deepseek
4 articles
InferenceX v2: NVIDIA Blackwell's Benchmark Massacre and AMD's Software Debt
SemiAnalysis benchmarked ~1,000 GPUs across NVIDIA and AMD lineups. GB300 NVL72 hits 100x over H100 — Jensen's 30x was an underestimate. AMD FP8 competes, but FP4+disagg+wideEP combo falls apart in software.
DeepSeek-R1 Grew Its Own Internal Debate Club — Nobody Asked It To
DeepSeek-R1 developed internal multi-agent debates through pure RL training — no one taught it to. Google researchers call this the 'Society of Thought.' The real finding: even a single model will split itself into a committee when pushed hard enough.
Anthropic Goes Public: DeepSeek, Kimi, and MiniMax Used 24,000 Fake Accounts to Steal Claude's Capabilities — 16 Million Conversations of Industrial-Scale IP Theft
Anthropic publicly accuses DeepSeek, Moonshot/Kimi, and MiniMax of using 24,000 fake accounts for 16M+ conversations to distill Claude's coding and agentic reasoning. MiniMax was caught mid-operation pivoting to Claude's latest model within 24 hours.
Andrew Ng: "America First" Is Accidentally Strengthening Global AI — What Is Sovereign AI and Why Should Taiwan Care?
Andrew Ng reports from Davos WEF that US export controls and America First policies are pushing countries toward Sovereign AI. DeepSeek and Qwen adoption is rising globally. For Taiwan, the question is sharp: you make the AI chips, but do you own AI sovereignty?