AI 新闻 —— April 18, 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。
-
Anthropic updates Claude Opus 4.7 system prompt with new tools and safety features
Anthropic has updated the system prompt for its Claude Opus 4.7 model, introducing changes that refine its behavior and capabilities. The update renames the "developer platform" to "Claude Platform" and adds new integrated tools like "Claude in Chrome," "Claude in Excel," and "C…
-
Luma Labs launches Luma Agents for cinematic video generation
Luma Labs has released Luma Agents, a new text-to-video generation model that allows users to specify cinematic styles. The model enables users to define a desired aesthetic and then generate video content that matches it. This tool aims to provide greater creative control over …
-
Airborne DNA detection offers new insights into biodiversity and ecosystem health
Scientists are increasingly exploring the potential of environmental DNA (eDNA) found in the air as a powerful tool for understanding ecosystems. This airborne DNA, shed from living organisms through various means, can be collected and sequenced to identify species present in an…
-
Researchers propose PU classification method using clustering and logistic model
Researchers have proposed a new algorithm for PU classification that addresses scenarios where the standard SCAR condition is not met. The method utilizes a two-step process: first, it employs a 2-means clustering algorithm to generate initial cleaning labels, and then it perfor…
-
Anthropic launches Claude Design for AI-generated websites
Anthropic has released Claude Design, a new product that generates production-ready websites, slide decks, and one-pagers from natural language prompts. This tool integrates with existing design systems by extracting color palettes, typography, and component patterns from codeba…
-
College instructor uses typewriters to combat AI-written assignments
A Cornell University instructor is using manual typewriters for assignments to combat student reliance on AI and digital tools. The exercise requires students to write without access to spellcheck, delete keys, or online resources, encouraging them to think more deeply and engag…
-
Trump convenes Iran situation room meeting amid renewed Hormuz crisis
Former President Trump convened a situation room meeting to address a renewed crisis in the Strait of Hormuz. The meeting focused on the escalating tensions and potential responses to the situation. Details regarding the specific nature of the crisis or the outcomes of the meeti…
-
Hugging Face releases BasketHAR, a multimodal dataset for basketball activity recognition
Researchers have introduced BasketHAR, a new multimodal dataset designed for human activity recognition specifically within basketball training scenarios. This dataset addresses the limitations of existing HAR datasets, which often focus on general activities rather than special…
-
Iran closes Strait of Hormuz, fires on tankers amid regional tensions
Iran has once again closed the Strait of Hormuz and fired upon tankers in the region. This action has significant geopolitical implications, potentially disrupting global oil supplies and increasing tensions in the Middle East. The specific details of the engagement and the imme…
-
Netanyahu's actions erode US support for Israel, especially among younger Americans
Israeli Prime Minister Benjamin Netanyahu's actions are significantly eroding American public and congressional support for Israel, particularly among younger demographics. This shift is evident in recent voting patterns, with a notable increase in Senate Democrats opposing arms…
-
Scientists observe rare peaceful queen succession in naked mole rat colony
Scientists have observed a rare peaceful succession of power within a captive colony of naked mole rats, a species typically known for aggressive conflicts over leadership. This event, documented over several years, suggests that non-violent transitions are possible under certai…
-
Simon Willison turns Anthropic's Claude system prompts into a git timeline
Simon Willison has developed a method to transform Anthropic's published system prompts for Claude into a git-like timeline. This approach breaks down the monolithic markdown into granular files, each representing a specific model revision with timestamped commits. This allows f…
-
New theory explores Hermite approximations with adaptive coordinate transformations
Researchers have developed new error estimates for approximating functions using Hermite expansions combined with adaptive coordinate transformations. Their analysis demonstrates an equivalence principle where approximating a function in a transformed basis is akin to approximat…
-
Soaring jet fuel costs accelerate airline shakeout, threatening travelers with higher fares
The U.S. airline industry is facing significant disruption due to soaring jet fuel prices, exacerbated by recent geopolitical events. This cost pressure is leading to higher fares for consumers and forcing airlines to re-evaluate their business models, with some already cutting …
-
AI-generated beauty filters create unrealistic standards, impacting industry
AI-generated images are creating unrealistic beauty standards within the beauty industry. These "digital fantasies" are making it harder for consumers to achieve the "perfect" look they see online. The technology is raising concerns about its impact on self-esteem and the percep…
-
Users report Claude 4.7 Opus overly cautious, limiting creative work
Users are reporting that Anthropic's Claude Code Opus 4.7 is exhibiting overly cautious behavior, refusing tasks it deems potentially related to malware or security bypasses, even for legitimate development work. This has led to user frustration, with some feeling controlled by …
-
vLLM releases v0.19.2rc0 with bug fixes for GLM-ASR
vLLM has released version 0.19.2rc0, which includes a bugfix for the k_proj bias in GLM-ASR models. This release is part of the ongoing development and maintenance of the vLLM project, a high-throughput and low-latency inference engine for large language models.
-
LLMs can now recognize demographic differences without causing harm via DART training
Researchers have developed a new training method called DART (Distill-Audit-Repair Training) to address "harm drift" in large language models. This drift occurs when models, in an effort to be safe, become overly cautious and avoid acknowledging factual demographic differences, …
-
Fireworks AI details training-inference parity challenges in MoE models
Fireworks AI has released learnings on achieving Training-Inference Parity in Mixture-of-Experts (MoE) models. The core challenge identified is that floating-point addition is not associative, meaning the order of operations can affect the final result. This technical insight is…
-
Simon Willison uses Claude Code to enhance blog-to-newsletter tool
Simon Willison details how he used Claude Code to update his blog-to-newsletter tool. He provided a prompt to modify the tool's HTML and JavaScript to include a new content type called 'beats', which captures external posts like project releases or museum visits. The prompt also…