News Hub

K-Search translates CUDA optimization knowledge to MLX strategies for Apple Silicon.

Researchers developed methods to teach large language models to update their beliefs efficiently during long-horizon interactions.

AI inference costs fell 50x median per year, with GPT-4-class capabilities dropping from $30 to under $1 per million tokens.

Berkeley AI Research Lab celebrates 2026 Ph.D. graduates whose work spans robotics, embodied intelligence, large language models, and reasoning.

Adaptive Parallel Reasoning proposed as new paradigm for efficient inference scaling.

Method identifies interactions at scale in LLMs to improve interpretability and transparency.

Berkeley AI Research developed information estimator to quantify measurement quality in imaging systems.

Berkeley AI Research proposes divide-and-conquer reinforcement learning without temporal difference learning for long-horizon tasks.

Researchers provide quantitative theory explaining what word2vec learns during representation learning.