Marc Romeyn (@MarcRomeyn) profile picture

Senior ML engineer @nvidia. Building open-source LLM tooling (NeMo & NeMo-Run). Deep Learning, ML-infra & Recsys. ex @spotify, @efLDN alum. 🇳🇱

Stockholm, Sweden
Joined May 2011
Photo by Marc Romeyn (@MarcRomeyn)Photo by Marc Romeyn (@MarcRomeyn)Photo by Marc Romeyn (@MarcRomeyn)Photo by Marc Romeyn (@MarcRomeyn)Photo by Marc Romeyn (@MarcRomeyn)Photo by Marc Romeyn (@MarcRomeyn)
Marc Romeyn retweeted
NVIDIA Nemotron 3 Ultra is now live! Frontier accuracy, 5X greater speed, 30% lower cost. Deploy however you need - on-premise, on the cloud, or at the edge. Model is live on HuggingFace under the OpenMDW 1.1 license. youtube.com/watch?v=D8LIIvQV…
25
74
14
370
95,895
Marc Romeyn retweeted
Nemotron 3 Super is here — 120B total / 12B active, Hybrid SSM Latent MoE, designed for Blackwell. Truly open: permissive license, open data, open training infra. See analysis on @ArtificialAnlys Details in thread 🧵below:
9
44
10
273
30,468
BREAKING: Warren Buffett's Berkshire Hathaway, $BRK.A, officially becomes the first US company outside of tech to hit $1 trillion in market cap.
47
266
34
1,987
149,987
Cool post from my old team at Spotify!
Crafting listening experiences for users through reinforcement learning. 🧠💡🎧 Explore how @Spotify leveraged the TensorFlow ecosystem to solve recommendation problems, design an extendable offline simulator, and train RL Agents to generate playlists. → goo.gle/48XiRDE
1
5
678
Very excited this is out! Showcases some new APIs of Merlin Models we have been working on last summer.
In #recsys you might want to predict different types of user events (click, like, share, purchase). In this post we introduce multi-task learning (MTL) for ranking models and how to easily build and train such models using #Merlin and #tensorflow. medium.com/nvidia-merlin/bui…
1
5
931
Good thread on some of the new features of Merlin that I have been working on for the last year. Stay tuned, much more to come!
I've been playing around with NVIDIA's Merlin Models library for the last couple days for the H&M Kaggle competition and here are my thoughts on the library! 🧵
6
I had fun time talking to @anders_arpteg on the AIAW podcast. We spoke about a wide variety of different topics, spanning from @nvidia, @Spotify, @joinef, Recsys, ML-ops & Web-3. It can be found as video (youtu.be/J3JUXZ5jjOw) + podcast platforms like spoti.fi/3EstvDF
6
I am a bit embarrassed that I never really knew what ASML did as a Dutch person. I learned a lot from this thread
ASML is the most important company you've never heard of. The $300B+ Dutch firm makes the machines that make semiconductors. Each one costs $150m and access to them are a huge geopolitical flashpoint. Here's a breakdown 🧵
7
I am very excited about this! We're currently working hard on getting the library ready for open-sourcing, so stay tuned for that
Incredibly excited to share that our @NVIDIAAI paper "Transformers4Rec: Bridging the Gap between NLP and Sequential/Session-Based Recommendation" which wraps @huggingface transformers for use in RecSys has been accepted to @ACMRecSys. See you in Amsterdam #RecSys2021!
1
7
1/ Personal update: after 4 amazing years at @spotify, I am very excited to start a new chapter today at @nvidia. I will be working on developer.nvidia.com/nvidia-…, an open source end-to-end GPU-accelerated recsys platform.
4
6
2
70
3/ Building large-scale recommender systems is notoriously hard, and open-source tooling is arguably behind compared to more mature ML-fields like Computer Vision & NLP.
1
6
4/ I can't wait to help trying to raise the bar of this tooling with the goal of advancing and democratizing Recsys for everyone.
1
6
Marc Romeyn retweeted
… and the final results for the creative track have just been published: recsys-challenge.spotify.com… Congrats to all!
8
10
Several machine learning researchers have signed a statement regarding the upcoming launch of Nature Machine Intelligence. If you agree, I encourage you to sign this as well. openaccess.engineering.orego…
29
913
188
1,510
Marc Romeyn retweeted
ImageNet 2017 results are out. Beijing-based momenta.ai team achieved top-5 error of 2.25%, down from ~3% last year. image-net.org/challenges/LSV…
4
199
13
329
Awesome stuff! Was thinking of something similar, happy I don't have to build it @jimmfleming nitter.gq/jimmfleming/status/885…
We're open-sourcing our project scaffolding tool TFStage for TensorFlow: buff.ly/2uOm1X7 #MachineLearning #TensorFlow
4
Marc Romeyn retweeted
Attention Is All You Need, from @GoogleBrain. No RNNs, No CNNs, Just Attention. Very fast training; SOTA on WMT'14. arxiv.org/abs/1706.03762
14
324
20
800
Clip from the latest mind-blowing neural network paper by @anorangeduck and team! Full video here: theorangeduck.com/media/uplo…