XOutlier
Download .zipAdd to Chrome

François Fleuret

@francoisfleuret55k followers

Research Scientist @meta (FAIR), Prof. @Unige_en, co-founder @neural_concept_. I like reality.

Open on X
Top outlier
191×
Best organic post
Peak views
3.9M
Single post
Bookmarks
46k
Across 92 organic posts
Last scan
Oct 16
Community cache
Posts 92 posts
#DatePost↓ OutlierViewsLikesRepostsBookmarks
1Dec 2, '23Mathematics will fall first. Then, a torrent of results will impact everything in theoretical science.191×3.9M4.5k3471.2k
2Mar 20, '242h of debugging. Whatever you say, that's counter intuitive.42.9×873k2.9k138619
3Jun 19, '24AGI happens in 3y, where should I invest my money?30×612k98636390
4Jul 19, '24This is the kind of stuff I do not understand. Is it the consequence of some purely financial cost-benefits analysis that totally ignores the "trust building" dimension because it'29.6×602k1.2k4976
5Dec 19, '23Those who say they lack the proper math background for AI, what are the pain points?29×590k84934770
6Apr 21, '23"The Little Book of Deep Learning" Consider this as a beta version rough on the edges. Comments are welcome. @unige_en @sciences_UNIGE28.9×588k2.9k5411.9k
7Oct 10, '23My deep learning course @unige_en is available on-line. 1000+ slides, ~20h of screen-casts. Full of examples in @PyTorch. And my "Little Book of Deep Learning" is available as a ph28.6×583k3.6k7084.1k
8May 11, '24Come on! Trying to bullshit people by making math scarier with symbols they are unfamiliar with. That's bad @balajis . Let me rephrase: How many people attacking him understand tha27×549k92020103
9May 24, '24I encourage everyone to pause for a second and realize how far in the future we are.26.3×535k3k1791k
10Mar 20, '25Hot take: Anybody should have the possibility to knock at a University door saying "I did this cool thing working at home the last three years, can you give me a phd title" Three e24.4×497k5.8k228726
11Sep 7, '24BTW, when something is bad in a course, and you have reasonable evidences that it's not "just you" and that the lecturer/prof is not a totally bad person, please let them know. You23.7×482k1.4k3799
12Jun 30, '24What I find slightly depressing is that none of what is written here is *complicated* and many 7yo could grasp this. There is a strange cultural systemic issue that leads us as a s22.8×463k2.2k152512
13Sep 7, '24Once I was teaching C++ to undergrads. I got this vibe that the full room has "lost connection". Like ~200 students. I stopped and asked "okay, who is totally lost". ~190 raised th22.7×461k4.3k144609
14Mar 30, '25This brilliant analysis begs the question: why not 310% tariff? We could make $10T. According to my own research, a mere 1e21% tariff would provide enough money to build a Dyson sp22.2×451k12k492332
15Oct 1, '25"the human brain doesn't need tons of training data to do stuff"22.2×451k3.2k3371k
16Feb 12, '23I copy-pasted into ChatGPT two paragraphs I wrote very carefully that I was happy with, asked it to "rewrite them in good English", and the result was clearly improved. That's depr20.6×419k1.2k41142
17Sep 30, '24I realize I have absolutely no clue how you train a single model with e.g. 10k h100s. What is where, updated when, with what?18.5×376k1.5k441.1k
18Nov 14, '24I'm a bit confused by the "scaling is over" thing. Nobody really believed we'll get the full-fledged reasoning and creative AI god machine by training a ginormous 2019 GPT on 100x 17.5×357k1.7k65295
19Aug 12, '23I absolutely do not understand how "predicting the next word" implies any bound on the cognitive capabilities of a system. Quite the opposite since, if it was indeed able to do it,17.3×351k64933110
20Feb 19, '25Serious take: how comes there is such a dominant silence from the humanities on what to expect from / how to shape a society with AIs everywhere.17.2×351k2.1k92376
21Nov 16, '24Linux, Python, Pytorch, Slurm, tex, latex, tikz, bash, zsh, ssh, screen, tmux, emacs, vi, gcc, gnuplot, matplotlib, scikit, firefox. Who am I missing.16.7×339k1.9k72894
22Dec 30, '24Serious question: Von Neumann is always described as an absolute genius. What did he do that would put him anywhere close to Turing, Godel, or anyone on the Solvay conference pic?15.8×322k65418403
23Nov 17, '23I cannot be blasé about this. No GPU. 5 lines in an xterm and a fully convincing LLMs up and running. The /.../ indicate some output removed for clarity.15.6×318k1k75962
24Dec 29, '23This speech is impressive.15.1×307k63822174
25Feb 1, '25Someone TL;DR why @NVIDIA GPU memory grows so slowly? The V100 released 7 years ago had 32gb, and we currently have 80gb on the H100s... Why?13.5×275k80328201
26May 16, '23So FlashAttention is used now *everywhere*, and the paper is cited only 50 times? Not very grateful, are we?13×264k80469260
27Apr 24, '24Okay , here we go!13×264k1.4k1172.2k
28Nov 23, '23As a researcher, I have two kinds of ideas: the bad ones, and those that other people had 5y ago.12.4×252k2.7k21294
29Aug 7, '25Ok, no leap.12.1×247k1.4k3371
30Jan 24, '25This being said, here is the TL;DR: On the model architecture side, @deepseek_ai v3/r1 is a standard GPT that is a "causal decoder only", hence an auto-regressive models made of ca10.5×214k83467768
31Sep 5, '23I tried for 1h, with Google, stackoverflow etc. Finally asked gpt4. It's insane.10.5×213k83533181
32Aug 20, '22Reminder that my deep learning course @unige_en is entirely available on-line. 1000+ slides, ~20h of screen-casts. Full of examples in @PyTorch10.3×03.5k7072.6k
33Jul 23, '25High quality meme.10.1×206k8.9k735331
34Sep 14, '24Let me repeat: Math will fall first! N.B. This does not mean that math will not work anymore, or that AIs evade math or whatever, it means that humans will be beaten there, and tha10×204k63038228
35Apr 26, '25What do you think are the most important improvements of the 2017 transformer architecture?9.7×197k58525492
36Jan 18, '24Information Theory is awesome so here is a TL;DR about Shanon's entropy. This field is about quantifying the amount "of information" contained in a signal and how much can be trans9.0×183k9331171.1k
37Aug 7, '25Guys, I understand you like drama, but this is a remark about the AI development at large. We are seeing the plateau: just scaling up is coming to an end. For EVERYONE, not one com8.7×178k1.2k48142
38Nov 3, '25Rarely a sentence combined so much pretentious intellectualism and abyssal stupidity at the same time.8.7×178k2.9k149359
39Mar 26, '25So it seems that "real CS" people got quite a huge result: anything that can be done in O(f(n)) compute can be done in O(sqrt(f(n))) memory. Wow.8.5×172k2.2k1911.5k
40Jun 7, '24Here comes σ-GPT! TL;DR: We added an output positional encoding, which allows to generate tokens in any order chosen on the fly. Tweet 5/6 is sweet.8.3×170k84474453
41Dec 25, '24You people have to understand that I come from an era when we were using PCA and linear regression to make the AI.8.0×163k1.6k65185
42Apr 23, '24Conquering the world, one reader at a time.7.6×155k1.3k58900
43Sep 22, '24C++ is a terrible and abusive programming language devoid of any trace of elegance. All those defending it are unpassionate programmers with Stockholm syndrome.7.5×153k5492070
44Dec 31, '23Wow, inline C++ extensions for @pytorch are that simple? I am extremely impressed. From7.4×151k1.2k112585
45Nov 18, '24How comes people are not hysterical about this?7.3×149k1.2k821k
46Jan 13, '251PFlop FP4 and 128Gb SVRAM for $3k? Why isn't everybody totally hysterical about that cutie?7.2×146k58333149
47Jan 17, '25BTW, until late in life I had mediocre grades in maths, physics, and writing.7.1×145k1.4k41293
48Jun 10, '23If you feel a bit cocky regarding your talent for figures and clear thinking, here is a nice baseline to recalibrate.7.1×145k92899759
49May 21, '23In a month there have been 235k downloads of "The Little Book of Deep Learning", from 77k unique IPs! After some needed proof-reading and polishing, it is now a preprint.6.7×135k594113403
50May 1, '25Hot take: GPTs do not think before they speak, and this is currently addressed by making them speak even more, which is misguided.6.4×130k82222109
51Dec 15, '2436 months max.6.1×125k68724172
52Jun 22, '23Nature, go home, you are drunk.6.1×125k6359387
53Oct 12, '24This thing got only 240 likes. You realize that a deep model *learned the 3d, the dynamic, and the rendering of counterstrike* from recordings of games alone???5.9×121k1.3k60306
54Nov 18, '24I understand you guys like lists. - MLP - LeNet - LSTM - AlexNet - ResNet - Unet - SDD - VAE - AlphaGo - Transformer - GPT - CLIP - ViT What am I missing?5.9×120k84243703
55Apr 17, '24The Figures from Fukushima 1980 are really something.5.9×119k68559257
56Sep 8, '25Very pretty book!5.7×116k2k1451.4k
57Feb 24, '25These days are a neverending succession of amazing things. How is that one even possible.5.5×113k1.2k1111.2k
58May 25, '23If you want a physical copy of "The Little Book of Deep Learning", you can get the current preprint for 8€ Some pictures from an earlier prototype. Layout should be slightly better5.3×108k801103374
59Apr 27, '25As expected, that was popular. Here is my attempt at consolidating all the answers into a list. - Prenorm: normalization in the residual blocks before the attention operation and t5.3×108k686641k
60Mar 27, '25TL;DR Scaling laws are very very important (figs from the llama3 paper) Imagine you want to train a $100M model over several months. How do you pick model size, nb of tokens, learn5.0×102k87366755
61Feb 14, '2523948h of debugging later. Kill me.5.0×102k1.3k34133
62Feb 2, '23What have we done?4.9×99k6104674
63Oct 5, '24Serious: this AI thing truly is going to be a defining moment of humankind. Being part of it in any way is exhilarating.4.8×97k9278196
64Feb 5, '24We do not have time to waste! My target are TikTok addicts.4.7×96k58127321
65Jan 28, '25"OMG a model trained on nvidia and running on nvidia is awesome!!! nvidia is like, finished"4.5×91k1.4k7262
66Sep 24, '24Just played with @openai voice interface. It makes such a extraordinary teaching assistant. This product and the similar ones that will follow are simply going to make humankind mo4.4×89k74740128
67Nov 23, '24What is the German word for the guilt of under-utilizing the GPUs you are monopolizing?4.2×86k1k1970
68Feb 28, '25Volodymyr Zelenskyy will remain in history forever as an exceptional leader. Vance will be forgotten as a garden variety classless politician. Slava Ukraini.4.0×81k1.9k7439
69Apr 5, '25We are all Lee Sedol.3.8×77k99893135
70Nov 1, '25View post on X3.8×77k97946848
71Apr 3, '25"Listen we are actually using a powerful AI model"3.8×77k1.6k8693
72Dec 9, '23For tomorrow's flight. The faith I have in myself is astounding.3.5×72k59821161
73Oct 6, '24We are missing a word for the special form of sociopathy that relies on ignoring the tragedy of the commons. That guy feeling genuinely astute, proud and happy to have the idea of 3.5×72k7043977
74Nov 1, '24Some career update! I am joining @meta (FAIR) to work in @syhw's team, I remain Prof part-time at @unige_en, and my affiliation with @epfl comes to an end. 🎶 We are going to make 3.3×68k1.2k1771
75Nov 21, '24There is something surreal that [if we forget the heavy machinery for parallel training] a sota llm is a few hundreds lines of python.3.3×68k65820119
76Oct 28, '25Whatever this paper is about, this is a gorgeous figure.3.2×66k93725216
77Sep 9, '24Remember! The phone-formatted pdf is entirely free! 643k downloads! Your competitors are reading it at this very moment!3.2×65k69073537
78Dec 29, '24Because learning to solve math and cs problems is not about learning to solve problems but about giving you the Inner Eye and make you see a universe of transcendent beauty and har3.0×61k63830136
79Jul 28, '24The four stages of the "Little Book of Deep Learning"3.0×61k57024209
80Jul 26, '25The AI field is now split into (A) a "traditional" ml/dl domain, and (B) a "psycho-AI" domain where innovation requires an understanding of / intuition about the cognitive capabili2.9×58k69748252
81May 14, '25Two things teach intellectual humility: people smarter than you, and maths. Doing math with people smarter than you is sort of a bit too much.2.9×58k1.8k153211
82Jul 9, '24Me looking at documents with tikz and matplotlib figures more tasteful than mine.2.7×56k1.1k7381
83Dec 26, '24The DeepSeekv3 paper, beside being full of great new ideas, reveals that we can "easily" gain a xN. The release of the model and paper just contracted the timeline we are facing.2.7×55k78081371
84Dec 12, '18My deep-learning course materials are now updated to PyTorch 1.0.0. Close to 900 slides! Comments are welcome. #deeplearning #idiap #epfl #pytorch2.6×01.3k375243
87Feb 14, '25View post on X2.6×52k7683640
88Sep 30, '22A recent exchange on Twitter made me realize that many may be using tensors without knowing how they are represented in memory, and how it makes some operations super fast. This is2.3×0910132541
90Dec 27, '24Let's try this: I'll put some of the code for the slides of my course open-source. We'll see where it goes. I was told you people like attention mechanisms.2.2×44k58277584
92Jun 22, '22(in case you never realized)1.9×0882119352
95May 26, '22Don't want to boast, but ain't it a beauty?1.5×094267164
96Sep 26, '22If @PyTorch was printing the sizes, dtype and devices of all the tensors involved in an operation that failed, we would be getting AGI ten years earlier.1.4×01.2k7932
97Aug 14, '22My toy diffusion model does what it should do!1.2×060655235
98Jun 12, '22Whatever you think, it is insanely cool to be researching A.I. at this moment of history. The closest is probably physics at the dawn of 20th century: progress at the human-life ti0.9×06805630

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