腾讯研究院AI速递 20260104

Group 1 - DeepSeek team released a new paper titled "Manifold-Constrained Hyper-Connections," co-authored by founder Liang Wenfeng, proposing the mHC scheme to stabilize large model training and enhance scalability [1] - The mHC scheme projects the residual mapping matrix onto a double-random matrix manifold space, preserving topological expressiveness while restoring the identity mapping property, controlling the signal amplification factor from 3000 to 1.6 [1] - Experiments with a 27B model show that mHC outperforms traditional HC across tasks like BBH and DROP, introducing only a 6.7% training time overhead, with a maximum improvement of 2.3 percentage points [1] Group 2 - Claude Code, launched 6 months ago, generated nearly $1 billion in annualized revenue, with project lead Boris Cherny confirming that 100% of code was completed by Claude Code in the past 30 days [2] - Key configurations include running 5 Claude instances in parallel on terminals, 5-10 Claude instances on the web, utilizing the Opus 4.5 model, and team collaboration through CLAUDE.md files integrated via GitHub actions [2] - Important techniques involve planning mode, slash command encapsulation for workflows, sub-agent handling of repetitive tasks, and PostToolUse hook for code formatting, with feedback loops for Claude to validate its work [2] Group 3 - Tesla's FSD V14.2 successfully completed a cross-country drive from Los Angeles to South Carolina in a 2025 Model 3, covering 2732.4 miles with zero human intervention, including parking and charging [3] - FSD V14.2 or pre-installed Grok shows significant enhancements in driving performance, perception capabilities, and decision-making logic, handling complex intersections and lane changes more decisively, resulting in a more human-like driving rhythm [3] - Tesla's end-to-end architecture contrasts with Waymo's modular approach, as demonstrated by a power outage in San Francisco that disrupted Waymo's operations, while Tesla's FSD remained largely unaffected [3] Group 4 - OpenAI is developing its first AI hardware, potentially a pen-shaped device or portable audio device, codenamed "Gumdrop," which integrates a microphone and camera to convert handwritten notes into text for ChatGPT [4] - The device is similar in size to an iPod Shuffle and aims to become the "third core device" following the iPhone and MacBook, initially planned for production by Luxshare Precision, later shifted to Foxconn, with manufacturing expected in Vietnam or the US [4] - OpenAI is also working on a new audio model architecture set to launch in Q1 2026, promising more natural emotional voices, more accurate and in-depth responses, and improved interruption handling capabilities [4] Group 5 - TSMC's N2 technology is set to enter mass production in Q4 2025, utilizing the first-generation nanosheet transistor (GAA) technology, achieving a 10%-15% performance improvement at the same power level compared to N3E, and a 25%-30% reduction in power consumption at the same speed [6] - The N2 process employs gate-all-around nanosheet transistors that wrap around the current channel, combined with SHPMIM capacitors, resulting in approximately a 20% increase in transistor density and over a 2x increase in capacitance density compared to N3E [6] - TSMC is expanding production simultaneously at its Kaohsiung and Hsinchu fabs, catering to both mobile and AI/HPC chip markets, with N2P and A16 expected to enter mass production in the second half of 2026 [6] Group 6 - Zhiyuan announced the launch of a "small-sized full-body force-controlled humanoid robot," named Q1, standing approximately 0.8 meters tall and capable of fitting into a 30-35L backpack, utilizing innovative materials and control algorithms to shrink QDD joints to "smaller than an egg" while maintaining full-size force control performance [7] - The Q1 robot employs advanced composite material technology for durability and is only 1/8 the size and weight of full-sized robots, with an open-source SDK and HDK supporting 3D printing for custom appearances [7] - It features the "Zhiyuan Lingxin" AI platform for natural conversation and encyclopedic Q&A, and through the "Zhiyuan Lingchuang" platform, users can arrange actions and logic like building blocks, positioning it as a desktop robot for individual creators [7] Group 7 - Elon Musk announced that Neuralink will begin large-scale production of brain-machine interface devices in 2026, transitioning to a streamlined, nearly fully automated surgical process, with electrode wires passing through the dura mater without the need for removal [8] - The new minimally invasive technology reduces costs, lowers risks, and shortens recovery times, making standardization more accessible; as of September 2025, Neuralink had served only 12 patients, increasing to 20 by December [8] - Founded in 2016, Neuralink focuses on treating neurological disorders such as paralysis, muscular atrophy, and Parkinson's disease, with the first patient, Noland Arbaugh, able to post and play games using only the brain chip post-surgery [8] Group 8 - Meta faced criticism from Turing Award winner LeCun after his departure, alleging that Llama 4's testing results were manipulated by using different models on various benchmarks to achieve better scores, leading to a loss of confidence from Zuckerberg in the original AI team [9] - LeCun criticized his 28-year-old supervisor, Alexandr Wang, for lacking research experience and understanding of research methodologies, asserting that Meta's hiring practices have led to a team overly influenced by large language models [9] - LeCun has founded AMI Labs, focusing on world models, with plans to release a "baby-level" model with preliminary physical intuition within 12 months, emphasizing the need for models to understand the physical world's operations rather than relying solely on language [9]