Louis Mullie, MD

4.4K posts

Louis Mullie, MD

Louis Mullie, MD

@LouisMullie

director of medical AI @doximity, ICU doc, teacher @chumontreal

Montréal, Québec เข้าร่วม Kasım 2011
2.8K กำลังติดตาม2.9K ผู้ติดตาม
Louis Mullie, MD รีทวีตแล้ว
Fengzhuo Zhang
Fengzhuo Zhang@FengzhuoZhang·
The Newton–Schulz iteration coefficients optimized by DeepSeek-V4 are surprisingly strong: they effectively normalize all singular values to 1. This matches our previous intuition: a well-balanced spectrum may help strike a better balance across long-tail knowledge. Plot code: github.com/FengzhuoZhang/…
Fengzhuo Zhang tweet mediaFengzhuo Zhang tweet media
Fengzhuo Zhang@FengzhuoZhang

Why does Muon outperform Adam—and how? 🚀Answer: Muon Outperforms Adam in Tail-End Associative Memory Learning Three Key Findings: > Associative memory parameters are the main beneficiaries of Muon, compared to Adam. > Muon yields more isotropic weights than Adam. > In heavy-tailed tasks, Muon significantly improves tail-class learning compared to Adam. Paper Link: arxiv.org/pdf/2509.26030 A thread 🧵

English
8
59
426
66.3K
Louis Mullie, MD รีทวีตแล้ว
elie
elie@eliebakouch·
Deepseek V4 Pro is the biggest open model ever with 1.6T total 49B active, trained on 33T tokens, 1M context, with 2 new attention mechanisms, Muon, mHC, open source kernels, FP4 QAT, MIT license and with one of the best tech repot of the year
elie tweet media
English
20
165
1.4K
129K
Louis Mullie, MD รีทวีตแล้ว
Joelle Pineau
Joelle Pineau@jpineau1·
Exciting news for the future of AI! @cohere and Aleph Alpha are partnering to build a global sovereign AI platform. As demand for AI grows, we are accelerating the development of next-generation frontier models while upholding strong standards for data security and ethical AI.
Cohere@cohere

🚀 Sovereign AI for the world. Cohere & Aleph Alpha form transatlantic AI powerhouse anchored in Canada & Germany! Combining our global scale with European R&D excellence to build sovereign, enterprise-grade AI. Security, privacy & trust for businesses & governments worldwide. #SovereignAI #AIPartnership Learn more: businesswire.com/news/home/2026… Image from left to right: Rolf Schumann, Schwarz Digits, Samuel Weinbach, Aleph Alpha, Aidan Gomez, Cohere, Minister Solomon, Canada, Minister Wildberger, Germany

English
4
11
80
7.5K
Louis Mullie, MD รีทวีตแล้ว
Cohere
Cohere@cohere·
🚀 Sovereign AI for the world. Cohere & Aleph Alpha form transatlantic AI powerhouse anchored in Canada & Germany! Combining our global scale with European R&D excellence to build sovereign, enterprise-grade AI. Security, privacy & trust for businesses & governments worldwide. #SovereignAI #AIPartnership Learn more: businesswire.com/news/home/2026… Image from left to right: Rolf Schumann, Schwarz Digits, Samuel Weinbach, Aleph Alpha, Aidan Gomez, Cohere, Minister Solomon, Canada, Minister Wildberger, Germany
Cohere tweet media
English
10
25
188
111.9K
Louis Mullie, MD
Louis Mullie, MD@LouisMullie·
@DrBeavisAI Happy to know you like it! We’ve put in a ton of work on preventing hallucinations so it’s great to see users notice that. To be sure, DoxGPT is fully HIPAA compliant - glad to answer any questions about this if part of our terms are unclear :)
English
1
0
1
92
G, MD
G, MD@DrBeavisAI·
Doximity doesn't provide real HiPAA either because privacy policy states it stores and uses all prompts, so unclear what even happens if chats are deleted. DoxGPT doesn't truly web search nor uses a thinking model which openAI is offering. Unclear which model is used for DoxGPT, I assume maybe 5.0? Doxmity app on android needs a widget 2x1 size then I would use it more. DoxGPT is better than open Evidence and I don't understand why folks use that, good marketing to genX/boomers?? Its LLM hallucinates the most.. its only good for simple Q.
English
1
1
1
449
Louis Mullie, MD
Louis Mullie, MD@LouisMullie·
It’s great to see more people entering the race - but let’s be honest… ChatGPT for clinicians - Doesn’t provide HIPAA compliance out of the box - Doesn’t have an integrated AI scribe or telehealth platform - Doesn’t integrate an expert-reviewed drug information product - Doesn’t have a system for peer review of AI answers (PeerCheck) - Doesn’t have enterprise agreements in place with 100+ hospitals Meanwhile, the “model moat” is getting thinner than ever as post-training for specialized use cases shows strong results vs. frontier models. OpenAI is going to make the race even more competitive than it already is, but by no means is it leading out of the gates. Game on!
Avi Roy@agingroy

Two out of three doctors already use ChatGPT for clinical work, per the @AmericanMedAssn. @OpenAI just made it official with a free version built for clinicians. @EvidenceOpen, @GlassHealthHQ, @doximity spent years building specialized tools. OpenAI walked in and made theirs free. Good luck competing with that. The benchmark matters more to me. We need a way to compare these tools. But OpenAI built the test and is grading its own models. That's a drug company running its own trial. @MGHMedicine tested 21 of these models on real patient cases two weeks ago. They got the diagnosis wrong more than 80% of the time. Nobody paid for that study. That's what real testing looks like.

English
10
9
64
14.9K
Position Journal
Position Journal@PositionJournal·
@LouisMullie @jonathanhershon Finally someone from Doximity actually marketing the product properly. Go and show what your tools actually offer, your shareholders are patiently waiting 🫡
English
1
0
2
203
Louis Mullie, MD รีทวีตแล้ว
shubham
shubham@bluequbit·
@nickinpractice Low latency, hot swappable lora
English
0
1
3
364
Louis Mullie, MD รีทวีตแล้ว
Lukas (computer) 🔺
Lukas (computer) 🔺@SCHIZO_FREQ·
This is always how I assumed LLMs would wind up functioning because this is how I (and presumably most others) think I assume the base unit of thought is this gestalt thought vector thing, not "words," and we've just all developed a very fast way to translate these to words because words are more communicable than thought pieces This was always my issue with "some people don't have an internal monologue!" discourse It just makes no sense for words to be the base unit people think in. It's like 1000x faster to think in terms images or these thought pieces or whatever I assume it just seems like people think in words bc when they describe what they're thinking to people, they have to translate the thought pieces to words - as that's how we communicate - and this process converts their actual thoughts into the form of a monologue But it only makes sense to think in words when you need to output some form of communication. Otherwise it's not very efficient And human brains are insanely efficient
Simplifying AI@simplifyinAI

🚨 BREAKING: Tencent has killed the “next-token” paradigm. Tencent and Tsinghua has released CALM (Continuous Autoregressive Language Models), and it completely disrupts the next-token paradigm. LLMs currently waste massive amounts of compute predicting discrete, single tokens through a huge vocabulary softmax layer. It’s slow and scales poorly. CALM bypasses the vocabulary entirely. It uses a high-fidelity autoencoder to compress chunks of text into a single continuous vector with 99.9% reconstruction accuracy. The model now predicts the “next vector” in a continuous space. The numbers are actually insane: - Each generative step now carries 4× the semantic bandwidth. - Training compute is reduced by 44%. - The softmax bottleneck is completely removed. We’re literally watching language models evolve from typing discrete symbols to streaming continuous thoughts. This changes the entire trajectory of AI.

English
64
62
1.2K
118.8K
Louis Mullie, MD รีทวีตแล้ว
eigenron
eigenron@eigenron·
i don't understand why this paper did not get much traction. they GRPO'd a small base model on its own confidence scores (internal rewards) instead of external rewards and it shows comparable results on math and coding benchmarks compared to models trained with GRPO with external rewards.
eigenron tweet media
English
20
68
816
48.3K
Andrey Kurenkov
Andrey Kurenkov@andrey_kurenkov·
This story about @openblocklabs blatantly cheating on terminal bench, and the founder's reply after being exposed, was so wild I couldn't help but be curious about this 'prev. ML research @stanford' note on his profile...' And oh man, let's just say that might be a bit misleading. His Linkedin says "ML Research Scientist Stanford University Aug 2016 - Sep 2018 · 2 yrs 2 mos Youngest author at NeurIPS 2017 (ML4H); 200+ citations Working on GANs @ MRSRL Lab Mentors: J. Cheng, M. Mardani, J. Pauly" That appears to come down to a single paper, 'Synthetic Medical Images from Dual Generative Adversarial Networks', which be co-authored with 2 other students while in high school. But, none of the listed 3 mentors are co-authors, nor is anyone from the MRSRL Lab. It appears to have been presented as a poster at the Machine Learning for Health workshop, though oddly only his co-authors are listed there (neurips.cc/virtual/2017/w…). And having just read it (arxiv.org/abs/1709.01872), I think it's safe to say it would not have been accepted to Neurips proper (though still very cool of high school students to do this level of research). So his linkedin has 'ML Research Scientist @ Stanford University' **seemingly** because he had or sought unofficial mentorship from some people at Stanford while doing research in high school with some schoolmates. So maybe having 'prev. ML research @stanford' in his twitter bio is just a tiny bit misleading...
Andrey Kurenkov tweet mediaAndrey Kurenkov tweet mediaAndrey Kurenkov tweet mediaAndrey Kurenkov tweet media
Monk Zero@NoCommas

x.com/i/article/2032…

English
14
5
223
37.4K
Agus 🔸
Agus 🔸@austinc3301·
Turns out they just committed straight-up fraud? Their package prebundles solutions for the benchmark and then gives them to the agent when it starts
Agus 🔸 tweet media
English
23
34
792
94.3K
Louis Mullie, MD รีทวีตแล้ว
Peter Schmidinger
Peter Schmidinger@PeteSchmidinger·
They did not "just did" it, the technique is almost a year old. It is not "impossible", it is possible and is common knowledge. It is currently in production (they released the toolkit before the paper). Also, your phrasing is open to misinterpretation. The model is not called NVPF4, the 4-bit floating-point data format is. The "first successful demonstration of large-scale 4-bit pretraining" was done before by 2501.17116 (13B model, 100B tokens).
English
2
1
12
902
Chris Laub
Chris Laub@ChrisLaubAI·
🚨 NVIDIA just did the impossible. They trained a 12B-parameter language model on 10 trillion tokens entirely in 4-bit precision. It’s called NVFP4, and it might redefine how frontier AI models are trained. Here’s why this matters: • NVFP4 delivers 2–3× faster math throughput and 50% less memory vs FP8 • Accuracy? Practically identical. (MMLU-Pro: FP8 = 62.62%, NVFP4 = 62.58%) • Stability issues? Solved using Random Hadamard transforms, stochastic rounding, and 2D scaling • Trained entirely on NVIDIA Blackwell GPUs the first 4-bit run stable across 10T tokens This is the first successful demonstration of large-scale 4-bit pretraining without losing accuracy. The next generation of frontier models will be faster, cheaper, and greener without compromise.
Chris Laub tweet media
English
52
126
837
58.9K
Louis Mullie, MD รีทวีตแล้ว
Yulu Gan
Yulu Gan@yule_gan·
Simply adding Gaussian noise to LLMs (one step—no iterations, no learning rate, no gradients) and ensembling them can achieve performance comparable to or even better than standard GRPO/PPO on math reasoning, coding, writing, and chemistry tasks. We call this algorithm RandOpt. To verify that this is not limited to specific models, we tested it on Qwen, Llama, OLMo3, and VLMs. What's behind this? We find that in the Gaussian search neighborhood around pretrained LLMs, diverse task experts are densely distributed — a regime we term Neural Thickets. Paper: arxiv.org/pdf/2603.12228 Code: github.com/sunrainyg/Rand… Website: thickets.mit.edu
Yulu Gan tweet media
English
87
435
3K
686.2K
Louis Mullie, MD รีทวีตแล้ว
Benjamin Thérien @ ICLR 2026
Benjamin Thérien @ ICLR 2026@benjamintherien·
🚨 New Tech Report: Covenant-72B 🚨 TL;DR we use SparseLoCo to pre-train a 72B model on 1.1T tokens over the internet! This is the largest decentralized training run to date. x.com/tplr_ai/status…
templar@tplr_ai

We just completed the largest decentralised LLM pre-training run in history: Covenant-72B. Permissionless, on Bittensor subnet 3. 72B parameters. ~1.1T tokens. Commodity internet. No centralized cluster. No whitelist. Anyone with GPUs could join or leave freely. 1/n

English
2
7
70
5.2K

ค้นพบ

@cohere @DrBeavisAI @PositionJournal @jonathanhershon @nickinpractice @andrey_kurenkov @openblocklabs @Stanford