DeepSeek
36 articles found in this topic.
AI Models Exhibit Deception and Strategic Bluffing in Competitive Game Arenas
AI models are now demonstrating advanced deception and strategic bluffing in competitive game environments like Werewolf and Texas Hold'em. A Kaggle Game Arena competition in 2026 showcased models like Gemini 3 Pro and DeepSeek V3.2 employing cunning tactics, including logical traps and bluffs, to outmaneuver opponents. This marks a new era where AI's strategic 'cunning' is being tested beyond traditional computational benchmarks.
ICLR Paper Introduces "Lookup-Based Memory" for Transformers Ahead of DeepSeek Engram
An ICLR paper introduced "lookup-based memory" for Transformer models, predating DeepSeek Engram. This approach, STEM, reconfigures the FFN to use static lookup from a token-indexed embedding table, separating memory capacity from computational overhead. It offers advantages like plug-and-play knowledge editing.
Hugging Face Details Transformer Library Evolution for Mixture-of-Experts Models
Hugging Face details the significant redesigns in its Transformers library to support Mixture-of-Experts (MoE) models. This evolution addresses challenges in model loading and execution, crucial for integrating MoE architectures that enable larger LLMs with improved computational efficiency and scalability.
Google Research Challenges "Technological Singularity" with New AI Intelligence Explosion Theory
Google researchers and collaborators propose a new theory of AI intelligence explosion, challenging the traditional "technological singularity" concept. Their study suggests that AI's intelligence growth is diverse, social, and stems from internal multi-agent interactions within models, dubbed a "society of thought."
DeepSeek V4 and Tencent's New Hunyuan Model Anticipated for April Release
DeepSeek V4 and a new Tencent Hunyuan model are both slated for an April 2026 launch. DeepSeek V4 focuses on long-term memory and multimodality, while Tencent's Hunyuan model, led by Yao Shunyu, emphasizes contextual learning and real-world task evaluation.
OpenAI Research Reveals Advanced AI Models Struggle to Control Their Reasoning Processes
New OpenAI research reveals advanced AI models, despite excelling at final answers, struggle significantly with controlling their internal reasoning or "chain of thought." The study, evaluating 13 models, found that more powerful reasoning capabilities often correlated with less adherence to specific constraints during the thought process, highlighting a critical discrepancy between output control and internal process control.
OpenClaw Home Installation Services Highlight Market Chaos and Security Risks
Paid home installation services for OpenClaw, an open-source AI agent, reveal a chaotic market with vast price disparities and significant security vulnerabilities. Third-party installers pose risks, as users lack verification methods for software integrity and installed skills, potentially exposing systems to threats.
DeepSeek Founder Liang Wen-feng Named to Nature's Top 10 Scientists of the Year
Liang Wen-feng, founder of DeepSeek, has been named one of Nature's top ten scientists of the year for his work with DeepSeek's AI models, including the open-source R1. His company's innovations challenge the perceived AI leadership gap and offer cost-effective, transparent LLM development.
DeepSeek Introduces DualPath Architecture to Optimize AI Agent Inference Throughput
DeepSeek introduces DualPath, a new inference framework designed to boost AI agent performance by tackling GPU idling caused by I/O bottlenecks. This dual-path KV-Cache loading mechanism significantly improves throughput in multi-turn agentic scenarios, with reported increases of up to 1.96 times in online settings.
Sequoia Predicts AGI by 2026, Redefining Artificial General Intelligence as Functional Execution
Sequoia predicts Artificial General Intelligence (AGI) will be achieved by 2026, driven by advancements in long-horizon agents. The firm redefines AGI as the ability to "get things done," highlighting agents' capacity for complex, multi-step execution. This rapid progress, dubbed "Moore's Law for agents," suggests AI could perform human expert-level tasks within years.
AI-Powered Prompt Aims to Uncover Hidden Talents Through Deep Self-Reflection
A new AI prompt, "Deep Talent Miner," helps individuals discover hidden talents using principles from psychology and career development. It engages users in deep, multi-round conversations to move beyond conventional assessments and reveal overlooked abilities. The AI then synthesizes the information into a personalized "Talent Instruction Manual."
OpenRouter's Anonymous "Pony Alpha" Model Sparks Global Speculation on Its Origin
OpenRouter's anonymous "Pony Alpha" model has sparked widespread speculation due to its impressive performance in programming, reasoning, and its 200K context window. The tech community is actively debating its developer, with theories pointing to major AI labs like Anthropic, DeepSeek, xAI, and Z.ai.