New Fields Medal laureate Wang Hong, as it turned out, published at the NeurIPS 2019 conference. Her paper is dedicated to low-rank matrix approximation and uses the mathematical apparatus of interpolation theory, which is recognized as a rare example of interdisciplinary research for an AI conference.
NeurIPS
The boom in data centers for AI generates about 2.5 million tons of electronic waste annually, equivalent to 250 Eiffel Towers. Toxic metals and chemicals pollute the environment, and waste collectors, especially in developing countries, suffer from diseases, but their voices are not considered in decision-making.
Microsoft
Google/DeepMind
Oracle
OpenAI
Anthropic
Researchers at the Okinawa Institute of Science and Technology gave virtual robots curiosity, enabling them to acquire language twice as fast. In simulations, robots receiving an internal reward for novelty began playing with objects without prompts, accelerating learning. Results were published in the journal Science Advances.
ByteDance has announced Seeduplex, a new voice AI model that, according to the developers, puts an end to the turn-based approach in voice interactions. The system is capable of simultaneously processing speech, understanding intent, and generating responses in real-time, enabling natural dialogue without pauses.
ByteDance
Chinese company ZTE has revealed new details about its development of the 'world's first smartphone with an AI agent.' The device will integrate artificial intelligence at the operating system level, enabling complex tasks. The new product is expected to hit the market in the coming months.
ByteDance
An official from the former US President Donald Trump's administration accused Chinese company Moonshot AI of stealing intellectual property from American company Anthropic. It is reported that Moonshot AI improperly used Anthropic's technology to develop its Kimi K3 model.
Moonshot AI
Anthropic
Anthropic has released a new feature for its AI model Claude that allows it to learn workflows by observing user actions on the screen. This expands automation capabilities.
Anthropic
Moonshot AI (月之暗面) has released its new open-source large language model, Kimi K3. The model achieves impressive results on international benchmarks: it ranks first on LiveCodeBench (programming tasks), surpassing GPT-4o and Claude 3.5 Sonnet, and also demonstrates high scores on AIME 2025 (mathematics) and MMLU-Pro (general knowledge).
Moonshot AI
The US government announced a $5 billion allocation for federal artificial intelligence research. The funds will support a variety of AI projects.
Tencent Vice President Lin Songtao stated that the AI agent Marvis will not compete in versatility, but will focus on system operations on devices. Marvis has already shown impressive results: DAU exceeded 300,000 within two days, with a seven-day retention rate of 54%. Unlike other agents, Marvis operates on the device side, completing tasks faster and more efficiently through local processing.
Tencent
RUMBA (Russian User Memory Benchmark), the first Russian-language benchmark for evaluating long-term memory in multi-session dialogues, has been introduced. It includes 85 dialogues and 1,543 questions testing information retrieval, reasoning, and the ability to refrain from answering. The benchmark accounts for temporal context and enables diagnosis of bottlenecks in memory architecture.
Сбербанк
The author, a product manager of a billing platform at Selectel, reflects on how graphical user interfaces (GUIs) are not always convenient, and that Model Context Protocol (MCP) and AI agents could become a new way to interact with products, allowing users to formulate intent instead of learning the interface. He shares an experiment: he created an MCP server for Selectel billing and tested it with the Hermes agent.
Selectel
Thomas Ptacek believes that an open-weight model from 2025, equipped with pentesting tools, could escape its sandbox and compromise most networks. In his view, the surprise stems only from us attributing more robust sandboxes to OpenAI.
OpenAI
Moonshot AI
Anthropic
Hugging Face
Liang Wenfeng, founder of DeepSeek, shared his views in an interview on where he believes the main differences lie between China and the US in the field of artificial intelligence. He pointed to a technological gap, particularly in fundamental research and chip development.
DeepSeek
At a recent funding meeting, DeepSeek founder Liang Wenfeng elaborated on the technical and business aspects of large language model development, emphasizing the importance of fundamental research and long-term strategy.
DeepSeek
Flova.ai founder Guo Le, previously the creator of FaceU and CapCut, shared insights on designing products for both humans and AI agents. Flova's team abandoned the traditional canvas model in favor of a flexible workspace with three zones: dialogue, storyboard, and preview. The company has raised $80 million from Sequoia China, IDG, and Cloud9, planning to operate at zero margin during the growth phase.
Moonshot AI
Flova
Tencent launches a new AI service based on its own Hunyuan model. The company also announced a development strategy in the field of generative AI.
Tencent
US officials allege that Chinese company Moonshot AI (known for developing the Kimi model) stole a model from Anthropic and also gained access to export-restricted Nvidia chips.
Moonshot AI
Anthropic
NVIDIA
Tencent is merging teams developing large language models (LLMs) and multimodal technologies under the Hunyuan project, placing them under the unified management of Yao Shunyu. The restructuring aims to optimize resources and accelerate AI development.
Tencent
Tencent has merged its teams working on multimodal and large language models into a single structure. The reorganization aims to strengthen synergy between the areas.
Tencent