🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

I realized much of the disagreement is about "what is the right baseline". If the claim is "<our RL> improves reasoning". The null hypothesis is RL just tuned the model to your specific inference hparams, so it's important to show changing hparams doesn't give the same gains.

🔥The debate’s been wild: How does the reward in RLVR actually improve LLM reasoning?🤔 🚀Introducing our new paper👇 💡TL;DR: Just penalizing incorrect rollouts❌ — no positive reward needed — can boost LLM reasoning, and sometimes better than PPO/GRPO! 🧵[1/n]

IMO, a major driver of PhD attrition is: do you have the willpower to finish a paper A full published econ/finance paper is like well over 100 pages of work after revisions and such, it's longer than the longest thing many students have ever done at that point in their lives

Really loved this eval! More people should work on this-- I will be (fingers crossed)

Announcing the release of the official Common Corpus paper: a 20 page report detailing how we collected, processed and published 2 trillion tokens of reusable data for LLM pretraining.

Thinking that mechanistic interpretability is the key to understanding AI safety is like thinking that neuroscience is the key to understanding political science.

Yep strong agree - 'having some way to consolidate experience is good. In the past, continual learning research never had a good response to “just retrain bro”, but maybe this will change in the age of experience'

Forecasting has a lot of potential, but lots of open problems to be tackled to get there. Really good reference for what critical problems it faces 👇

Elon Musk’s recent Twitter meltdown over the $2.4 trillion infrastructure bill seems contradictory given his own AGI timeline. If Musk genuinely believes AI will surpass human intelligence by late 2025, then he must also acknowledge the unprecedented productivity boom AGI would

Chat is this true?

Qwen3 Embedding Paper Embedding and reranking pipelines still struggle with task diversity and multilingual consistency. This report introduces a multi-stage training framework that uses Qwen3 LLMs to synthesize rich, multilingual relevance data, followed by contrastive

If you need some synthetic personas/people for your data projects, my team just put out a CC-BY-4.0 open dataset on HF of 100k folks matching US-census distributions for your free use. nvidia/Nemotron-Personas

New episode with Kenneth S Rogoff, former chief economist of the IMF. Ken predicts that, within the next decade, the US will have a debt-induced inflation crisis, but not a Japan-type financial crisis (the latter is much worse, and can make a country poorer for generations). Ken

I missed this when it came out and I think it is important. It seems that generally AI systems are not stress tested enough before publication/ press releases!

I'm in Nashville for CVPR and wow, the Music City name is not exaggerated. If you're around, we'll be presenting our work on temporal model merging with Vishaal Udandarao✈️CVPR'25, Karsten Roth, and Ameya P. (at CVPR) on Saturday 5-7 pm in ExHall D (poster #445). Come say hi!

Why can AIs code for 1h but not 10h? A simple explanation: if there's a 10% chance of error per 10min step (say), the success rate is: 1h: 53% 4h: 8% 10h: 0.002% Toby Ord has tested this 'constant error rate' theory and shown it's a good fit for the data chance of

moonshot doing as they do

my maybe most heretical startup opinion is more founders should quit sooner and do something else. yeah yc loves to talk about how long it took airbnb to take off but we don’t have the counterfactual of the better company chesky could have built instead if he quit.