XOutlier
Download .zipAdd to Chrome

Soumith Chintala

@soumithchintala329k followers

Building new things @thinkymachines. Also dabble in robotics at NYU. Cofounded @PyTorch. AI is delicious when it is accessible and open-source.

Open on X
Top outlier
34.9×
Best organic post
Peak views
2.5M
Single post
Bookmarks
17k
Across 87 organic posts
Last scan
Oct 16
Community cache
Posts 87 posts
#DatePost↓ OutlierViewsLikesRepostsBookmarks
1Nov 6, '25Leaving Meta and PyTorch I'm stepping down from PyTorch and leaving Meta on November 17th. tl;dr: Didn't want to be doing PyTorch forever, seemed like the perfect time to transitio34.9×2.5M11k5582.2k
2Jul 30, '23No More GIL! the Python team has officially accepted the proposal. Congrats @colesbury on his multi-year brilliant effort to remove the GIL, and a heartfelt thanks to the Python St19.3×1.4M4.5k1k499
3Apr 27, '24apparently Google laid off their entire Python Foundations team, WTF! ( @SkyLi0n who is one of the pybind11 maintainers just informed me, asking what ways they can re-fund pybind1117.1×1.2M3.6k5011.4k
4Jan 18, '24Can finally talk some GPU numbers publicly 🙃 By the end of the year, Meta will have 600k H100-equivalent GPUs. Feel free to guess what's already deployed and being used 😉!9.6×695k1.8k171283
5Feb 15, '24Based on all the user-request videos that @sama's been posting, it looks like sora is powered by a Game Engine, and generates artifacts and parameters for the Game Engine. 🤔8.5×612k1.3k67283
6Feb 1, '24If you have questions about why Meta open-sources its AI, here's a clear answer in Meta's earnings call today from @finkd8.1×584k2.4k372724
7Nov 17, '23* In 2016, I thought OpenAI was just shady, with highly unrealistic statements * In 2020, I thought OpenAI was doing awesome work, but a bit too hypey, and the AGI bonds were weird7.1×514k1.1k70136
8Mar 5, '24reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins I spend everyday on catching up with what's going on in AI. So much alpha, organized hierarchicall7.0×503k78971777
10Nov 1, '23In 270 days, the Department of Commerce will determine whether they will allow open-weights or not. if you support open model weights and want something actionable to do, then figu6.5×469k668151127
11Dec 18, '23PyTorch's design origins, its connection to Lua, its intertwined deep connection to JAX, its symbiotic connection to Chainer The groundwork for PyTorch originally started in early 6.5×467k1.4k256947
12May 25, '23I'm fairly puzzled by $NVDA skyrocketing. GenAI inference and fine-tuning will significantly outweigh GenAI training in overall compute. When it comes to inference and fine-tuning,5.8×418k979130478
13Aug 24, '23CodeLlama -- a version of Llama2 that was fine-tuned for code tasks is live now. Available in 7B, 13B and 34B.5.7×409k901194226
14Jul 23, '24Why do 16k GPU jobs fail? The Llama3 paper has many cool details -- but notably, has a huge infrastructure section that covers how we parallelize, keep things reliable, etc. We hit5.3×381k1.3k201700
15Oct 15, '25Sometimes we forget that NVIDIA wins because it's a software company. DGX Spark is a reminder of that. It's a CUDA dev machine that's beautiful enough and small enough to be on my 4.9×358k1.1k69207
16Jan 19, '22It’s been 5 years since we launched @pytorch. It’s much bigger than we expected -- usage, contributors, funding. We’re blessed with success, but not perfect. A thread (mirrored at 4.6×02.3k264199
17Dec 6, '23Seems to solidly compete with GPT-4 on benchmarks. Google has existing customers and surfaces to start the feedback loop, without worrying about adoption. And Google will use TPUs 4.4×317k89780120
18Jan 25, '25i'm comically impressed that people are coping on deepseek by spewing bizarre conspiracy theories -- despite deepseek open-sourcing and writing some of the most detail oriented pap4.3×314k4k388469
19Mar 12, '24Here's details on Meta's 24k H100 Cluster Pods that we use for Llama3 training. * Network: two versions RoCEv2 or Infiniband. * Llama3 trains on RoCEv2 * Storage: NFS/FUSE based on4.3×311k1.4k187646
20Nov 20, '23to my researcher friends (at @openai and those who watched this saga unfold) -- if you want to focus on doing solid *open* and published research, have access to lots of hardware -4.2×305k64852153
21Jun 30, '23Here's @MosaicML showcasing their results with @PyTorch 2.0 + AMD for LLM Training. They made "Zero Code Changes" to run on AMD. MI250 is already trending well, and IMO MI300X will3.9×279k5097493
22May 20, '20This is a new Microsoft. - WSL CUDA/GPU Support - native `winget` package manager - VSCode, Edge, buying Github They are listening. They are admitting failure. They are marching on3.9×02.2k24534
23Dec 6, '19The first full paper on @pytorch after 3 years of development. It describes our goals, design principles, technical details uptil v0.4 Catch the poster at #NeurIPS2019 Authored by 3.8×01.6k404165
24Nov 4, '25I love how @FFmpeg is conducting themselves online. It's a project that's created billions of dollars in actual value and has captured practically none of it. Their stance seems to3.6×261k1.8k60131
25Oct 16, '21PyTorch co-author Sam Gross (@colesbury) has been working on removing the GIL from Python. Like...we can start using threads again instead of multiprocessing hacks! This was a mult3.6×01.5k298227
26Apr 10, '24Meta announces 2nd-gen inference chip MTIAv2. * 708TF/s Int8 / 353TF/s BF16 * 256MB SRAM, 128GB memory * 90W TDP. 24 chips per node, 3 nodes per rack. * standard PyTorch stack (Dyn3.5×255k1.1k125280
27Mar 15, '23People getting mad about @OpenAI not releasing GPT4's research details.... the only tangible way to get back is to surpass GPT-4's results and release the details of how it was don3.4×249k7796465
30May 17, '24holding your (already earned) stock compensation hostage over signing a non-disparagement clause is **not** normal. IMO its pretty sketchy and sleazy. Its sleazier when a super-maj3.3×239k87451149
31Oct 13, '16Linux kernel push confirms, Intel to add dedicated neural network hardware ops into their future processors3.1×07357251
32Jan 4, '24Perplexity has become my most used AI app by the end of 2023. I use it for fact-seeking questions -- including recent news / facts, summarizing opinions and recommendations on prod3.0×218k74850214
33Nov 20, '23this weekend has been very sad. My friends at @OpenAI swore that it had become a magical place, with the talent density, velocity, research focus and (yet) a product fit that is re3.0×215k1.1k5766
34Dec 20, '24i rabbit-holed into the Genesis Sim codebase because the website is hypey and unclear; and I didn't want to just blindly retweet. It's basically like Mujoco but with more advanced 2.9×213k53364230
35Mar 21, '23The race has publicly begun! - GPT still blows everyone else out of the water wrt coding abilities, with Claude coming a close second. - Bard can't do coding yet (according to thei2.8×203k59350123
36Apr 6, '22anyone else feel burned out by a new AI breakthrough every week? 🤯 Trying to keep up but it goes by so fast most of it is not easily or locally reproducible which adds to the stre2.8×01.6k9354
37Dec 16, '22can @huggingface quickly start a Twitter clone for AI/ML folks 🤪2.7×194k7023034
38Feb 14, '18Tensor Comprehensions: einstein-notation like language transpiles to CUDA, and autotuned via evolutionary search to maximize perf. Know nothing about GPU programming? Still write h2.6×01k41013
39Sep 16, '17NIPS Conference Registrations 2002 thru 2019. [2018] War erupts for tickets [2019] AI researchers discover time travel2.6×01.1k3713
40Jul 18, '23LLaMa-2 from @MetaAI is here! Open weights, free for research and commercial use. Pre-trained on 2T tokens. Fine-tuned too (unlike v1). 🔥🔥🔥 Lets gooo.... The paper lists the ama2.5×182k1.1k173135
41Jul 18, '21Deep Learning is not yet enough to be the singular solution to most real-world automation. You need significant prior-injection, post-processing and other engineering in addition. 2.4×01.2k15987
42Mar 19, '19Pretty excited about NVIDIA's Jetson Nano. 5W, $99 and a 128-core Maxwell sounds pretty great. This year is so good for embedded deep learning!2.4×01k26857
43Dec 21, '22Had a practical search today that Google totally failed to answer, failed too, but ChatGPT gave an answer that sounds plausible. Now, the conundrum is that I don't know whether Cha2.3×169k5403572
44Oct 16, '25MacStudio you ask? Apple Engineering's **actual** time spent on PyTorch support has't given me confidence that PyTorch Mac experience would get anywhere close to NVIDIA's any time 2.2×160k76555161
45Nov 30, '2310 crazy years -- pytorch, detectron, segment-anything, llama, faiss, wav2vec, biggraph, fasttext, the Cake below the cherry, and so much more. Can't say we didn't change AI and to2.2×156k7785493
46Aug 11, '21Maybe I can finally use matplotlib now without spending half a day googling the exact syntax and options!2.1×01.1k10990
47Jun 28, '23No, GPT-3 wasn't trained in 11 minutes. The GPT-3 architecture was trained on the C4 dataset to 2.69 log-probability in 11 minutes on 3584 H100 GPUs. Don't focus on the "11 minutes2.1×149k5214370
48Nov 8, '18Rethinking floating point for deep learning - Jeff Johnson at FAIR - proposes non-linear floating point math -- more energy efficient, accurate - no retraining or quantization befo2.1×083025474
49Feb 11, '202013 me would not have had that loud GPU desktop from craigslist if Colab was around. Colab Pro at $9.99 / month to get prioritized 24-hour stints of fast GPUs is a steal. It is a 1.8×01k8256
50Jun 26, '24My general rule of thumb for assessing hardware startups (is this flowchart) Hardware has a long lead-time, so startups hustle and bullshit and make over-the-top claims all the tim1.8×132k73769383
51Oct 4, '24yes, Meta released a full scientific paper on MovieGen, with a lot of details that'll help the field move forward.1.8×130k68256157
52Oct 3, '22We just released AITemplate -- a high-performance Inference Engine -- similar to TensorRT but open-source. It is really fast! On StableDiffusion, it is 2.5x faster than the XLA bas1.8×0692112173
53Sep 23, '2510GW is about $340B of nvidia h100 at $30k/gpu (assuming 20% of power for non-gpus). if openai got a 30% volume discount, they'd pay nvidia $230b probably. so instead, maybe openai1.7×125k46417139
54Oct 22, '25. @PyTorch has snowballed into something never imagined, while still keeping our core values intact! feels incredible. -with @ezyang and @apaszke1.7×123k7321854
55Sep 1, '22PyTorch M1 GPU support (still in alpha) already being put to good use:1.7×0722104144
56Apr 22, '24Llama3-70B has settled at #5. With 405B still to come next... I remember when GPT-4 released in March 2023, it looked like it was nearly-impossible to get to the same performance. 1.6×118k6564681
57Sep 26, '17NVIDIA releases an open-source Deep Learning Inference chip design (based on Xavier), with full verilog source:1.6×05362876
58Apr 18, '24Llama3 8B and 70B are out, with pretty exciting results! * The ~400B is still training but results already look promising. * Meta's own Chat interface is also live at * TorchTune i1.6×115k67989103
60Apr 8, '20ML Code Completeness Checklist: consistent and structured information in the README makes your code more popular and usable. Sensible advice, backed by data. Proposed by @paperswit1.5×0620143103
61Nov 14, '17use gradient magnitude as a signal for gradient importance. Sort your gradients, find a threshold, clip your gradients, exchange sparse gradients, win.1.5×06192271
62Aug 23, '18"Everybody Dance Now" from Caroline Chan, Alyosha Efros and team transfers dance moves from one subject to another The only way I'll ever dance well. Amazing work!!!1.5×059019321
63May 13, '24really exciting demos from @OpenAI establishing new expectations for AI. Lot of work to do on the Llama train -- which isn't going to stop until we catch up! My personal feeling wi1.5×106k7062861
65May 22, '20Debates on PyTorch vs TensorFlow were fun in 2017. There was healthy competition to innovate, and philosophical differences like Theano vs Torch, Emacs vs vim, or android vs iOS. N1.4×07669614
66Feb 25, '19cancel meetings on your calendar until it sparks joy!1.4×07131188
67Jun 7, '18googled for an error message in pytorch, read my own answer from a year ago. full circle! Then dug around some stats. There are 34,300 posts on the PyTorch forums, viewed 7.6 milli1.4×0875522
69Mar 2, '17FAIR releases faiss. Many uses: text2image by searching through 1B or 100B images? RL Agent with VERY LARGE memory?1.3×04852360
70Jul 10, '24Meta's PyTorch teams published their actual roadmap documents publicly for the first time: Our engineers' incentives and objectives are laid bare for folks to read 🙂 While all PyT1.3×95k818125374
71Mar 6, '24the PyTorch codebase is 27 million tokens.1.3×92k5922571
72Dec 5, '19Super excited to welcome the PFN team to the @PyTorch community. With Chainer, CuPy, Optuna, MNCore, their innovations need no introduction. The community is going to get even more1.3×062313913
73May 27, '20- Take FasterRCNN - Remove clunky NMS, Proposals, ROIAlign, Refinement and their gazillion hyperparameters - Replace with Transformer - Win! Simplifies code and improves performanc1.3×060510250
74May 25, '21Getting @satyanadella to talk about PyTorch...✅ . . (Satya and I went to the same high school, Hyderabad Public School)1.2×07284438
75Sep 12, '22Big announcement: PyTorch Foundation! PyTorch has large core investments from many companies. So, we're creating a neutral foundation for securing assets and interests. Technical G1.2×06139919
76Oct 2, '24"How to train a model on 10k H100 GPUs?" has now been immortalized on my blog:1.2×88k58376518
78Jul 22, '17Congrats to the DenseNets authors for winning the CVPR best paper award. Elegant work, well deserved!1.2×04891760
79Dec 4, '17We're launching a FAIR Residency Program: a 1yr fixed-term research training program where you will work closely with researchers at FAIR. Deadline for applications is January 26, 1.1×04901610
80Jul 16, '25considering Muon is so popular and validated at scale, we've just decided to welcome a PR for it in PyTorch core by default. If anyone wants to take a crack at it...1.1×82k83757246
82Dec 21, '18The speech team @ FAIR released wav2letter++: - a fully convolutional speech recognition system - a C++ ML library on top of ArrayFire They recognized very early that C++ was their1.1×047911151
84Nov 1, '21Two exciting news from our robotics research today. 1/ DIGIT: a vision-based touch-sensor Projects light into a gel in the "finger-tip". A camera + model rapidly estimates the chan1.1×04778351
85Aug 7, '18Dont miss out on internships. Get as many under your belt as you can before joining full-time! I always wish I could go intern at Deepmind, Brain or NVIDIA for the summer just to k1.1×0583697
86Feb 12, '18Cloud TPUs are out, we'll start sketching out @PyTorch integration. The cost is $6.50 per TPU-hour right now. Hopefully when they get affordable, we will be ready with PyTorch supp1.1×05411080
87Feb 3, '22This is not a research paper, this is a real-world product. Wow! Been following @runwayml from their early days (and visited their offices last year). Great set of people, strong c1.0×05285752
88Sep 15, '22Small perks of joining the Linux Foundation! We spoke about ML Accelerators and Linux driver-land issues :D1.0×05983837
89May 15, '20me watching @PyTorch be dragged into the AGI tweetstorm1.0×0662186
90Jul 20, '24I'm giving the opening Keynote at ICML 2024 on Tuesday the 23rd @ 9:30am CEST. I'll try empower folks to get Open Science back on track -- the free discussion of ideas is such an i1.0×72k6526883
91May 18, '22Everyone's been waiting for it! Thanks to Apple, working in closing collaboration with the core team for making this happen!1.0×05884421
92Oct 11, '23AI is just matrix multiplications. Brains are just biochemical interactions. Must be simple.0.9×68k4984945
94Sep 23, '19very excited about the @GoogleAI office in Bangalore!0.8×0520379
95Apr 2, '21Over the past year, I've been doing robotics at FAIR. It's been lots of fun. My personal research goal is to build home robots: cooking, cleaning, etc. (1/x)0.8×04732236
96Apr 26, '22Twitter, if you're listening -- 1. get rid of the crypto bots 2. a feed that does some kind of Tweet TF-IDF. If someone tweets once a month, I want to see that as visibly as someon0.8×0487257
97Oct 13, '24SpaceX continues to be incredible. The engineering feats are absolutely nuts. What I find more incredible is SpaceX as an organization -- can execute structured long-term research 0.7×51k6513337

Scan another creator

@
@karpathy@levelsio@paulg@naval@marclou@illyism

Run this on any account.

XOutlier builds this table for every public profile you open on x.com, with live data and every post, not just the top 50.

Add to Chrome Download .zip