten at understanding text... it could barely read anything back in March.

9

7

134

1

0

15

Last Seen Profiles

@akaoni_______

@27K6185336CSE

@jeyke_heyyy

@Veteriner16

@madisonwest39

@EduardococaGrch

@yngsnooze

@asitakhle

@YKS_White_Kings

@HunnyyyBayeeee

@yukiiii_222

@SsdjNj2BotkyGJX

@bigzedick

@fatalspo

@SKZOO_CAFE

@kkristensoria

@dgtz

@ShuttUpImTalkin

@breakfastnick

@DoodChillBro

@auroraeuproject

@aquko_chan

@preset_2m3

@Wulfeee

@Sps_8O

@edgarchucky1

@8oo33

@maitake_web

@Kaoru8284

@FWBSMTR

@PooRKOpP0VZYzM0

@ashlyngoetz

@Tantedesijanda

@HWThYKjwFe18738

@gee_jefe

@Kelly_BeanWms

V

@_VatsaDev_

2 months

@mcneilly_alex Unspoken MIT rizz moment confirmed

0

82

V

@_VatsaDev_

2 months

My dad absolutely KO'ed my programming as usual: python DSA -> nah boi he used pascal three.js -> nah he made a torus in c++ use nvim -> oh yeah I remember using that when I debugged errors in C arch btw -> oh you mean formatting linux at 3am and rawdogging the servers at 6 ...

4

0

68

V

@_VatsaDev_

3 months

ML grind day 3 - presented some RNA research work at prof group meet - kept going through nnfs - on jjk ch 190 Need to get into more projects

1

53

V

@_VatsaDev_

2 months

Looking for a Internship, contact me if you have a spot. open to general cs, ml or robotics roles. I've worked on projects at nous research and optimized product search for I also have research lab experience in bio AI

4

6

51

V

@_VatsaDev_

3 months

growth by getting all the highbie chads

1

0

30

V

@_VatsaDev_

21 days

If asked why bullish on @cohere :

3

0

24

V

@_VatsaDev_

1 month

@vikhyatk You've literally announced m87's existence please give the company a linkedin profile so my best internship isnt a default building 😭😭😭

1

0

22

V

@_VatsaDev_

3 months

I've Finally got the blog post up! take a look at some experiments on layers and layer width at thanks to @cto_junior @notyamisukehiro @twofifteenam for reviewing it over and making it much better than I could have alone.

2

3

20

V

@_VatsaDev_

21 days

Just turned 18 yesterday

7

0

17

V

@_VatsaDev_

1 month

Everybody else: Moondream is wayyy to good for its size how tf did they do this meanwhile the channels:

3

0

16

V

@_VatsaDev_

2 months

Finally out of Vacay mode, locked in again: skill maxxing day 12: - learning Windows x86 Assembly and Ocaml: (legit asm is a footgun)

3

0

16

V

@_VatsaDev_

14 days

Google deepmind screenAI best case (left) v. my stuff at moondreams worst case (right)

0

15

V

@_VatsaDev_

11 days

me looking at myself compared to boon, damn i'm dying in the lowbie trenches out here:

boon

@iamyourboon

11 days

anyway, curious what the payout is, finally eligible

3

0

14

2

0

15

V

@_VatsaDev_

20 days

speaking of this @iamyourboon you should call up google deepmind and see what they do

V

@_VatsaDev_

20 days

CMU+quant really hits different huh

1

0

10

1

0

15

V

@_VatsaDev_

1 month

I've hit 300, ama for my followers, my chronic followers: @Kshitijjkapoor @iliekcomputers @angkul07 @snowclipsed @wateriscoding

5

0

15

V

@_VatsaDev_

7 months

Hmm @Euclaise_ @4evaBehindSOTA @Teknium1 @NousResearch Lilith looks promising at scale, maybe useable for future training runs?

1

2

14

V

@_VatsaDev_

2 months

Jane street! Ocaml! -> nah mf we used fortran, that stuff was fun bit manipulation -> that stuff great in cobol I'm starting an AI research lab -> haha nice me and the bois used to research compilers at 2am too state school CS :( -> (NIT undergrad->IIT masters) I cannot winnn

1

0

14

V

@_VatsaDev_

1 month

Bro wants to work at moondream maybe??? 🥹🥹🥹

snow

@snowclipsed

1 month

went through the entire moondream repo today, finally got a plan down on how to go all about it in pic: struct for weights

8

2

69

1

0

13

V

@_VatsaDev_

14 days

Mid level in three months, the mans a programming beast

boon

@iamyourboon

14 days

first, i am a new grad (i'm 3 months out of college), i want to maximise learning and becoming a subject matter expert in something i enjoy, swe is more versatile, engaging, and something i am more personally passionate about

3

0

102

1

0

13

V

@_VatsaDev_

20 days

I have questions. 1 yr and L5? 320k tc is a downgrade after 1yr nahhh im so dead 😭

boon

@iamyourboon

20 days

one downside is i actually downgrade my pay and leave a life of trading behind 🫡 (which is better for wlb, worse for being unable to retire in ten years)

4

0

124

1

0

12

V

@_VatsaDev_

1 month

@_xjdr That has got to be stupidest take I've ever seen, guarantee this guy Thinks OAI=GPU google=TPU, and nothing else. If I train a transformer with the same settings on a t4 or an h100, its going to be 99% the same

1

0

13

V

@_VatsaDev_

2 months

Jeremy howard follows my lab wtf got to ship faster now

2

0

13

V

@_VatsaDev_

2 months

Skillmaxxing day 15: - mostly a planner day, reflections on the AIMO - converting pythonic OOP to Ocaml is kind of weird - is the most satisfying file name ever new nvim theme fits though, monokai to chadracula is nice

1

0

13

V

@_VatsaDev_

22 days

is there like less competition for these or something? Is this a target-uni only thing/special stuff they notice you by? @iamyourboon @anpaure I need to know 🥹

boon

@iamyourboon

1 month

also ppl be looking down on citadel fall/spring internships, it’s just 85% waterloo kids

0

22

2

0

12

V

@_VatsaDev_

3 months

ML grind day 2 - 40 problems solved at project euler - making my first neuron all over again - started reading jjk, im 100 chapters in

1

0

13

V

@_VatsaDev_

3 months

omg thank for the god raichu follow @twofifteenam

2

0

12

V

@_VatsaDev_

3 months

Finally have a follower/following ratio > 1, AMA my 3 viewers?

3

0

11

V

@_VatsaDev_

1 month

Bullish on Noam arch coming to Gemini @_xjdr thoughts?

TDM (e/λ)

@cto_junior

1 month

Suddenly bullish on Gemini Most likely they'll hire Ilya as well once Sam takes away the Figure AI bots keeping tabs on him

2

0

47

1

0

11

V

@_VatsaDev_

1 month

nah my employer got noted its over, his reps ruined get the allegations out there now

vik

@vikhyatk

1 month

gotta admit he has a point

261

745

20K

1

0

11

V

@_VatsaDev_

20 days

CMU+quant really hits different huh

boon

@iamyourboon

20 days

good morning, twitch called, got the job, do I be a twitch engineer??

77

1

1K

1

0

10

V

@_VatsaDev_

3 months

ML grind day 3 - more in NNFS, stuck building simple nn's for now, must accelerate - Did some computational biology research work - JJK ch 155!

1

0

10

V

@_VatsaDev_

18 days

22yrs at Amazon ... damn bro probably looks at the juniors like "I can smell the burnout in you"

Grant Slatton

@GrantSlatton

18 days

LinkedIn resume of the most talented engineer I ever worked with at AWS Confirms all my priors

65

508

18K

2

1

9

V

@_VatsaDev_

2 months

I'm safe, I'm not unpaid 🥹🥹🥹🥹🥹

vik

@vikhyatk

2 months

@adamcohenhillel unpaid interns are the backbone of the economy and I will not be taking further questions at this time

1

0

14

0

9

V

@_VatsaDev_

2 months

Introducing @lucky86877 , or Lucky Bamboo research labs. The names from lots of bamboo metaphors, like NNs being grown and a model picking responses is like one path of many in the bamboo maze. We are looking at model evals, lora, model interfaces, and maybe some mech interp ...

3

0

9

V

@_VatsaDev_

6 months

@stablequan

How to Convert Any Text Into a Graph of Concepts

A method to convert any text corpus into a Knowledge Graph using Mistral 7B.

towardsdatascience.com

0

1

9

V

@_VatsaDev_

17 days

The duality of man

0

9

V

@_VatsaDev_

2 months

Skillmaxxing day 11: - mostly on 1 ML experiment, blog post is now up: - more fiddling with nvim, the lower ram usage is nice.

1

0

9

V

@_VatsaDev_

3 months

@anpaure economists liked this tweet

0

9

V

@_VatsaDev_

11 days

guy in my lab running randomized hyperparameter search and its just kept improving for 3 wks now (improving at least 1k times in a row) is he the hakari of neural nets???

3

0

8

V

@_VatsaDev_

21 days

Legendary shitposting out here @cto_junior

1

0

8

V

@_VatsaDev_

18 days

@alexthegrreat Goddamm I have never seen such a blanket "sucks will never work" and then anyone who sounds like they know what they're doing straight up agree with you

1

0

8

V

@_VatsaDev_

3 months

@iamgingertrash @egrefen @ssi @nearcyan NFGD has some great ones? Figma/Stripe/11labs/wnb/suno/perplexity/pika?

0

8

V

@_VatsaDev_

3 months

@LatentPepe @teortaxesTex @AkshGarg03 @siddrrsh @OpenBMB bruh this wrong on so many levels. how does one flirt with llama-3v?

0

8

V

@_VatsaDev_

1 month

@Kshitijjkapoor @iliekcomputers @angkul07 @snowclipsed @wateriscoding My occasional banger group: @vikhyatk @arpitingle @dejavucoder @cto_junior @Aryvyo

1

0

7

V

@_VatsaDev_

2 months

The guy asking the question died when he realized stable people on sf are real

Telt 🍕

@twofifteenam

2 months

@Welt__Meister__ 💍

0

8

0

8

V

@_VatsaDev_

1 month

@Kshitijjkapoor @iliekcomputers @angkul07 @snowclipsed @wateriscoding @vikhyatk @arpitingle @dejavucoder @cto_junior @Aryvyo And ofc the only here for the best tweets: @twofifteenam @_xjdr @teortaxesTex

0

8

V

@_VatsaDev_

2 months

I was looking for pokemon red "source code" as an 8 yr old and found a github repo full of asm and c. 8 yr old me ran away, only to use python 2 days later I do ml now ig

GitHub

@github

2 months

How did you get into coding?

927

131

2K

0

8

V

@_VatsaDev_

1 month

MIT students realizing they have free riches waiting for them

mcneilly ☃️

@mcneilly_alex

1 month

holy moly

18

4

685

0

1

7

V

@_VatsaDev_

3 months

@yacineMTB I've been bullish on them after the command R releases, been really strong

0

7

V

@_VatsaDev_

3 months

skillmaxxing Day 5 - more NNFS+nvim, testing claude 3.5 - made a dataset of challenging math problems, around like 5k QA pairs of AMC->AIME->IMO level stuff, way too much webscraping - Up to date with JJK now, I see why all the stans hype this everything cooks peak

0

7

V

@_VatsaDev_

1 month

We mogging bois

vik

@vikhyatk

1 month

@dejavucoder

1

0

5

0

7

V

@_VatsaDev_

2 months

Average 1AM+claude+neovim

0

7

V

@_VatsaDev_

7 months

These are Lilith runs at bs(batch_size) 48, versus an AdamW run at bs 180, the Lilith step takes 70 ms vs adam taking 300 ms per step. 4x less memory for batches while 5x faster.

0

6

V

@_VatsaDev_

2 months

The influence grows

0

7

V

@_VatsaDev_

13 days

HRT seems to grab every non jane street

mcneilly ☃️

@mcneilly_alex

13 days

in other news...

8

5

131

1

0

7

V

@_VatsaDev_

1 month

It'd be hilarious if @vikhyatk wrote a "what the interns have wrought" for m87, buts its just me pretending to be worth 3 jane street interns

Yaron (Ron) Minsky

@yminsky

1 year

Once again, it's time for Jane Street's "what the interns have wrought" post.

4

18

149

0

7

V

@_VatsaDev_

3 months

This is a real question guys. @dejavucoder @cto_junior @justalexoki @untitled01ipynb you're the best shitpoasters I know please help me.

V

@_VatsaDev_

3 months

@LatentPepe @teortaxesTex @AkshGarg03 @siddrrsh @OpenBMB bruh this wrong on so many levels. how does one flirt with llama-3v?

0

8

4

0

7

V

@_VatsaDev_

8 months

@burkov No 16k and 32k are just as possible, just more compute hungry, which means people like Google/OAI do it. Now when you see 128k or 200K, thats prob. rope/alibi, etc.

0

7

V

@_VatsaDev_

2 months

personally I'm really worried over the cycle that is: -> dpo the models making you look as good as possible -> it actually sucks on you irl -> consumer frustration -> make the model realistic -> "why does the model make ugly now ?!" ->

TDM (e/λ)

@cto_junior

2 months

Shopping at Zara online is going to be so much fun in a few years

20

80

1K

1

0

7

V

@_VatsaDev_

1 month

chaddream not moondream

vik

@vikhyatk

1 month

i have trained the same model 300 times so far this year

19

1

145

0

6

V

@_VatsaDev_

1 month

ML bros and Algo bros need to learn from minecraft bros @snowclipsed @iamyourboon probably get this

YUNG

@yungnickyoung

1 month

first day implementing some of the pieces for the Better Mineshafts redesign. Got a basic version of the surface entrance + main shaft working. Integration with surroundings in the main shaft is already looking wayyy better imo. Thoughts?

36

94

1K

2

0

7

V

@_VatsaDev_

15 days

@giffmana creating the real startup reef here, from doubtsuu-VLM to @moondreamai

0

1

7

V

@_VatsaDev_

3 months

@lelouchdaily @wateriscoding @gi0nyx @angkul07 I'm finally starting my own daily progress log too. Trying to get cracked by end of the summer and college Day 1 - Started reading through/working on NNFS ch 1 - Started on the Euler project, want to finish by the end of summer

4

0

6

V

@_VatsaDev_

5 days

imagine the waterloo intern at citadel on his way to invent the worlds fastest rejection setup, parses resume in 1 nanosec, rejcts you the other

anpaure

@anpaure

5 days

job application rejection world record

74

150

4K

2

1

6

V

@_VatsaDev_

6 months

Got transformers, nothing bigger than 20m parameters, doing math

0

2

6

V

@_VatsaDev_

3 months

Skillmaxxing Day 8: - the blog post is out! - Almost done with the paper for the neurips highschool track - AIMO looks like i get a bronze medal ig? (23) - Still working on the sass idea, describe it when its more fleshed out

0

6

V

@_VatsaDev_

1 month

If I have to measure up with this guy im: freshman, noname uni, 2x ML startup internships, 1x ML+product+search internship, offer from stanford startup I'm not "him", but I'm getting there

anpaure

@anpaure

1 month

he is why you don't have a job

65

106

6K

0

6

V

@_VatsaDev_

12 days

@_xjdr look id just like to know what 100 mil ctx effectiveness even is here, theres no evals for anything that big, and I cant even think of what 100m ctx len data even looks like.

3

0

6

V

@_VatsaDev_

2 months

great things happening out here

vik

@vikhyatk

2 months

> train the same model 3 times with slightly different data mixes > average the checkpoints > better performance than any individual checkpoint what is this black magic?

97

28

1K

0

6

V

@_VatsaDev_

3 months

googling @_sholtodouglas just brought up the other dude this was unexpected:

1

0

6

V

@_VatsaDev_

2 months

Everyday I yearn to learn Jax and fill out the TPU RC g-form and use the so-called legendary stability in these TPUs

xjdr

@_xjdr

2 months

i had to venture back in to GPU cluster training again these past few weeks. I figured "Lets do what the people do, pytorch, triton, FSDP, Ray" ... on the rare occasion that all my accelerators were available for training, it worked ok, i guess? HOW DO Y'ALL LIVE LIKE THIS?!

5

0

83

0

6

V

@_VatsaDev_

20 days

@agihippo google engineers otw to shard their kid across 1024 TPU slices and pool him back together again (child speedruner)

0

2

6

V

@_VatsaDev_

3 months

@quantymacro damn @quantymacro if a quants TC cant afford this...

1

0

6

V

@_VatsaDev_

16 days

@mcneilly_alex another mit student caught by the quant bait

0

6

V

@_VatsaDev_

3 months

I'M NOT WORKING ENOUGH

snow

@snowclipsed

3 months

I'M NOT WORKING ENOUGH

1

0

11

0

6

V

@_VatsaDev_

2 months

skillmaxxing day 14: - More Ocaml, check out: - Reflecting and regearing around AIMO, hope to improve on our score from here on with our own thing.

dungeon.ml

GitHub Gist: instantly share code, notes, and snippets.

gist.github.com

0

6

V

@_VatsaDev_

2 months

Malloc fails here

nahr (yapping)

@nahrzf

2 months

building ur mom in c from scratch

34

89

1K

0

6

V

@_VatsaDev_

5 days

bros pfp change was so big I thought a new anon happened

boon

@iamyourboon

5 days

new killua pfp vibe

31

1

263

0

6

V

@_VatsaDev_

22 days

Literally this but my dms are full of images with all 8 transformer paper authors and the rest have no rizz while aidan gomez has rizz

Ivan Zhang

@1vnzh

22 days

If Aidan Gomez has a million fans, then I am one of them. If Aidan Gomez has ten fans, then I am one of them. If Aidan Gomez has only one fan then that is me. If Aidan Gomez has no fans, then that means I am no longer on earth. If the world is against Aidan Gomez, then I am

6

4

103

1

0

6

V

@_VatsaDev_

5 days

wait damn hes actually former citadel @graffioh @Aryvyo @arpitingle this is peak alpha knowledge, hack your nearest google datacenter today, intercept that request, get that job people

binal

@binalkp91

5 days

@_VatsaDev_ Colocate your rejection server at Google’s data center to reduce latency

1

0

3

2

0

6

V

@_VatsaDev_

2 months

@dejavucoder Bro finishes projects instead of feature creep wtf

1

0

6

V

@_VatsaDev_

1 month

literally me rn

John

@jrysana

1 month

crazy how people will publish a research paper and its just blatantly false

15

5

103

2

0

6

V

@_VatsaDev_

1 month

This is what Non-MIT kids write on their cover letters to get the OAs

near

@nearcyan

1 month

when i worked at jane street i earned the nickname jane for my proficiency in predicting the street

4

2

187

0

6

V

@_VatsaDev_

3 months

@yacineMTB Yess yacine join the fellow desis man

0

6

V

@_VatsaDev_

2 months

My top website visitors are split between America and China (both countries need my ml alpha) but there's 31 ugandans? what are you doing?

0

5

V

@_VatsaDev_

2 months

Ocaml is kind of quirky, but nice gist with Intro code to like learn the very basics of Ocaml, getting use to syntax and stuff Also Ocaml needs to become the machine learning language, its literally got the mandate of heaven with the .ml extension

dungeon.ml

GitHub Gist: instantly share code, notes, and snippets.

gist.github.com

1

0

5

V

@_VatsaDev_

2 months

The avg @kaiokendev1 or @qtnx_ tweet is either peak esoteric experiment content or good theorizing. then half the replies are just "quantum blockchain tokenizer incoming" Its saddening

0

4

V

@_VatsaDev_

3 months

What if claude 3.5 is just claude 3 with all the logic/reasoning/math neuron activations boosted?

0

5

V

@_VatsaDev_

3 months

@mcneilly_alex Dude most of them are crypto shills. You can tell when the pfp is also just boredape/pepe/pixelboy with the .eth

0

5

V

@_VatsaDev_

1 month

This is what creator of llama file thought lmao

vik

@vikhyatk

1 month

gotta admit he has a point

261

745

20K

0

5

V

@_VatsaDev_

3 months

Thread thoughts: Pytorch visualizer - like penzai? whats it monitor? Pytorch to Manim - ooh thats epic if it works, though you might need paper for eqns Kino Compass - yess if only this existed the the main time was called the grandline so I could just have one piece tpot ...

snow

@snowclipsed

3 months

some ideas i am working on this summer

37

16

927

3

0

5

V

@_VatsaDev_

7 days

In Trading, Machine Learning Benchmarks Don’t Track What You Care About

Doing good work at HRT means constantly learning and improving. As part of that ongoing education, many people at HRT follow academic research – whether to keep up with the latest developments in...

www.hudsonrivertrading.com

0

5

V

@_VatsaDev_

18 days

There is no way the rich dude who can summon most waterloo with a tweet can complain about unfinished project ideas, 1 day spent articulating and posting them, and it's done

Nat Friedman

@natfriedman

19 days

I have too many project ideas

160

41

1K

0

5

V

@_VatsaDev_

2 months

Thundermoon release is out!!!!

vik

@vikhyatk

2 months

Released a new version of moondream today, with significant improvements in OCR and document understanding! You can actually run this model locally, unlike the other one that dropped today. 😅

24

46

453

0

5

V

@_VatsaDev_

2 months

The world if pytorch had q1/q2/q4 datatypes

vik

@vikhyatk

2 months

the world if ML libraries preserved backward compatibility over minor version updates

1

12

0

5

V

@_VatsaDev_

5 months

All right its official I need to build and run the pipelines, weekend lock in.

1

0

5

V

@_VatsaDev_

18 days

there's two wolves that you chose from, @yacineMTB with 2070s mem constrained, or @_xjdr where you start doing the scaling and end at "our run will fail every 8 minutes"

kache

@yacineMTB

18 days

@_xjdr once a shill always a shill (they do just work)

1

0

3

0

4

V

@_VatsaDev_

3 months

Late day 4 post, didn't get to do much - cont. on NNFS - started using neovim, my ram is thanking me after running vs bloat (I don't even know why its slow I have a new laptop and 5 extensions) - working on Mech Interp, have my own lab when the paper happens - jjk ch 200

0

4

V

@_VatsaDev_

3 months

@snowclipsed @teortaxesTex Ilya and the cult was the magic? Sama is money man - eliminated Murati management - gone GDB wonders - well they happened after the model was made, so also gone

0

5