AI Daily – 2026-01-04(Evening)
Semiconductor Giants’ Showdown at CES 2026: AI Chips Enter a New Era of Performance and Efficiency: As CES 2026 approaches, the three giants—NVIDIA, AMD, and Intel—are collectively “unveiling their we
AI Daily – 2026-01-04(Morning)
MIT Releases Recursive Language Model (RLM): Breaking the 10 Million Token Processing Limit: Researchers at the Massachusetts Institute of Technology (MIT) have proposed a new paradigm of Recursive La
AI Daily – 2026-01-03(Evening)
DeepSeek Releases mHC Technology: Manifold-Constrained Hyper-Connections Reshape Model Architecture: DeepSeek’s recent release of the mHC (Manifold-Constrained Hyper-Connections) paper has sent shockw
AI Daily – 2026-01-03(Morning)
Meta Llama 4 Benchmark Manipulation Confirmed: LeCun’s Departure Reveals Inner Workings: Turing Award winner Yann LeCun, upon leaving Meta, publicly admitted that the base model test results for Llama
AI Daily – 2026-01-02(Evening)
DeepSeek Releases mHC Architecture, Attempting to Reconstruct Transformer Residual Connections: DeepSeek published the paper “mHC: Manifold-Constrained Hyper-Connections,” proposing a manifold-constra
AI Daily – 2026-01-02(Morning)
DeepSeek Releases mHC Architecture to Improve Transformer Residual Paths: DeepSeek has published research on “Manifold-constrained Hyper-connection” (mHC), aimed at addressing the limitations of resid
AI Daily – 2026-01-01(Evening)
DeepSeek Releases mHC Paper: Restructuring Residual Connections for Large-Scale Training Stability: The DeepSeek team has released research on Manifold-constrained Hyper-connection (mHC), aiming to so
AI Daily – 2026-01-01(Morning)
South Korea Launches $140 Million “Sovereign AI” Initiative to Build Local Ecosystem: South Korea’s Ministry of Science, in collaboration with five giants including SKT, LG, and Naver, is investing ap
AI Daily – 2025-12-31(Evening)
DeepSeek R1 Raid and the Reinforcement Learning Paradigm Shift: The open-sourcing of DeepSeek-R1 marks a direct impact of Chinese AI forces on Silicon Valley. The model achieved reasoning performance