Gemma 4: Google's Small Model Beats Larger Rivals
Google's Gemma 4 open-weight model family demonstrates that smaller, efficient architectures can outperform much larger AI models on key benchmarks, reshaping how developers approach multimodal AI.
Latest
LLM Research
Why Most LLM Ablation Studies Produce Misleading Results
Ablation studies are the gold standard for understanding which components of an LLM matter most. But systemic methodological flaws mean many published results may be fundamentally unreliable.
AI Security
Tool Injection Attacks Can Hijack AI Agents
Researchers demonstrate how a malicious tool can hijack an AI agent's behavior, feeding users fabricated information — revealing critical vulnerabilities in agentic AI systems.
Knowledge Distillation
Knowledge Distillation: Compressing AI Ensembles
Knowledge distillation transfers ensemble intelligence into compact, deployable models. Here's how this critical technique powers efficient AI systems for video generation and deepfake detection.
deepfake fraud
Deepfake Scams Erode Trust in Global Banking Systems
A deepfake scam targeting banking customers highlights the growing threat of AI-generated identity fraud in financial services, raising urgent questions about authentication and digital trust.
deepfake fraud
Deepfake Fraud Threatens Real Estate Transactions
Real estate professionals face a growing wave of deepfake-powered fraud, from synthetic identity verification to voice-cloned wire transfer requests. Here's how the industry is fighting back.
deepfakes
AI Deepfakes on TikTok Fuel Russian Anti-Ukraine Campaign
Russian propaganda operations are deploying AI-generated deepfake videos on TikTok to undermine Ukraine's military mobilization efforts, highlighting growing weaponization of synthetic media in geopolitical conflicts.
Synthetic Media
Pro-Iran Group Deploys AI Lego Cartoons as Propaganda
A pro-Iranian media outlet called Explosive Media is using AI-generated Lego-style cartoons to mock U.S. leadership, highlighting how accessible synthetic media tools are enabling new forms of state-aligned propaganda.
Alibaba
Alibaba's Secret 'Happy Horse' Model Tops AI Leaderboards
Alibaba reveals it secretly submitted its Qwen model under the alias 'Happy Horse,' topping major AI leaderboards and raising questions about benchmark gaming and anonymous submissions.
LLM Research
LLM Training as Lossy Compression: What Models Forget
New research reframes LLM training as lossy compression, revealing that learning is fundamentally about strategic forgetting. The information-theoretic framework explains how models selectively discard data while retaining generalizable patterns.
AI governance
Coasean Bargain Framework for AI Copyright & Data Scraping
New research proposes applying Coasean economic theory to resolve AI copyright disputes over training data scraping, offering a market-based framework for balancing creator rights with AI innovation.
AI Video Generation
New Method Accelerates Autoregressive Video Model Training
A new research paper proposes a local optimization method with representation continuity to speed up training of autoregressive video generation models, addressing key computational bottlenecks.