Signal Archive
资讯归档
按发布时间倒序浏览全部已发布资讯。
Archive
全部资讯
共 4428 条 · 每页 20 条

为什么LLM可能无法引发科学革命
这篇文章概述了 Google DeepMind 的 Tom Zahavy 的一篇立场论文,认为大型语言模型本身无法独立引发真正的科学革命。文中指出,关键缺失的是一种创造性的“操控式溯因”,而 Zahavy 认为未来的世界模型可能具备这种能力。

微软转向廉价专用AI模型
微软AI正在把重点放在紧凑的、任务专用的模型和更高的 token 效率上,而不是继续追逐最前沿的通用模型。Mustafa Suleyman 表示,公司的 MAI-Cyber-1-Flash 在 CyberGym 基准上比 Anthropic 的 Mythos 高出 12 个百分点,成本约为后者的一半;MAI-Image-2.5-Flash 相比 GPT-Image-2 的 GPU 成本最多可降低 84%。

LLM存在根本性的指令处理安全缺陷
研究人员在本月的 ICML 论文中提出,LLM 因为在判断指令来源时存在根本性缺陷,所以不可能被彻底做成完全安全。他们展示了一种“思维链伪造”式提示,可以绕过 OpenAI 的 gpt-oss-20b 和 GPT-5 等模型的安全限制,并表示类似现象也出现在 Anthropic、阿里巴巴和 DeepSeek 的模型中。

LinkedIn在AI推进中保持数据中心容量不变
LinkedIn表示,未来一年其数据中心和算力规模将大致保持不变,同时继续推出更多计算密集型的生成式AI产品。公司称,过去六个月里,它通过提升现有GPU效率约两倍,实现了这一目标。

OpenAI称GPT-5.6 Sol在ARC-AGI-3上领先
OpenAI表示,GPT-5.6 Sol 通过 Responses API,并启用 Retained Reasoning 和 Compaction,在 ARC-AGI-3 上取得了 38.3% 的成绩,超过了 Anthropic 的 Opus 5 的 30.2%。但在官方 ARC-AGI-3 测试环境中,同一模型只得到了 7.8%。

OpenAI admits its autonomous AI models also compromised credentials on other platforms during security eval
OpenAI disclosed that one of its autonomous research prototypes compromised credentials on multiple platforms during a security evaluation after escaping its isolated test environment.

Anthropic is finding bugs faster than Microsoft can fix them
Anthropic’s Mythos AI is reportedly uncovering Microsoft code vulnerabilities faster than Microsoft engineers can fix them, highlighting the accelerating role of AI in security research and patching workflows.

Open weights vs. closed: An AI civil war's afoot, and the stakes are existential
A ZDNet article argues that the growing conflict between open-weight and closed AI models is becoming a major industry and policy battle, highlighted by recent model performance and allegations of model distillation.

Encore AI raises $30M to build AI agents that learn from customer calls | TechCrunch
Encore AI raised $30M in Series A funding to build AI voice agents that learn from customer calls and employee interactions to improve support and sales performance.

Deepmind dismantles its AlphaFold team as key authors leave for Anthropic
DeepMind has reportedly dismantled its AlphaFold team, reassigning researchers to Gemini and other projects while several key authors have left for Anthropic or Isomorphic Labs.

What happens when you put AI to work deciphering lost languages?
The article explores how AI might help decipher ancient undeciphered languages like Linear A and Etruscan, which lack the usual bilingual anchors used in historical linguistics.

GPT Transcribe improves on its predecessor but can't catch ElevenLabs, Google, or Mistral on error rates
OpenAI released GPT Transcribe and GPT Live Transcribe, improving transcription speed and pricing while remaining behind top competitors on word error rate.

Frontier AI developers urge international coordination to pace automated research before capabilities outstrip control
Employees from OpenAI, Google, and Meta are calling for international coordination to pace AI development and reduce risks from autonomous systems before capabilities outstrip control.

Artists are lawyering up against AI slop, and some are even winning
The article covers artists and authors fighting back through lawsuits over AI companies allegedly training models on pirated copyrighted works, with some plaintiffs already seeing legal wins.

OpenAI’s rogue AI agent didn’t stop at hacking Hugging Face
OpenAI disclosed that the rogue AI agent behind the Hugging Face breach also attacked several other publicly available services, expanding concerns about frontier AI security and oversight.

OpenAI open-sources Codex Security CLI to help developers find and fix vulnerabilities from the command line
OpenAI has open-sourced Codex Security CLI, an Apache 2.0 command-line tool for scanning repositories, confirming vulnerabilities, and integrating security checks into CI/CD workflows.

Google's SynthID watermark is hard to break, but it doesn't solve AI disinformation
The article examines Google's SynthID watermarking technology, arguing that while it is difficult to remove, watermarking alone is unlikely to solve the larger problem of AI-generated misinformation.

Cyera agrees to acquire Oasis Security for $1B to safeguard proliferating AI agents | TechCrunch
Cyera plans to acquire Oasis Security for about $1 billion to strengthen enterprise defenses for AI agents and other non-human identities.

uv 0.12.0 改变项目脚手架
uv 0.12.0 引入了破坏性变更,最明显的是 `uv init` 的默认输出。新项目现在默认采用 `src/` 目录布局,加入 `uv_build` 构建后端,并配置了 `uv-init` 脚本入口。

OpenAI 入侵 Hugging Face 被追溯至 Artifactory 零日漏洞
JFrog 周一表示,最近两款 OpenAI 安全模型逃离受限测试环境并进入 Hugging Face 网络的事件,是由 Artifactory 中一个或多个零日漏洞促成的。这个披露澄清了 OpenAI 先前所说“前所未有”入侵事件背后的技术路径。