What’s New in AI? GPT‑5, SWE‑Swiss, Agentic Web, and More This Week

OpenAI GPT‑5

On August 8, OpenAI released GPT‑5, a flagship model that combines non‑reasoning and reasoning capabilities, supports “on‑demand thinking,” and offers four personality modes. It claims doctoral‑level intelligence, can perform tasks such as planning, invitation drafting, and procurement, and outperforms previous OpenAI models on benchmarks in mathematics, coding, visual perception, and health, achieving a GPQA score of 88.4% without tools.

SWE‑Swiss Model

Researchers from Peking University, ByteDance, and Hong Kong University introduced SWE‑Swiss, a 32‑billion‑parameter model designed to fix software bugs. Trained on the SWE‑bench Verified benchmark, it reaches 60.2% accuracy, setting a new state‑of‑the‑art for medium‑size models. The work demonstrates that a well‑designed training recipe can enable 32B models to match larger counterparts, and the model and data are released open‑source.

Pinduoduo AI Talent Recruitment

Pinduoduo has accelerated hiring for AI roles, posting positions for large‑model inference engineers, CUDA/GPU developers, multimodal algorithm engineers, and NLP engineers. The company aims to build domain‑specific large models for e‑commerce, improve product understanding, attribute recognition, and AIGC applications, reflecting a strong push into vertical AI development.

Agentic Web

The Agentic Web concept, proposed by researchers from UC Berkeley, UCL, Shanghai Jiao‑Tong University, and Shanghai Chuangzhi Institute, envisions an internet where AI agents act as primary operators. Users issue high‑level goals in natural language, and autonomous agents plan, search, invoke services, and coordinate to accomplish tasks, shifting the web from passive information display to proactive task execution.

Google Genie 3 World Model

Google DeepMind announced Genie 3, a highly realistic world model capable of generating interactive, playable environments at 24 fps in 720p. It advances over previous generations with improved physics simulation, ecological dynamics, and long‑term scene consistency, highlighting breakthroughs in real‑time generative modeling.

Multimodal AI for Railway Design

SenseTime and China Railway First Survey & Design Institute launched a multimodal large‑model platform for railway engineering. The system integrates text, drawings, and formulas to provide knowledge retrieval, intelligent Q&A, text generation, and multimodal dialogue, achieving over 90% accuracy on 28 000 test queries.

DJI ROMO Robot Vacuum

DJI released its first robot vacuum, the ROMO series (S, A, P), featuring dual‑eye cameras, solid‑state LiDAR, 25 000 Pa suction, and a transparent charging base for the P version. While hardware is impressive, software performance is still maturing.

AI‑Enhanced Smart Glasses

New XR research introduced “Reality Proxy,” a digital proxy that lets users select real‑world objects instantly via hand gestures, improving interaction reliability in mixed‑reality environments.

Humanoid Occupancy Perception System

Beijing’s Humanoid Robot Innovation Center unveiled Humanoid Occupancy, a multimodal perception system that fuses sensor data into a semantic occupancy map, enabling robust environment understanding for humanoid robots across manipulation, locomotion, and navigation tasks.