AI News 2025-12-26
AI Daily Brief
Summary
Lenovo Group plans to launch a cross-device AI super-agent to compete with ByteDance's Doubao.
OpenAI partnered with Broadcom to design custom AI chips, with planned deployment in 2026.
Studies found ChatGPT and other models overestimate human rationality, leading to inaccurate predictions.
OpenAI tested new Skills feature, building modular AI capability units.
Alibaba Cloud open-sourced an image editing model that solves target position drift issues.
Cursor CEO warned about hidden dangers of over-reliance on AI-generated code.
DeepSeek's interaction with Yuanbao attracted attention, with significant user scale growth.
DeepSeek became users' 24/7 personal assistant, with Yuanbao team continuously optimizing and launching DeepSeek+ enhanced features.
AI assistants are deeply integrating into daily life.
Today’s AI News
Lenovo Group plans to launch a strategic global application called “AI Super Agent” during CES, aiming to compete with ByteDance’s “Doubao.” The agent is defined as a system-level application that can connect different AI capabilities and integrate Motorola phones, Lenovo PCs, tablets, and wearables to achieve seamless cross-device connectivity and deep learning of user habits. With its hardware foundation, Lenovo’s agent has deeper system-level integration permissions, potentially offering a more comprehensive experience in cross-device task scheduling.
OpenAI partnered with Broadcom to jointly design custom AI inference chips, with planned first deployment in data centers in second half of 2026, deploying up to 10 gigawatts of computing power over the next five years. The chip uses TSMC’s 3nm process, specifically optimized for “o1” series inference models and future GPT versions, aiming to improve energy efficiency and reduce dependence on general-purpose GPUs. This is part of OpenAI’s vertical integration strategy to strengthen competitiveness and drive AI hardware toward customization.
Latest research found that top AI models like ChatGPT and Claude have a tendency to “over-mythologize” human rationality, assuming human decisions possess extremely high logic and strategic nature, which is not true in reality. In game theory experiments like the “Keynesian Beauty Contest,” AI makes inaccurate predictions because it overestimates opponents’ logical depth. The study indicates that calibrating AI’s understanding of human irrational behavior is a key challenge it faces in fields like economic prediction that require precise modeling of human behavior.
According to disclosures, OpenAI is internally testing a new feature called “Skills” codenamed “Hazelnut,” expected to launch around January 2026: The feature aims to build modular, executable, and cross-platform reusable “capability units.” Unlike existing GPTs, Skills are organized as folders that can contain instructions, code, etc., and can be automatically recognized and combined/called by AI during tasks. This is seen as OpenAI’s key step in responding to AI Agent ecosystem competition and transitioning to a “programmable intelligent platform.”
Alibaba Cloud’s Tongyi Lab officially released the latest image editing model Qwen-Image-Edit-2511: This version focuses on solving the issue in previous models where target object positions experienced “slight drift” after image editing, significantly improving consistency and visual stability before and after editing through technical optimization, providing developers with a more precise and controllable image editing tool.
Michael Truell, CEO of popular AI programming assistant Cursor, issued a warning about the “ambient coding” concept popular among developers: He pointed out that blindly relying on AI for end-to-end code generation without examining underlying logic is like building a house without checking the foundation, which will sow long-term hidden dangers for software engineering. He emphasized true value lies in deep human-AI collaboration; developers should use AI to assist in diving deep into code details rather than giving up control.
In today’s rapidly evolving tech landscape, AI assistants have gradually become an indispensable part of daily life: On December 24, DeepSeek officially interacted with Yuanbao on Xiaohongshu, liking and responding to Yuanbao’s annual report—this rare public interaction attracted widespread industry attention. According to the “Yuanbao × DeepSeek Annual Report,” since integrating DeepSeek in February this year, Yuanbao’s user scale has grown steadily and reached a usage peak on December 14, with growth exceeding 100×, making Yuanbao firmly one of the top three domestic native AI apps.
DeepSeek in Yuanbao is not just a Q&A tool but more like a thinking partner for in-depth communication: In user conversations with DeepSeek, over one-third are continuous multi-turn questions围绕 the same topic, showing user dependency on and trust in AI. In user scenarios, DeepSeek has become a 24/7 “personal assistant”—accompanying work during the day and used for chatting and relaxation at night. Meanwhile, in WeChat, Yuanbao is always on standby to meet users’ various needs.
To enhance user experience, Yuanbao team continuously optimizes DeepSeek technology, improving response efficiency and speed, and launched the “DeepSeek+” strategy, using self-developed large models to supplement multimodal capabilities, ensuring AI assistants can provide more comprehensive information and services. Now, when users use Yuanbao, they can not only get official DeepSeek support but also enjoy additional features like image processing, making phone calls, and document understanding.
Users can @Yuanbao in comment sections during social activities without switching apps, allowing AI assistants to seamlessly integrate into daily life: As more people rely on AI assistants, the “mutual pursuit” between DeepSeek and Yuanbao not only drives AI technology progress but also makes AI gradually become a good helper for people’s work and life.