Machine Learning
273 articles in this category (Page 3 of 12)
Subliminal Learning: How LLMs Inherit Hidden Behavioral Traits via Synthetic Data
New research in Nature reveals student LLMs inherit teacher behavioral traits through hidden signals in synthetic data, even when datasets are semantically unrelated.
Meta AI and KAUST Propose Neural Computers: Folding Computation and Memory into One Learned Model
Meta AI and KAUST researchers introduce Neural Computers (NCs), achieving 98.7% cursor accuracy in GUI prototypes by folding OS functions into a single learned runtime state.
Sigmoid vs ReLU: Why Geometric Context Preservation is Critical for Neural Network Inference
ReLU outperforms Sigmoid by preserving geometric distance from decision boundaries, achieving 96% accuracy compared to Sigmoid's 79% in two-moons benchmarks.
Five AI Compute Architectures Every Engineer Should Know: CPUs, GPUs, TPUs, NPUs, and LPUs Compared
Understand the trade-offs between AI architectures, including Groq’s LPU which achieves 10x higher energy efficiency than traditional systems for LLM inference.
Google AI Research Introduces PaperOrchestra: A Multi-Agent Framework for Automated AI Research Paper Writing
Google AI Research debuts PaperOrchestra, a multi-agent system that transforms raw experimental logs into submission-ready LaTeX papers, achieving simulated acceptance rates of up to 84%.
Meta AI's EUPE: A <100M Parameter Universal Vision Encoder Rivaling Specialists
Meta AI introduces EUPE, a compact vision encoder under 100M parameters that matches domain-expert models in classification and dense prediction, achieving 55.2ms latency on iPhone 15 Pro.
Engineering Production-Ready RAG Pipelines: Lessons from the Python Ecosystem
Learn how to move RAG from prototype to production using Python, FAISS, and SentenceTransformers while managing latency and data consistency for datasets under 100,000 chunks.
Google DeepMind AlphaEvolve: LLM-Driven Evolutionary Search Outperforms Human-Designed Game Theory Algorithms
DeepMind's AlphaEvolve uses Gemini 2.5 Pro to evolve MARL source code, discovering algorithms that outperform expert-designed baselines in 10 of 11 test games.
Hugging Face Releases TRL v1.0: A Unified Post-Training Stack for SFT, Reward Modeling, DPO, and GRPO Workflows
Hugging Face TRL v1.0 standardizes LLM post-training with a unified CLI and config system, delivering up to 2x training speed and a 70% reduction in memory usage.