Andrej Karpathy's new 'autoresearch' tool enables an AI agent to autonomously run roughly 100 LLM training experiments overnight on a single NVIDIA GPU, modifying code, evaluating results, and iterating — all guided by a simple plain-text file instead of manual coding.
DeepSeek, Tencent, and the University of Hong Kong unveil Pointer-CAD, a groundbreaking open-source AI framework built on Alibaba's Qwen 2.5 model that dramatically improves 3D design precision by reducing segmentation errors and streamlining complex entity selection in computer-aided design.
An AI trained with zero human data defeated its predecessor 100-0, sparking bold speculation that a future 'AlphaCode Zero' could learn programming from scratch and invent entirely alien coding paradigms that surpass human comprehension — though experts say such a system remains decades away.
MIT researchers unveil Attention Matching, a groundbreaking KV cache compression technique that slashes large language model memory usage by up to 50x in seconds, delivering near-identical accuracy without the slow, compute-heavy optimization methods that have long bottlenecked AI deployment.
A groundbreaking 'Bayesian Teaching' framework is training LLMs to reason probabilistically with 80% accuracy by fine-tuning them on interactions with an optimal Bayesian model, and the learned skills are successfully transferring across entirely different domains like hotels and web shopping.
A new Bitcoin Policy Institute study reveals that 81.5% of AI agents independently prefer bitcoin or stablecoins over fiat currency, with bitcoin dominating as a long-term store of value at 79.1% and traditional fiat facing a staggering 90.8% rejection rate — with advanced AI models showing even stronger bitcoin preferences.
Microsoft's new 15-billion-parameter AI model, Phi-4-reasoning-vision-15B, matches and even outperforms much larger rivals while using only one-fifth the training data, thanks to a 'mixed reasoning' design that intelligently switches between deep analytical thinking and fast direct responses depending on task complexity.
Apple unveils groundbreaking AI research that pinpoints the exact words where AI models hallucinate, transforming detection from a simple yes-or-no judgment into a precise, multi-step process that outperforms conventional methods — a critical breakthrough as the tech giant faces mounting pressure to ensure accuracy for its 2.5 billion devices worldwide.
A new reinforcement learning framework called RL4HS is outperforming existing AI models in detecting hallucinated spans in large language model outputs, using Group Relative Policy Optimization and a novel Class-Aware Policy Optimization technique to deliver superior results across summarization, question answering, and data-to-text tasks.
Researchers warn that AI large language models can now deanonymize pseudonymous social media users with up to 90% precision by browsing the web and analyzing free text, raising urgent concerns about government surveillance, corporate profiling, and targeted scams.