# ![@marc_lelarge Avatar](https://lunarcrush.com/gi/w:26/cr:twitter::567539642.png) @marc_lelarge Marc Lelarge 🌻

Marc Lelarge 🌻 posts on X about code, attention, python, in the the most. They currently have [-----] followers and [--] posts still getting attention that total [---] engagements in the last [--] hours.

### Engagements: [---] [#](/creator/twitter::567539642/interactions)
![Engagements Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::567539642/c:line/m:interactions.svg)

- [--] Month [------] +4,825%
- [--] Months [------] +200,387%
- [--] Year [------] +40%

### Mentions: [--] [#](/creator/twitter::567539642/posts_active)
![Mentions Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::567539642/c:line/m:posts_active.svg)


### Followers: [-----] [#](/creator/twitter::567539642/followers)
![Followers Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::567539642/c:line/m:followers.svg)

- [--] Month [-----] +1.60%
- [--] Months [-----] +11%
- [--] Year [-----] +14%

### CreatorRank: [---------] [#](/creator/twitter::567539642/influencer_rank)
![CreatorRank Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::567539642/c:line/m:influencer_rank.svg)

### Social Influence

**Social category influence**
[technology brands](/list/technology-brands)  5% [celebrities](/list/celebrities)  2.5% [travel destinations](/list/travel-destinations)  2.5%

**Social topic influence**
[code](/topic/code) 12.5%, [attention](/topic/attention) 10%, [python](/topic/python) 7.5%, [in the](/topic/in-the) 7.5%, [bridge](/topic/bridge) 7.5%, [the first](/topic/the-first) 5%, [have the](/topic/have-the) 5%, [links](/topic/links) 5%, [make no](/topic/make-no) 2.5%, [close](/topic/close) 2.5%

**Top accounts mentioned or mentioned by**
[@sri26762339](/creator/undefined) [@sethhweidman](/creator/undefined) [@grumpyreviewer2](/creator/undefined) [@gabrielpeyre](/creator/undefined) [@haggaimaron](/creator/undefined) [@lipmanya](/creator/undefined) [@davidpicard](/creator/undefined) [@sama](/creator/undefined) [@alexundersky](/creator/undefined) [@ollama](/creator/undefined) [@noureddinsadawi](/creator/undefined) [@blogbinaire](/creator/undefined) [@googlecolab](/creator/undefined) [@pytorch](/creator/undefined) [@coqlang](/creator/undefined) [@ejgallego](/creator/undefined) [@poor__rican](/creator/undefined) [@varadrane17](/creator/undefined) [@katek_warren](/creator/undefined) [@k7agar](/creator/undefined)

**Top assets mentioned**
[Cipher (CPR)](/topic/cipher)
### Top Social Posts
Top posts by engagements in the last [--] hours

"New course on"  
[X Link](https://x.com/marc_lelarge/status/2021879148097286406)  2026-02-12T09:28Z [----] followers, 36.7K engagements


"Here is the code to learn Flash Attention: https://github.com/dataflowr/gpu_llm_flash-attention https://github.com/dataflowr/gpu_llm_flash-attention"  
[X Link](https://x.com/marc_lelarge/status/2021887041768763892)  2026-02-12T10:00Z [----] followers, [----] engagements


"Are you allowed to use Triton https://x.com/marc_lelarge/status/2021879148097286406s=20 welcome to the mistral interview process. implement flash attention from scratch. starting [--] minute timer now. make no mistakes https://t.co/gZTU7X22Bp https://x.com/marc_lelarge/status/2021879148097286406s=20 welcome to the mistral interview process. implement flash attention from scratch. starting [--] minute timer now. make no mistakes https://t.co/gZTU7X22Bp"  
[X Link](https://x.com/marc_lelarge/status/2022056631811387447)  2026-02-12T21:14Z [----] followers, [----] engagements


"@SethHWeidman It depends on your background. For me and my students Triton is easier (close to Python). I think knowing a bit of Triton allows them to understand how GPUs are working with a minimal effort. Closer to CUDA I did some Numba too: https://x.com/marc_lelarge/status/1999467482025767113s=20 Learn  from the ground up: begin with Numba for low-level control then progress to Triton to write high-performance kernels in a Python-like language. A hands-on Jupyter notebook to get you started quickly. https://t.co/VwpyHZSqAw https://x.com/marc_lelarge/status/1999467482025767113s=20 Learn "  
[X Link](https://x.com/marc_lelarge/status/2021987533471854926)  2026-02-12T16:39Z [----] followers, [----] engagements


"@Sri26762339 I did not try on RTX3060 but the first part in Triton is working on T4 on Colab so it will probably be OK. For the last part full Flash Attention in Triton it is probably easier on H100"  
[X Link](https://x.com/marc_lelarge/status/2022301601550008421)  2026-02-13T13:27Z [----] followers, [---] engagements


"#AdventOfCode - Day [--] by #ChatGPT 🤯"  
[X Link](https://x.com/marc_lelarge/status/1599100010226135045)  2022-12-03T17:55Z [----] followers, [--] engagements


"Reviewer 2: "In terms of underlying techniques ChatGPT is not particularly innovative (.) It's nothing revolutionary although that's the way it's perceived in the public. It's just that you know it's well put together it's nicely done." @GrumpyReviewer2"  
[X Link](https://x.com/marc_lelarge/status/1617795181822234626)  2023-01-24T08:03Z [----] followers, [----] engagements


"The : a fundamental result in matrix theory with applications to probability theory (ergodicity of Markov chains); to the theory of dynamical systems to economics demography and to Internet search engines (PageRank). #AlgoNetworks"  
[X Link](https://x.com/marc_lelarge/status/1681165901758857216)  2023-07-18T17:37Z [----] followers, 16.6K engagements


"Hyperlink-Induced Topic Search (HITS or hubs and authorities) is a link analysis algorithm that rates Web pages developed by Jon Kleinberg: the authority weight is computed using the power iteration method to compute the first eigenvector of the co-citation matrix. #AlgoNetworks"  
[X Link](https://x.com/marc_lelarge/status/1681890174215634945)  2023-07-20T05:01Z [----] followers, [----] engagements


"is a variable-length prefix coding algorithm used for lossless data compression where shorter codes are assigned to more frequent symbols ensuring optimal compression by minimizing the average code length. Mathematical proof👇 #InformationTheory"  
[X Link](https://x.com/marc_lelarge/status/1731553417879761368)  2023-12-04T05:58Z [----] followers, [----] engagements


"() can be used to save space. This is particularly useful when working on a cluster where your HOME is limited in size"  
[X Link](https://x.com/marc_lelarge/status/1731954059228135880)  2023-12-05T08:30Z [----] followers, [---] engagements


""The surprising effectiveness of untrained message-passing GNN. For untrained MPNNs we (.) only optimize for the output layer(s) used for the final prediction. (.) Untrained MPNNs show competitive performance as trained MPNNs given sufficiently large hidden dimensionality.""  
[X Link](https://x.com/marc_lelarge/status/1732441340485775560)  2023-12-06T16:46Z [----] followers, [----] engagements


"@gabrielpeyre @HaggaiMaron @lipmanya Thanks. I would not bet on your last statement. Graph isomorphism is not known to be in P nor known to be NP-complete but the quasi-polynomial time algo of Babai moved it closer to P"  
[X Link](https://x.com/marc_lelarge/status/1737391440848638065)  2023-12-20T08:36Z [----] followers, [---] engagements


"Les Vacances de Monsieur Hulot (Jacques Tati)"  
[X Link](https://x.com/marc_lelarge/status/1743752155499057304)  2024-01-06T21:51Z [----] followers, [--] engagements


"ET (Steven Spielberg)"  
[X Link](https://x.com/marc_lelarge/status/1743752158078533752)  2024-01-06T21:51Z [----] followers, [--] engagements


"Weekend puzzle: There are [--] boxes in a line. A cat sleeps in one of these boxes during the day. Every night the cat moves to either of the adjacent boxes (except for boxes [--] and [--] which are not adjacent). You can look inside one box each day. How can you find the cat"  
[X Link](https://x.com/marc_lelarge/status/1783759274256052590)  2024-04-26T07:25Z [----] followers, [----] engagements


"TIL: the Riemann zeta-function (s) = n is capable of fitting any arbitrary smooth function Voronin [--] Given any non-vanishing analytic function f(z) on a disk U in the strip 1/2Re(z)1 and any [--] we can find some real t such that for all zU (z+it)f(z) "  
[X Link](https://x.com/marc_lelarge/status/1798025346304053372)  2024-06-04T16:13Z [----] followers, 47.9K engagements


"TIL: Nicolaas Govert de Bruijn an eclectic mathematician: In mathematical logic the de Bruijn index is a tool for representing lambda calculus terms without naming the bound variables. In statistics De Bruijn's identity links the Shannon entropy and the Fisher information"  
[X Link](https://x.com/marc_lelarge/status/1812910988468883656)  2024-07-15T18:03Z [----] followers, [----] engagements


"@david_picard @sama All of them soon AGI is coming"  
[X Link](https://x.com/marc_lelarge/status/1843756374158913951)  2024-10-08T20:52Z [----] followers, [---] engagements


"Thanks to @alexUnder_sky for noticing that A* is always more efficient than Dijkstra when the heuristic is consistent. What about an admissible heuristic It is known that if the heuristic is admissible then A* returns an optimal path. Can you find a maze where Dijkstra's algorithm is more efficient (fewer nodes expanded) than A* with the Euclidean distance as heuristic https://t.co/rtOeMj16pk It is known that if the heuristic is admissible then A* returns an optimal path. Can you find a maze where Dijkstra's algorithm is more efficient (fewer nodes expanded) than A* with the Euclidean"  
[X Link](https://x.com/marc_lelarge/status/1876669707786576027)  2025-01-07T16:38Z [----] followers, [---] engagements


"*  is an informed search algorithm: starting from a starting node of a graph it aims to find a path to the given goal node having the smallest cost by maintaining a tree of paths extends those paths one edge at a time until the goal is reached"  
[X Link](https://x.com/marc_lelarge/status/1877778979647537254)  2025-01-10T18:06Z [----] followers, [----] engagements


""To equip more efficient smaller models with reasoning capabilities like DeekSeek-R1 we directly fine-tuned open-source models like Qwen and Llama using the 800k samples curated with DeepSeek-R1 (.) For distilled models we apply only SFT and do not include an RL stage (.)""  
[X Link](https://x.com/marc_lelarge/status/1881359618380878086)  2025-01-20T15:14Z [----] followers, [---] engagements


"The DeepSeek-R1 distilled 32B model found the cipher algorithm used for the demo of OpenAI o1-preview: https://openai.com/index/learning-to-reason-with-llms/ You can play with these models thanks to @ollama https://t.co/JwVOu2usMR https://t.co/YwUxbPBUtQ https://openai.com/index/learning-to-reason-with-llms/ You can play with these models thanks to @ollama https://t.co/JwVOu2usMR https://t.co/YwUxbPBUtQ"  
[X Link](https://x.com/marc_lelarge/status/1881462842886647844)  2025-01-20T22:04Z [----] followers, [---] engagements


"The optimization landscape of . Introducing Gaussian GRPO: a very simple GRPO algorithm to understand the basics of GRPO and the impact of the reward function on the global virtual reward maximized by GRPO. Explanations and code available at: https://github.com/dataflowr/notebooks/blob/master/llm/Gauss_GRPO.ipynb https://github.com/dataflowr/notebooks/blob/master/llm/Gauss_GRPO.ipynb"  
[X Link](https://x.com/marc_lelarge/status/1888930357497536862)  2025-02-10T12:37Z [----] followers, 14.7K engagements


"This summer saw significant progress at the intersection of LLMs and formal mathematics. Here are my slides:    from foundational concepts to recent advances. https://docs.google.com/presentation/d/1SZYrdKV4qZJW7y6gg5__bgtMUwazE9O25DnOHA9J0hM/editusp=sharing https://docs.google.com/presentation/d/1SZYrdKV4qZJW7y6gg5__bgtMUwazE9O25DnOHA9J0hM/editusp=sharing"  
[X Link](https://x.com/marc_lelarge/status/1962839282554425714)  2025-09-02T11:25Z [----] followers, [----] engagements


"@NoureddinSadawi Yes this course took place at ENS in Paris https://www.linkedin.com/posts/ecolenormalesuperieure_%C3%A0-lens-lintelligence-artificielle-est-activity-7398752752446902272-zAsQutm_source=share&utm_medium=member_android&rcm=ACoAAAHw_DQBr7ftBPWOqI4MsOGkfZ3ENIiUr1w https://www.linkedin.com/posts/ecolenormalesuperieure_%C3%A0-lens-lintelligence-artificielle-est-activity-7398752752446902272-zAsQutm_source=share&utm_medium=member_android&rcm=ACoAAAHw_DQBr7ftBPWOqI4MsOGkfZ3ENIiUr1w"  
[X Link](https://x.com/marc_lelarge/status/1996670682290061617)  2025-12-04T19:59Z [----] followers, [--] engagements


"Faire communiquer un modle de langage et un assistant de preuve - @Blog_Binaire https://www.lemonde.fr/blog/binaire/2025/05/23/faire-communiquer-un-modele-de-langage-et-un-assistant-de-preuve/ https://www.lemonde.fr/blog/binaire/2025/05/23/faire-communiquer-un-modele-de-langage-et-un-assistant-de-preuve/"  
[X Link](https://x.com/marc_lelarge/status/1926910494398394513)  2025-05-26T07:57Z [----] followers, [---] engagements


"Coding a python simulator for the Brownian bridge using Mistral's new "Vibe" CLI coding agent is super easy The  is a Brownian motion conditioned to have the same value [--] at time T as at the origin. The expected value of the bridge at any time t in the interval 0T is zero with variance t(T-t)/T. Increments are not independent: https://t.co/Gp923bjnUz The  is a Brownian motion conditioned to have the same value [--] at time T as at the origin. The expected value of the bridge at any time t in the interval 0T is zero with variance t(T-t)/T. Increments are not independent: https://t.co/Gp923bjnUz"  
[X Link](https://x.com/marc_lelarge/status/1998677658025025956)  2025-12-10T08:54Z [----] followers, [----] engagements


"Ready for tomorrows lecture:"  
[X Link](https://x.com/marc_lelarge/status/1999234521389293602)  2025-12-11T21:47Z [----] followers, 29.2K engagements


"Code (ready to run on @GoogleColab ): https://github.com/dataflowr/notebooks/blob/master/ModuleGPU/GPU_programming_basics.ipynb https://github.com/dataflowr/notebooks/blob/master/ModuleGPU/GPU_programming_basics.ipynb"  
[X Link](https://x.com/marc_lelarge/status/1999467484408348951)  2025-12-12T13:12Z [----] followers, [----] engagements


"Que signifie vraiment dcouvrir une particule invisible Du boson de Higgs au seuil des cinq sigma mon article sur le test dhypothse et la prudence statistique au cur de la physique des particules. https://larecherche.fr/physique/lhomme-invisible-de-la-physique https://larecherche.fr/physique/lhomme-invisible-de-la-physique"  
[X Link](https://x.com/marc_lelarge/status/2002679070258339851)  2025-12-21T09:54Z [----] followers, [----] engagements


"🌻 🧵Thread with links to all slides videos notebooks and @PyTorch code. Solutions are now available on the repo of the course: 1/15 https://github.com/dataflowr/notebooks https://github.com/dataflowr/notebooks"  
[X Link](https://x.com/marc_lelarge/status/1671080753193426944)  2023-06-20T09:01Z [----] followers, 75.8K engagements


"Theory and code for a simple denoising score matching algorithm for energy-based models (running in a Jupyter notebook locally). Links below"  
[X Link](https://x.com/marc_lelarge/status/1927293351969100276)  2025-05-27T09:18Z [----] followers, [----] engagements


"Course 👉 Code 👉 https://github.com/dataflowr/notebooks/blob/master/Module18/Denoising_Score_Matching_Energy_Model_sol.ipynb https://dataflowr.github.io/website/modules/18a-energy/ https://github.com/dataflowr/notebooks/blob/master/Module18/Denoising_Score_Matching_Energy_Model_sol.ipynb https://dataflowr.github.io/website/modules/18a-energy/"  
[X Link](https://x.com/marc_lelarge/status/1927293355127435338)  2025-05-27T09:18Z [----] followers, [---] engagements


"New course:   It covers Hypothesis Testing Bayes' Theorem Tree-based methods Convolutional neural network Attention and Transformers and more"  
[X Link](https://x.com/marc_lelarge/status/1996595306268328076)  2025-12-04T14:59Z [----] followers, [---] engagements


"Material is available here: https://github.com/mlelarge/ens-ml4sd https://github.com/mlelarge/ens-ml4sd"  
[X Link](https://x.com/marc_lelarge/status/1996595308537487868)  2025-12-04T14:59Z [----] followers, [---] engagements


"The  is a Brownian motion conditioned to have the same value [--] at time T as at the origin. The expected value of the bridge at any time t in the interval 0T is zero with variance t(T-t)/T. Increments are not independent:"  
[X Link](https://x.com/marc_lelarge/status/1998478891451625693)  2025-12-09T19:44Z [----] followers, [----] engagements


"Here are the prompts: "make a simulator for a Brownian bridge in python" "you need to activate a conda env before running python" "first create a conda env called mistral with the required python version and packages" "Make a dynamic simulation for the brownian bridge""  
[X Link](https://x.com/marc_lelarge/status/1998683757394493926)  2025-12-10T09:18Z [----] followers, [---] engagements


"Learn  from the ground up: begin with Numba for low-level control then progress to Triton to write high-performance kernels in a Python-like language. A hands-on Jupyter notebook to get you started quickly"  
[X Link](https://x.com/anyuser/status/1999467482025767113)  2025-12-12T13:12Z [----] followers, 56.7K engagements


"Want to learn formal mathematics with @CoqLang Start with Volume [--] of Software Foundations. Ive created a small Rocq tactics cheatsheet for the first chapters where you can interactively build your own proofs directly in the browser: Thanks @ejgallego https://mlelarge.github.io/rocq-proofs/ https://mlelarge.github.io/rocq-proofs/"  
[X Link](https://x.com/anyuser/status/2008965756591968377)  2026-01-07T18:15Z [----] followers, [---] engagements

Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing