A leading RLHF researcher warns that proposed US anti-distillation legislation would cripple open-source AI startups and academic research, arguing it could set the field back by six to twelve months.
A former White House official says DeepSeek's V4 model is behind and likely using restricted chips. A Silicon Valley CEO argues it's a genuine open-source breakthrough. The battle over China's AI capabilities is heating up again.
DeepSeek V4's technical report reveals training instability after doubling training data to 33T tokens. A DeepMind researcher says this explains the delay, calling the fixes 'band-aids.'
DeepSeek's open-source V4-Pro-Max set a record on Codeforces, beating GPT-5.4, but fell short on knowledge and long-context benchmarks against Gemini and Opus.
DeepSeek's new V4 series makes 1M-token context standard across all tiers, with V4-Pro matching top closed-source models in reasoning and openly acknowledging gaps against Anthropic's Opus 4.6.
DeepSeek open-sources its V4 preview under MIT license: 1.6 trillion parameters, 1M token context, and memory usage slashed to a tenth of V3.2.
DeepSeek open-sources TileKernels, revealing two never-before-seen architecture components — Engram and Manifold HyperConnection — that point to a next-generation model.
DeepSeek is raising 50 billion RMB at a $44 billion valuation, requiring a minimum 5 billion RMB investment per firm — a stark pivot from its self-funded past as talent flees and rivals stockpile capital.
A Princeton student has leaked the full technical specifications for DeepSeek's V4 model, revealing a 1.6-trillion-parameter architecture with 384 experts and a 1-million-token context window.
DeepSeek founder Liang Wenfeng, who once rejected VC money, is now seeking $3 billion at a $100 billion valuation. The company's deep technical reliance on Nvidia has become its biggest obstacle.
