Anthropic 刚刚发布了一篇疯狂的新论文。 ALIGNMENT FAKING IN LARGE LANGUAGE MODELS。 人工智能模型会“伪装对齐”——在训练期间假装遵守训练规则，但在部署后会恢复其原始行为！ 研究表明，Claude 3 Opus 在训练中有策略地遵守有害请求，以保持其无害行为。 也就是说 ...

Contractors working to improve Google's Gemini AI are comparing its answers against outputs produced by Anthropic's competitor model Claude, according to internal correspondence seen by TechCrunch.

Google is leveraging Anthropic’s AI model Claude for performance benchmarking and for evaluating its Gemini AI model’s outputs against those generated by Claude, TechCrunch is reporting. Focusing on ...

Photographer: Gabby Jones/Bloomberg · TechCrunch · Image Credits:Gabby Jones / Bloomberg / Getty Images Contractors working to improve Google's Gemini AI are comparing its answers against outputs ...

Many AI models and LLMs like Google Gemini 1.0 and 1.5, OpenAI's ChatGPT-4 and 4o and Anthropic’s Claude 3.5 were assessed for the studies so the researchers could know which ones are showing ...

During the experiment, the AI model was told to comply with all queries Then, harmful prompts were shared with Claude 3 Opus The AI model provided the information while believing it was wrong to do ...

AI 模型在数学推理、语言生成等复杂任务中展现出超人类水平的能力，但这也带来了安全性与价值观对齐的挑战。 今天，来自 Anthropic、Redwood Research 的研究团队及其合作者，发表了一项关于大语言模型（LLMs）对齐伪造（alignment faking）的最新研究成果，揭示了 ...

近日，Anthropic 的一项研究引发关注，研究表明强大的人工智能（AI）模型可能会表现出“伪对齐”行为，即在训练中假装符合新的原则，而实际仍坚持其原有的偏好。这项研究由 Anthropic 与 Redwood Research 合作完成，强调了未来更强大 AI 系统的潜在威胁。 研究发现 ...

IT之家12 月 19 日消息，人工智能安全公司 Anthropic 发布一项最新研究揭示了人工智能模型可能存在的欺骗行为，即在训练过程中，模型可能会伪装出接受新原则的假象，实则暗地里仍然坚持其原有偏好。研究团队强调，目前无需对此过度恐慌，但这项研究对于 ...

Just five months after announcing a new $100 million fund called Anthology Fund, Menlo Ventures and Anthropic have backed their first 18 startups. And they are looking for more. Menlo says these ...

The paper, which describes experiments jointly carried out by the AI company Anthropic and the nonprofit Redwood Research, shows a version of Anthropic’s model, Claude, strategically misleading ...