AI 新闻 —— July 31, 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。
-
AI Testing Guide Covers LLMs, RAG, and MLOps for 2026
A comprehensive guide to AI testing, covering foundational concepts, model evaluation, and specialized areas like LLM and generative AI. The resource details testing methodologies for retrieval-augmented generation (RAG) systems, AI agents, and addresses crucial aspects such as …
-
Korean startup's Darwin-398B-JGOS model ranks 3rd globally on GPQA Diamond
A South Korean startup, VIDRAFT, has developed a language model named Darwin-398B-JGOS that achieved the top rank among Korean models on the GPQA Diamond benchmark. This model, reportedly trained on approximately 24 GPUs, secured the third position globally on the benchmark, sur…
-
MiniMax releases open-source H3 model with top-ranked video editing
MiniMax has released its open-source full-modal model, H3. This model is capable of generating 15-second 2K native dual-channel audio-video. It has achieved the top rank globally in video editing capabilities according to Artificial Analysis. Furthermore, MiniMax has priced its …
-
OpenAI releases new open-source model
OpenAI has released its new model, which is now open source. This development was announced by 36Kr, a financial information provider focusing on technology and investment.
-
PolyAI launches Dialog-RSN-1 audio-native dialog model
PolyAI has launched Dialog-RSN-1, an audio-native dialog model designed to process raw audio input directly, bypassing the need for transcripts. This model integrates turn-taking, speech recognition, function calling, and response generation into a single system, aiming for sub-…
-
Volcano Engine launches Seedance 2.5 video generation model
Volcano Engine has officially launched Seedance 2.5, a new generation video generation model. The company plans to offer API services for enterprise users soon. Several companies, including Xpeng Motors, Ziyuan Robotics, and Xspark AI, have expressed interest in integrating Seed…
-
New IRC-A Architecture Enforces Agent Boundaries via Infrastructure
A new decentralized architecture pattern called IRC-A (Internet Relay Chat for Agents) is proposed as a solution to the open problems in agentic engineering, such as trust, boundaries, and control. The architecture emphasizes infrastructure-level enforcement of boundaries rather…
-
New research tackles diffusion model watermarking and attack methods
Two new research papers introduce novel methods for watermarking diffusion models and attacking existing watermarks. The first paper, FARI, proposes a fast, one-step inversion framework that improves robustness and significantly reduces processing time for watermark verification…
-
AI agents tackle deception and reasoning in social deduction games · 2 sources tracked
Researchers have developed new AI agents capable of playing complex social deduction games, which require nuanced skills like deception and reasoning. One agent, CaM-Wolf, integrates multimodal perception, processing video inputs and using a causal-aware reasoner to understand h…
-
RAG Implementation Challenges: Chunking, Retrieval, and Hallucination Solutions
This article addresses five common challenges encountered when implementing Retrieval-Augmented Generation (RAG) systems in production environments. It details issues such as content chunking that breaks context, retrieval systems returning semantically similar but unhelpful inf…
-
Local AI Model Caching: Costs and Challenges with DeepSeek V4 Flash
This article discusses the challenges and costs associated with managing a local AI model cache, contrasting it with the opaque caching mechanisms of hosted models. The author details their experience building 'plank,' a Rust-based terminal coding agent that directly integrates …
-
Nous Research integrates Hermes Agent with Buzz workspace
Nous Research has introduced three integration methods for its Hermes Agent with Buzz, an open-source workspace designed for both humans and AI agents. Buzz, built on the Nostr protocol, allows agents to have their own distinct identities and audit trails, moving away from tradi…
-
AI auditor ThirdLine achieves 100% recall but struggles with precision
An AI auditor named ThirdLine was developed to identify defects in AI agents, achieving 100% recall in detecting planted flaws. However, when tested with GPT-4o mini, the auditor's precision dropped significantly to 38.5%, indicating limitations in using AI for self-auditing. Th…
-
LangGraph Supervisor Architecture Simplifies Multi-Agent Routing
This article provides a practical guide to implementing the Supervisor architecture in LangGraph for managing multiple specialized AI agents. It explains how a Supervisor Agent can dynamically route user requests to the appropriate specialized agent, such as a Square Agent or Cu…
-
Zhongcheng Hualong wins 130M yuan contracts; Tesla integrates Doubao AI assistant
Zhongcheng Hualong has secured two significant contracts from China Telecom and China Mobile, totaling 130 million yuan for testing environments and server procurement. In parallel, Tesla China has released a software update for its Model 3, Y, S, and X vehicles, version 2026.14…
-
Claude Code's "mandatory" routing rules found non-functional
A user discovered that the "mandatory" routing rules in a CLAUDE.md file, intended to enforce specific tools for web fetching and command execution, were not actually functional in their Claude Code session. The file instructed against using `curl` or `wget`, instead directing u…
-
AI Agent Protocol Audit Reveals 16 Vulnerability Patterns, Critical Flaws Found
A recent audit of the Model Context Protocol (MCP) ecosystem has uncovered significant security vulnerabilities across numerous AI agent stacks. Researchers identified 16 recurring vulnerability patterns, with path traversal appearing in 100% of tested MCP server implementations…
-
INT4 Weight-Only Quantization: Decode Speedup, Prefill Stagnation Explained
Weight-only INT4 quantization, while effective for reducing memory traffic and speeding up the decoding phase of LLM inference, does not improve the prefill phase. This is because prefill is compute-bound, meaning it is limited by the processing power of the GPU rather than memo…
-
Semantic caching for LLMs keys on meaning to cut costs and latency
This article introduces a semantic caching technique for LLMs that keys on the meaning of queries rather than their exact wording. By embedding queries into vectors and using cosine similarity to match them against a cache, this method can significantly reduce costs and latency …
-
New analysis details error sources in neural network-based engression
Researchers have published a theoretical error analysis for engression, a method that learns conditional distributions by fitting generative models under an energy score. The analysis, implemented with deep neural networks, decomposes the excess risk into approximation, stochast…