Less than a week left to #MLRC2025! Here is the agenda for the day. Location - Friend Center @ Princeton University maps.app.goo.gl/Cq39Ju3qNYCMy3…

The Transformer is not quite a Ship of Theseus, but, since 2017, best practices have radically changed for positional encodings, layer norms, attention, residual streams, and MLP components. Architecture work is not dead, and scaling is not all you need to respect scaling laws.

I’m happy to share that our paper "Controlling Multimodal LLMs via Reward-guided Decoding" has been accepted to #ICCV2025! 🎉 w/ Pierluca D'Oro, Koustuv Sinha, Adriana Romero-Soriano, Michal Drozdzal, and Aishwarya Agrawal 🔗 Read more: arxiv.org/abs/2508.11616 🧵 Here's what we did:

#MLRC2025 happening this week (August 21st)! The first keynote talk [1/4] will be from Arvind Narayanan at 9.45AM EST

#MLRC2025 happening this week (August 21st)! The second keynote talk [2/4] will be from Soumith Chintala at 10:30AM EST

#MLRC2025 happening this week (August 21st)! The third keynote talk [3/4] will be from Stella Biderman at 1pm EST

#MLRC2025 happening this week (August 21st)! The fourth keynote talk [4/4] will be from Jonathan Frankle at 1.45PM EST

Today we're releasing NVIDIA Nemotron Nano v2 - a 9B hybrid SSM that is 6X faster than similarly sized models, while also being more accurate. Along with this model, we are also releasing most of the data we used to create it, including the pretraining corpus. Links to the

Excited for tomorrow’s The ML Reproducibility Challenge #MLRC2025 event!! Really nice setup by Princeton Laboratory for Artificial Intelligence Princeton University Friend Center - looking forward to see you here tomorrow 9AM!

Awesome #MLRC2025 talks kicking us off this morning! I'm learning lots The ML Reproducibility Challenge about science with ML and reproducibility for real world applications (Arvind Narayanan), and software/firmware and data concerns for reproducibility (Soumith Chintala) Slides coming soon!

Its a wrap - the first in-person event for #MLRC2025 successfully concluded yesterday - we witnessed some of the best talks I have ever heard on reproducibility issues in AI, ranging from issues regarding leakage and irreproducibility in ML-based science (Arvind Narayanan),

Now that school is starting for lots of folks, it's time for a new release of Speech and Language Processing! Jim and I added all sorts of material for the August 2025 release! With slides to match! Check it out here: web.stanford.edu/~jurafsky/slp3/

On the contrary, when history remembers people who stood in the face of unbridled hype constructively and in line with the scientific method, let David be thus-remembered over the perpetual goalpost-moving Garys of the world.

Quality > Quantity. I’d love to see efforts to further reduce the size of pretraining data while keeping the downstream evals constant. BabyLM was a step in this direction, but the focus was to design the right architecture. We need a GPT speedrun equivalent of this, where the

So happy to share what we have been working on for the past 2 years Ai2. Data-driven discovery sits at the core of Asta, along with amazing literature discovery tools! You probably have read our works, now see them in action 🛼🛼

I’m excited to announce I’ve started a YouTube channel. I plan to publish videos regularly explaining my views on AI and its present and future impacts. My first video asks: What happens if there’s an AI crash? youtube.com/watch?v=VDfyuB… This is my first foray into video (beyond

Happy to share a side project: Diffusion Language Models Know the Answer Before Decoding. Diffusion LMs are often dismissed as slow. But what if they already *know* the answer halfway through? 1. Early Answer Convergence: Our new paper shows that in many cases, they do,

Subscribed!

LLMs are great at single-shot problems, but in the era of experience, interactive environments are key 🔑 Introducing * Multi-Turn Puzzles (MTP) * , a new benchmark to test multi-turn reasoning and strategizing 🔗 Paper: huggingface.co/papers/2508.10… 🫙Data: huggingface.co/datasets/arian…

Super excited about this event! I will give an updated version of my talk on the Simulation Hypothesis - i.e. Machine Learning in the upcoming era of extremely fast computers. How can we do science that stands the test of time when compute capacity is accelerating?