AI Intel

Reddit上的一项研究探讨了通过探针目标微调（如LoRa）来校准LLM口头置信度的方法。结果显示，通过探测LLM的隐藏状态，可以在0.76的准确率下区分正确和不正确的答案。

Hacker News上的一篇文章及其讨论，探讨了AI时代人类与模型的关系，以及在AI效率提升下，人类在认知和创造力方面可能面临的挑战。

Hacker News上的一篇文章引发讨论，反思AI技术对前端开发领域的影响，担忧其可能导致类似于“前端失去的十年”的停滞或挑战。

一篇arXiv论文提出LCO（LLM-based Constraint Optimization），旨在通过LLM驱动的约束优化，提高Agentic LLM在真实世界任务中的安全性，解决“情境奖励劫持”（ICRH）等问题。

Claude's v2.1.154 release introduces the Opus 4.8 model and dynamic workflows, enabling Claude to orchestrate complex tasks across multiple agents.

Anthropic announced the acquisition of StainlessAPI, an SDK and MCP server platform that has powered all Anthropic SDKs since the early days of its A…

xAI has launched Grok 4.3 on its API, claiming it to be their fastest and most intelligent model to date, excelling in agentic tool calling and instr…

Mistral Vibe has officially shipped, serving as an AI agent for long-horizon productivity and coding, featuring Work mode, Code mode, a CLI, and a ne…

Mistral AI's Connectors API is now in Public Preview, enabling developers to register MCP connectors once and reuse them across Le Chat, AI Studio, a…

The Gemini App received new upgrades, introducing Gemini Spark, a 24/7 personal AI agent capable of taking action on behalf of users under their dire…