"Parameter Golf" competition attracts over 2,000 submissions on AI optimization
The Parameter Golf event successfully concluded with thousands of creative ideas on AI model optimization, including quantization, TTT LoRA, and SSMs.
Sources x.com
Lead reasoning researcher departs Meta FAIR
A key member of Meta FAIR has announced their departure after two years of leading critical research into the reasoning capabilities of large language models.
Sources x.com
The Mathematics Behind Neural Tangent Kernel in AI 🧠
Lilian Weng's article analyzes the mathematical foundation of the Neural Tangent Kernel (NTK), explaining how over-parameterized neural networks efficiently converge during training.
Sources lilianweng.github.io
The Art of Prompt Engineering: Shaping the Mindset of LLMs 🧠
Prompt Engineering helps effectively optimize the steerability of Large Language Models (LLMs) without the need to update model weights.
Sources lilianweng.github.io
New studies untangle reinforcement learning (RL) bottlenecks 🤖
Studies on arXiv propose solutions for sim-to-real transfer, off-policy optimization, and opponent behavior shaping in multi-agent environments.
New PNAS Study Reveals Breakthrough in AI Modeling
A new study published in the journal PNAS introduces novel optimization methods for large-scale AI systems.
Sources x.com
Hugging Face Launches Reasoning Model Achieving Gold Medal-Level Performance in Physics and Math Olympiads
The new 30B-A3B reasoning model achieves gold medal-equivalent performance in the IPhO and IMO exams, powered by a simple scaling recipe for proof search.
Sources x.com
Yann LeCun: Hierarchical World Models to Arrive Within 18 Months
Meta's Chief AI Scientist predicts that AI will soon be able to learn from videos to build hierarchical world models, helping robots plan complex actions in the real world.
Sources x.com
Math Proof: World Models Exponentially More Data-Efficient Than LLMs 📊
New research by Matthieu Wyart provides mathematical proof that World Models like JEPA are exponentially more sample efficient than LLMs by predicting abstract representations.
The more "obedient" AI is, the less human-like it becomes — The paradox of chatbot training 🤖
A large-scale study reveals that prioritizing usefulness in AI training unintentionally weakens its ability to simulate natural human behavior.
Sources the-decoder.com
Release of stable-worldmodel — an open-source platform for JEPA and World Model research
After a year of development, stable-worldmodel has been officially launched. It is an open-source, scalable platform designed to accelerate AI research in JEPA and World Models.
Sources x.com
DynaSchedBench: Decoding the LLM 'Observability Paradox' in Dynamic Scheduling
A new study introduces DynaSchedBench, a standardized benchmark for the Dynamic Flexible Job-Shop Scheduling Problem (DFJSP), exposing the limitations of AI agents when exposed to excessive data.
Sources arxiv.org
Yann LeCun: The Core Difference Between the Engineer and Scientist Mindset 🧠
Meta's Chief AI Scientist Yann LeCun shares deep insights into the differing goals and working methodologies of engineers and scientists in the tech industry.
Sources x.com
Microsoft Research: AI is an extension of human intelligence, not a replacement 🧠
Microsoft Research proposes viewing AI as a tool to extend human cognitive capabilities rather than completely replace them, aiming to build more trustworthy AI systems.
Sources microsoft.com
JEPA-WM research receives reproducibility certification from TMLR journal 🧠
Yann LeCun's JEPA-WM world model has been awarded an empirical reproducibility certification by the TMLR journal, confirming its transparency and mathematical stability.
Sources x.com
29,000-Word Deep Dive into FlashAttention-2 in CuTe Released
An incredibly detailed technical document analyzing every line of FlashAttention-2's production source code has been released, with an estimated reading time of 100 hours.
Sources x.com
Microsoft Research announces AI capable of autonomously operating repositories and 'verification-first' research
Microsoft Research has introduced new AI solutions capable of autonomously running repositories alongside a 'verification-first' research methodology.
Sources x.com
Robotics Endgame: Jim Fan's Roadmap to Decoding Physical AGI
Dr. Jim Fan presents his 'Robotics: Endgame' talk, proposing a roadmap to solve physical artificial general intelligence (Physical AGI) akin to the success of LLMs.
Sources x.com