腾讯研究院AI速递 20260309

Group 1: Generative AI Developments - OpenAI released the GPT-5.4 series, integrating Computer Use capabilities, combining code, reasoning, and desktop control into a unified model [1] - The OSWorld desktop control evaluation scored 75.0%, surpassing the human benchmark of 72.4%, while GDPval professional work evaluation reached 83.0% [1] - Standard API pricing is set at $2.50 per million inputs and $15 per million outputs, with a Pro version priced at a 12x premium targeting complex agent scenarios [1] Group 2: OpenAI Initiatives - Peter Steinberger, founder of OpenClaw, joined OpenAI and launched the "Codex for Open Source" project, offering free API credits and 6 months of ChatGPT Pro access to open-source maintainers [2] - The application criteria target core maintainers and widely used public project operators, with non-standard projects eligible if they play a significant role in the ecosystem [2] - Steinberger claims to balance responsibilities between OpenAI and OpenClaw, aiming to support as many open-source contributors as possible [2] Group 3: Tencent Innovations - Tencent's Mix Yuan introduced the HY-WU paradigm, generating personalized LoRA parameters in real-time during inference, replacing traditional static fine-tuning methods [3] - This approach was applied to an 800 billion parameter image editing model, outperforming closed-source models in multiple metrics, with only a 0.11 point gap from GPT Image 1.5 [3] - The paradigm is designed for cross-modal applicability, with plans to expand functional memory to video generation, multi-modal alignment, and edge deployment [3] Group 4: Xiaomi's AI Agent - Xiaomi launched the miclaw mobile AI Agent product based on the MiMo model, encapsulating over 50 system-level tools for autonomous task orchestration [4] - The AI can interact with the entire home IoT ecosystem and supports third-party applications through an SDK [4] - It features self-evolution capabilities, allowing it to create sub-agents and continuously adapt based on user preferences and experiences [4] Group 5: Karpathy's Autoresearch - Karpathy released the autoresearch project, consisting of only 630 lines of code, enabling an AI agent to autonomously execute code editing, model training, evaluation, and iteration without human intervention [5] - Each training session lasts 5 minutes, using val_bpb as a unified evaluation metric, with the agent submitting improvements via Git [6] - Karpathy is running an enhanced version on eight H100 GPUs, positioning the project as a proof of concept for self-evolving LLMs, with potential for expansion into various research fields [6] Group 6: Security Innovations - Illia Polosukhin, co-author of the Transformer paper, rewrote OpenClaw in Rust, launching the secure version IronClaw with a four-layer defense architecture [7] - Key security features include WASM sandbox isolation, AES-256-GCM encrypted credential vaults, and a trusted execution environment (TEE) [7] - The project aligns with NEAR Protocol's "user-owned AI" strategy, establishing an AI cloud platform and a marketplace for intelligent agents [7] Group 7: Multiplayer Video World Model - The team led by Xie Sainin introduced Solaris, the first multiplayer video world model capable of generating consistent first-person perspectives among multiple players, validated in Minecraft [8] - They developed the SolarisEngine for data collection, creating a dataset of 12.64 million frames, the first annotated dataset for training multiplayer world models [8] - The model incorporates a multi-player self-attention layer to facilitate information exchange among players, significantly outperforming previous solutions [8] Group 8: AI in Theoretical Physics - Google Research utilized Gemini Deep Think, tree search, and automatic numerical feedback to solve the unresolved problem of cosmic string gravitational radiation power spectrum [9] - The AI explored approximately 600 candidate paths, with 80% pruned by an automatic verifier, ultimately identifying six solutions, with the Gergenbauer method being the most elegant [9] - The final closed-form solution was achieved through human-AI collaboration, showcasing a reusable AI-driven research paradigm [9] Group 9: Labor Market Impact of AI - Anthropic's labor market report indicates that AI is subtly impacting young people's first jobs, with a 14% decrease in the proportion of 22-25-year-olds entering high AI exposure occupations [10] - The AI task coverage for computer programmers reached 74.5%, but actual coverage across industries is only about one-third of theoretical values, indicating significant untapped potential [10] - Companies are shifting investments from "future human assets" to "immediate computational assets," leading to the disappearance of entry-level positions and emphasizing decision-making, aesthetic engineering, and AI collaboration skills as core competencies [11] Group 10: OpenClaw's Global Impact - OpenClaw's global popularity surged, with over 1,300 attendees at a New York gathering, where Huang Renxun described it as "the most important software release in history" [12] - Observations from the event indicated users spending an average of $1,000 to $2,000 monthly on model costs, with some burning 1 billion tokens daily [12] - Security concerns emerged as the primary issue, with no one believing the system is 100% secure, highlighting the genuine demand for personal intelligent agents and marking the onset of the consumer AI agent era [12]