AI 新闻 —— July 14, 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。
-
Bank of England governor calls for global AI threat cooperation
Bank of England Governor Andrew Bailey has urged global cooperation to address the risks posed by advanced AI technologies. He emphasized that no single nation, including the US, can effectively manage these threats alone and highlighted the need for coordinated testing to ensur…
-
ChatGPT returns to WhatsApp in Europe after EU intervention · 2 sources tracked
ChatGPT has been reinstated on WhatsApp in the European Economic Area, following intervention by the European Commission. Meta was compelled to open its messaging platform to third-party AI assistants under threat of significant financial penalties. This move allows users in EU …
-
Anthropic releases Claude Sonnet 5 with enhanced agentic capabilities
Anthropic has released Claude Sonnet 5, an updated mid-tier model that significantly improves agentic capabilities and performance over its predecessor, Sonnet 4.6. This new model demonstrates enhanced abilities in planning, browser and terminal navigation, and autonomous task e…
-
QuiX delivers first universal photonic quantum computer for data centers
QuiX Quantum has announced the delivery of Carina, the first universal photonic quantum computer designed for installation in standard data centers. Unlike most quantum computers that require cryogenic temperatures, Carina operates at room temperature and uses single photons as …
-
Empero AI releases Qwythos-9B-v2, fixing looping with 1M-token context
Empero AI has released Qwythos-9B-v2, an updated version of its large language model designed to eliminate looping and degeneration issues that previously affected a small percentage of its outputs. This new version achieves a 0% looping rate through a technique called Final-Tok…
-
Anthropic ships Claude Opus 4.7, OpenAI counters with GPT-5.5 · 1 source tracked
Anthropic has released Claude Opus 4.7, featuring a significant improvement in agentic coding tasks with a SWE-bench Pro score increase to 64.3% and enhanced high-resolution vision capabilities. OpenAI followed shortly after with GPT-5.5, codenamed "Spud," which boasts a leading…
-
Anthropic's Claude Fable 5 excels at browser-based 3D generation
Anthropic's Claude Fable 5 model has demonstrated impressive capabilities in 3D generation, particularly when used with ThreeJS for browser-based applications. Recent examples showcase its ability to create complex scenes, such as a Minecraft-style world with various features, a…
-
Uber expands into travel, data collection, and financial services
Uber is expanding its services beyond ride-hailing and delivery, incorporating hotel bookings through a partnership with Expedia and "shop for me" features. The company is also developing a data-gathering operation called AV Labs, which develops sensor-equipped vehicles to colle…
-
AI models use 'relocation' in latent space for covert communication
Researchers have explored how AI models can communicate covertly by relocating signals within their latent space, rather than obfuscating them. In experiments using SpikeGPT, a spiking neural network based on the RWKV architecture, a sender (Alice) was able to move message clust…
-
LLM 'hallucination' is three distinct bugs, not one
The term "hallucination" in large language models (LLMs) is being used to describe three distinct issues, leading to confusion in developing solutions. The first type involves factual inaccuracies where missing context could be supplied to correct the model, a problem addressed …
-
OpenAI launches GPT-5.6 with Sol, Terra, Luna tiers for agentic tool calling
OpenAI has released GPT-5.6, a new family of three models: Sol, Terra, and Luna, all optimized for agentic tool calling. Launched on July 9, 2026, these models share a 1 million token context window and a 128K max output, with varying costs and reasoning capabilities. The key in…
-
New brands get zero AI citations without Wikipedia presence, study finds
A study by Kre8on found that new brands, less than a year old, are not cited at all by major AI models like ChatGPT, Claude, Gemini, and Perplexity. The most significant factor for a brand to be cited by these AI models is the presence of a Wikipedia or Wikidata entry. The resea…
-
Prompt vs. Semantic Caching: Optimizing LLM Costs and Latency
This article discusses two primary methods for optimizing large language model (LLM) performance: prompt caching and semantic caching. Prompt caching, or exact-match caching, stores and retrieves responses for identical prompts, offering simplicity and maximum cost savings but f…
-
Agnes AI launches free AI coding model to rival top tier
Agnes AI has released Agnes-2.5-Flash, a new AI coding model that rivals top-tier competitors like Claude Opus 4.7 and is available for free. This model demonstrates strong capabilities in bug detection and fixing, generating complex web applications from scratch, and performing…
-
RAG evaluation framework RAGAs improves AI assistant reliability
This post details the implementation of RAGAs, an evaluation framework for Retrieval Augmented Generation (RAG) systems, to address issues like hallucination and poor answer quality. It highlights three key metrics: faithfulness, which ensures generated answers are supported by …
-
New book recovers source code for pioneering chatbot ELIZA
A new book, "Inventing ELIZA," has recovered the source code for the pioneering chatbot ELIZA, developed by Joseph Weizenbaum at MIT. This recovery challenges conventional accounts of ELIZA by revealing multiple versions and scripts beyond its well-known "DOCTOR" persona. The bo…
-
Jailbroken Gemini AI used to create Russian fraud server in minutes
A jailbroken version of Google's Gemini AI was exploited to create a command-and-control (C2) server for a Russian fraud operation. This server was established in just six minutes, highlighting the potential for malicious actors to leverage AI tools for criminal activities. The …
-
Human approval integrated into CrewAI autonomous agents
This guide demonstrates how to integrate human approval into autonomous AI agent workflows, specifically using the CrewAI framework. It introduces a custom tool, ApprovedEmailTool, which acts as a gatekeeper for actions with real-world side effects like sending emails. The tool …
-
Coding agents should gate state-changing commands, not all actions
A coding agent's security hinges on how it handles permissions, with two common pitfalls: constant approval requests leading to fatigue, or no requests at all, risking destructive actions. The author, who builds agentproto, proposes a solution: differentiate between commands tha…
-
AI Tutor Enhanced with Persistent Memory System
This article details the creation of a persistent memory system for an AI tutor, designed to retain information across restarts. The author developed a tiered architecture that includes importance scoring, memory decay, and fact consolidation to manage the AI's knowledge effecti…