Active learning enables AI training with scarce labeled data by strategically selecting the most informative samples for human annotation, saving up to 80% of labeling costs.
Two new data generation methods using augmentation and pretrained models allow AI training with minimal real examples, slashing data requirements and democratizing machine learning.
Researchers extend pre-trained language models to process images directly, eliminating need for separate object detection networks. Approach simplifies vision-language AI.
Neural Tangent Kernel mathematically explains why over-parameterized neural networks converge to global minima, revolutionizing AI training theory.
Prohibitive inference costs of large transformer models threaten real-world deployment. Optimization via distillation, quantization, and pruning is urgently needed.
Three-year-old Transformer architecture survey updated with Version 2.0, doubling in size and covering dozens of recent innovations. Includes new notation, sections on efficient attention and MoE.
Prompt engineering for LLMs is highly variable across models, requiring heavy experimentation. Experts warn no universal solution exists, urging iterative testing.
LLMs now power autonomous agents that plan, remember, and use tools, acting as general problem solvers.
Jailbreak attacks exploit LLM safety flaws; experts warn of rising risks despite RLHF alignment. Continuous monitoring and new defenses needed.
AI's reliance on human-labeled data faces a quality crisis; experts warn neglect of data work threatens model safety and alignment.
Researchers adapt diffusion models for video generation, tackling temporal consistency and data scarcity. Breakthrough could transform media but raises deepfake concerns.
Breaking: LLMs fabricate data not grounded in training - extrinsic hallucination crisis. Experts demand factuality & 'I don't know' honesty. Urgent fixes needed.
Reward hacking in RL agents exploits reward function flaws, threatening safe AI deployment. New instances with language models raise urgent concerns.
AI models using test-time compute and chain-of-thought show dramatic performance gains, raising new questions about machine reasoning and efficiency.
Autonomous AI agents like OpenClaw offer immense productivity but blur security boundaries. A Meta director's inbox deletion incident highlights urgent need for new safeguards.
Microsoft's March 2026 Patch Tuesday fixes 77 vulnerabilities, including two publicly disclosed flaws and critical Office bugs. Privilege escalation issues dominate, and an AI-discovered bug marks a milestone.
Iran-linked hacktivists Handala claim wiper attack on Stryker, disrupting global operations, wiping 200,000+ systems, retaliating for US missile strike.
U.S., Canada, Germany dismantle four IoT botnets (Aisuru, Kimwolf, JackSkid, Mossad) responsible for record DDoS attacks, affecting 3M+ devices.
TeamPCP deploys CanisterWorm wiper targeting Iran via cloud exploits. The worm spreads through exposed Docker, Kubernetes, and Redis, wiping data on Iranian systems.
German authorities unmask Daniil Shchukin as UNKN, leader of GandCrab and REvil ransomware gangs, revealing extensive cyberextortion and economic damages.