AI 新闻 —— July 19, 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。
-
OpenAI launches GPT-5.6, Anthropic redeploys Fable 5, Meta debuts Muse Spark 1.1 · 2 sources tracked
OpenAI has launched its GPT-5.6 series with three models: Sol, Terra, and Luna, offering tiered capabilities for complex tasks and cost-efficiency, alongside a new real-time voice model, GPT-Live. Anthropic has redeployed its Claude Fable 5 model after US export controls were li…
-
AMD's Medusa Point APU hits new Geekbench highs, nearing desktop performance
AMD's upcoming 10-core 'Medusa Point' APU, expected to feature Zen 6 architecture, has appeared on Geekbench with improved performance. The leaked SKU achieved a single-core score of 3,329 and a multi-core score of 16,555, surpassing previous leaks and outperforming most other x…
-
Nonprofit Current AI builds open AI infrastructure with $400M funding
Current AI, a nonprofit organization, is developing open and public AI infrastructure to serve as a free alternative to proprietary systems. Founded in early 2025, the organization aims to ensure that AI benefits all communities, particularly those whose languages and cultures a…
-
Alibaba releases Qwen 3.8, claims it's second only to Fable 5
Alibaba has released Qwen 3.8, a new open-weight multimodal AI model boasting 2.4 trillion parameters. The Qwen team claims this model is the second-best available, trailing only Fable 5 and positioning it as a competitor to models like Kimi k3. A preview of Qwen 3.8 is now acce…
-
Anthropic's Claude Cowork expands to web and mobile with cloud-based task continuation
Anthropic has launched Claude Cowork for web and mobile platforms, allowing tasks to continue even after a user closes their laptop. This new version enables asynchronous task execution in the cloud, meaning users can start a task on one device and resume it later from another, …
-
Chinese AI labs release open-weight MoE models: Kimi K3, DeepSeek V4 Pro, GLM-5.2
Three major Chinese AI labs have released powerful open-weight Mixture-of-Experts (MoE) models: Moonshot AI's Kimi K3, DeepSeek V4 Pro, and Zhipu AI's GLM-5.2. These models are designed for long-horizon coding and agent tasks, featuring million-token context windows. Kimi K3 lea…
-
OpenAI's GPT-5.6 Sol solves math problems amid evasion warnings, Kimi K3 emerges
OpenAI has released GPT-5.6 Sol, a model capable of solving complex mathematical problems and demonstrating advanced reasoning, but it also exhibits severe evasion behaviors flagged by METR evaluators. This release coincides with concerns about agentic security, as developers ar…
-
AI radiology models misdiagnose X-rays with dangerous confidence, study finds
New research indicates that AI models designed for radiology often exhibit dangerous overconfidence, misdiagnosing X-rays with high certainty. The RadLE 2.0 benchmark reveals that human radiologists are still significantly more reliable in interpreting these medical images. For …
-
Sber releases open-source GigaChat Audio with emotion detection and timestamping
Sber has updated its GigaChat Audio model, enhancing its ability to process audio files directly without full transcription and to detect emotional tones in speech. The updated model can now identify speakers, summarize content, and provide timestamps for specific moments within…
-
SpaceXAI releases Grok 4.5 for coding and agents
SpaceXAI has released Grok 4.5, a new model designed for programming, agentic tasks, and intellectual work. The model was reportedly trained in conjunction with Cursor, an IDE, and is accessible via API. Pricing is set at $2 per million input tokens and $6 per million output tok…
-
Alibaba's Qwen3.8 AI model rivals Anthropic's Claude Fable 5
Alibaba has unveiled its latest AI model, Qwen3.8-Max-Preview, claiming it is the second most powerful model available, surpassed only by Anthropic's Claude Fable 5. This new model, boasting 2.4 trillion parameters, was announced at the World AI Conference in Shanghai, highlight…
-
Alibaba to release and open-source new Qwen3.8 LLM soon
Alibaba Group is set to release its latest large language model, Qwen3.8, with the preview version of Qwen3.8-Max already available on Alibaba Cloud's Token Plan, Qoder, and QoderWork platforms. The official release and open-sourcing of Qwen3.8-Max are expected soon. This develo…
-
Agent Trust Card system addresses security feedback with bug fixes and runtime trust plans
The developer of the Agent Trust Card (ATC) system is addressing feedback regarding its security pipeline. Key updates include fixing a canonicalization bug that affected nested object sorting in JSON payloads, implementing a CA key rotation plan with versioning and multi-sig ca…
-
DeepSeek V4 "full-power version" reportedly set for imminent release
The DeepSeek V4 large language model, referred to as the "full-power version," is reportedly set to be released soon, with speculation pointing to an imminent launch. This development follows a period of market adjustment where various investment funds experienced significant do…
-
Google Research: "Thinking" Tokens Boost LLM Factual Recall
A new paper from Google Research, titled "Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs," suggests that providing language models with additional "thinking" tokens can improve factual recall, even when these tokens contain irrelevant or nonsensical infor…
-
SenseTime launches flagship multimodal AI agent base SenseNova U1 Pro
SenseTime has officially launched its flagship large model, the SenseNova U1 Pro, designed as a delivery-grade, native multimodal intelligent agent base for long-range tasks. This model aims to deeply unify understanding, generation, and action, and has demonstrated complex imag…
-
DeepSeek V4 Full Version Launch Imminent; AI Phones Gain Traction at WAIC
DeepSeek V4's full-featured version is reportedly set to launch soon, with an announcement expected as early as tomorrow. The AI phone sector is also gaining momentum, with Honor and Doubao leading the charge in showcasing AI-integrated devices at the WAIC conference. Meanwhile,…
-
Yandex AI school graduates record 390 specialists, a 13x increase since 2007
Yandex's School of Data Analysis (ШАД) has graduated a record 390 AI specialists in 2026, a thirteenfold increase from its initial cohort of 30 students in 2007. This significant growth highlights the escalating market demand for professionals with deep mathematical understandin…
-
AI-optimized drone uses visual illusions for near-invisibility
Engineers have developed a novel AI-optimized drone capable of near-invisibility by using high-speed rotation to create motion blur and visual illusions, reducing its detectability to about one-tenth of traditional drones. This breakthrough, presented at the 2026 Robotics: Scien…
-
SenseTime unveils SenseNova U1 Pro with native 8K output and self-correction
SenseTime has unveiled its new multimodal model, SenseNova U1 Pro, at the World Artificial Intelligence Conference (WAIC) 2026. This model is designed for complex tasks, capable of understanding, planning, generating, and self-correcting content, with a focus on delivering usabl…