腾讯研究院·2025-12-18 16:01Group 1 - Google is advancing the "TorchTPU" strategy to enable PyTorch to run smoothly on TPU chips, aiming to eliminate migration barriers for developers and considering partial open-sourcing of the software [1] - Google is negotiating a collaboration with Meta to provide Meta with more TPU access, allowing Meta to reduce inference costs and dependence on NVIDIA by adapting software for TPU [1] - Wall Street analysts believe that CUDA is NVIDIA's strongest defense, and Google's previous reliance on its internal Jax framework has widened the gap with external customer usage habits [1] Group 2 - The ChatGPT app store has officially launched, categorizing applications like Adobe Photoshop and Canva, with users triggering them via "@app name" [2] - Developers can submit applications for review on the OpenAI developer platform, which offers a comprehensive resource system including best practice guides and open-source sample applications [2] - OpenAI plans to raise new funding at an estimated valuation of around $750 billion, potentially reaching $1 trillion, attempting to replicate the Apple App Store model in the AI era [2] Group 3 - Google has released the Gemini 3 Flash model, achieving a score of 33.7% on the Humanity's Last Exam benchmark, while Gemini 3 Pro scored 37.5% and GPT-5.2 scored 34.5% [3] - This model maintains the Flash series' extreme native speed, outperforming Gemini 2.5 Pro while tripling the speed, priced at $0.50 per million tokens for input and $3 for output [3] - Gemini 3 Flash is now the default model for Gemini applications and search AI modes, with response times generally under one second, available globally through Google AI Studio and Vertex AI [3] Group 4 - ByteDance has launched the universal Agent model Seed1.8, which integrates search, code, and GUI Agent capabilities, automatically adjusting processing methods based on task complexity [4] - In GUI Agent evaluations, Seed1.8 surpassed Seed1.5-VL, demonstrating reliability in multi-step tasks across computer, web, and mobile environments, scoring 67.6 on the BrowseCompen benchmark [4] - The model achieved a top score of 11.0 on ZeroBench and 87.8 on VideoMME for long video understanding, incorporating the "VideoCut" video tool [4] Group 5 - The Step-GUI cloud model has been fully upgraded, supporting over 200 task scenarios and usable across mobile, PC, and automotive platforms, with deployment of an "AI phone" possible in as little as 10 minutes [5][6] - This model features longer reasoning steps, enhanced semantic understanding, and generalization capabilities, autonomously asking questions when user instructions are vague [6] - The GUI-MCP protocol is open for end-cloud collaboration, with APIs temporarily available for free, and a call for users to create showcases and develop applications [6] Group 6 - xAI has officially released the Grok Voice Agent API, making its real-time voice capabilities available to developers for voice-first application scenarios [7] - The API includes various built-in voices and companion personalities, allowing developers to finely control system commands and behavior parameters [7] - It supports real-time voice recognition and synthesis with a streaming audio design, enabling search capabilities during conversations and significantly reducing interaction latency [7] Group 7 - Apple is reportedly abandoning its VR headset project in favor of developing AI smart glasses, with a projected launch in late 2026 or 2027 [8] - The company has paused its AR/VR headset initiatives and plans to reintroduce the iMac Pro, which has been off the market for over four years, potentially featuring the M5 Max chip [8] - A 20th-anniversary edition iPhone is expected in 2027, featuring a curved design that wraps around the device edges and a front camera positioned under the display [8] Group 8 - a16z partners assert that the AI bubble has not yet burst, as it has not reached a point where investments are wasted [9] - They believe that if companies cease developing larger models and rely solely on existing models, they could quickly achieve profitability at current profit margins [9] - Predictions indicate that GDP could grow by several percentage points by 2030, with a reasonable lower limit of 30% growth if AGI is achieved, though outcomes could vary widely [9]