AI 新闻 —— July 29, 2026
PulseAugur 当天浮现的 20 条头条故事 —— 综合实验室、论文及开发者社区的信号进行排序。
-
Value Generalisation: A New Approach to AI Alignment
Researchers are proposing a new approach to AI alignment called "value generalisation," which aims to create AIs that can reliably extend human values and preferences to novel situations. This capability is seen as a critical missing piece in current AI systems, which often fail…
-
BYD unveils HyWorldVLA, entering autonomous driving foundation models
BYD has unveiled its AI team and a new hybrid world model called HyWorldVLA. This model utilizes a combination of pixel-latent world modeling and a vision-language-action (VLA) architecture. HyWorldVLA has achieved state-of-the-art performance on the NAVSIM v1 benchmark, scoring…
-
Anthropic's Claude Opus 5 launches on DigitalOcean AI Inference Cloud
Anthropic's latest flagship model, Claude Opus 5, is now available on DigitalOcean's AI Inference Cloud. This model is designed for intensive agent workflows, offering advanced capabilities in software engineering, scientific research, and enterprise automation. Claude Opus 5 pr…
-
Moonshot AI releases 2.8T parameter Kimi K3 model, surprising Silicon Valley
Moonshot AI has officially open-sourced its Kimi K3 model, boasting a massive 2.8 trillion parameters. This release is expected to significantly impact the global AI landscape, potentially surprising major players in Silicon Valley. The announcement was made alongside other tech…
-
RMB strengthens against USD; Kimi K3 open-sourced with 2.8T parameters
On July 29th, the central parity rate of the RMB against the US dollar was adjusted upwards by 29 basis points to 6.7899. In market news, the Hang Seng Index opened up 0.7%, with the Hang Seng Tech Index rising 1.02%. The software and services sector led the gains, with MINIMAX …
-
Kimi K3 open-sourced with 2.8T parameters; Claude designer launches new tool
Kimi K3, a new large language model with 2.8 trillion parameters, has been open-sourced, reportedly impressing Silicon Valley giants. Concurrently, a tool designed by a former Claude designer aims to attract a million users, suggesting a competitive landscape driven by rapid inn…
-
Moonshot AI open-sources Kimi K3 with 2.8T parameters, surprising Silicon Valley
Moonshot AI has officially open-sourced its Kimi K3 model, boasting 2.8 trillion parameters. This release is positioned as a significant global offering, with reports suggesting it has surprised major tech players in Silicon Valley. The move by Moonshot AI aims to make advanced …
-
Moonshot AI open-sources Kimi K3 with 2.8T parameters
Moonshot AI has officially open-sourced its Kimi K3 model, boasting a massive 2.8 trillion parameters. This release is positioned as a significant move, aiming to challenge major players in the Silicon Valley AI landscape. The announcement comes alongside other tech news, includ…
-
Moonshot AI releases 2.8T parameter Kimi K3 model, challenging tech giants
Moonshot AI has officially open-sourced its Kimi K3 large language model, boasting a massive 2.8 trillion parameters. This release aims to challenge established tech giants and is expected to significantly impact the AI landscape. The move follows a period of intense development…
-
New AI models improve fall impact detection accuracy and efficiency · 2 sources tracked
Two new research papers propose advanced methods for detecting the precise moment of impact during falls, a critical factor for timely medical intervention. The first paper utilizes Spatio-Temporal Graph Convolutional Networks (STGCN) combined with Gated Recurrent Unit (GRU) and…
-
AI models show mixed results in assisting scientific research, study finds · 2 sources tracked
A new two-part study published on arXiv explores the capabilities of large language models (LLMs) in assisting scientific research. The first paper details how mid-2025 models like ChatGPT, Claude, and DeepSeek performed in generating project plans and evaluating proposals, find…
-
New benchmarks and methods advance drone-view geo-localization across conditions
Researchers have introduced new methods for drone-view geo-localization, a task that involves identifying a drone's location using its imagery. The first approach, MASTR-Net, addresses the challenge of varying illumination conditions by creating a unified benchmark called IRCHN.…
-
New research tackles evaluation and architecture for masked diffusion language models
Two new research papers introduce novel evaluation protocols and architectures for masked diffusion language models (MDLMs). The first paper, "CaRE," proposes a compute-aware framework to standardize evaluations, revealing that factors like temperature and step counts significan…
-
AI governance frameworks risk failure in public sector with rise of GPAI
Two new arXiv papers published on July 28, 2026, highlight significant challenges in applying existing AI governance frameworks to public sector organizations, particularly with the rise of general-purpose AI (GPAI). The first paper proposes a typology of AI-driven cyber governa…
-
Anthropic's Claude Opus 5 API errors linked to disabled thinking and effort settings
Anthropic's Claude Opus 5, launched on July 24, 2026, has specific API requirements that can lead to HTTP 400 errors if not met. Specifically, disabling the 'thinking' parameter while requesting 'xhigh' or 'max' effort levels will cause a failure. Developers migrating to Opus 5 …
-
AI leaders urge US government to control AI development pace; Kimi AI raises $3.5B
Over 1,100 employees and executives from leading AI companies, including Anthropic and OpenAI, have signed an open letter urging the U.S. government to regulate the pace of advanced AI development. The letter, supported by figures like Anthropic CEO Dario Amodei and OpenAI Chief…
-
Anthropic's Mythos AI finds crypto weaknesses in NIST-tested algorithms
Anthropic's AI model, Mythos, has identified previously unknown weaknesses in two cryptographic algorithms. While these findings do not compromise current security systems, they offer methods to reduce the effort required to break them. The identified vulnerabilities were found …
-
Dual-engine architecture for embedding systems proposed
This article discusses the architectural design of embedding systems, proposing a dual-engine approach. It suggests using a lean engine for steady embedding loads and a faster engine for handling burst loads. The core challenge highlighted is ensuring consistency in vector outpu…
-
Google launches Lyria 3.5 music model with selective track editing
Google has launched Lyria 3.5, an updated music generation model integrated into its Flow Music platform. This new version allows for the creation of tracks ranging from 30 seconds to 3 minutes in length. A key feature, "Selective Section Painting," enables users to modify speci…
-
Tencent Hunyuan open-sources AngelSpec framework and Hy3-A21B weights
Tencent's Hunyuan team has officially open-sourced AngelSpec, a speculative decoding framework designed for drafter training, architecture design, and online deployment. Alongside AngelSpec, they have also released the MTP and DFly drafter weights and training code for Hy3-A21B.…