OpenAI周四推出最新AI模型Astra，称其为公司迄今最强大且对齐度最高的模型1。该模型将首先向Daybreak网络安全项目用户开放，随后一周内扩展至OpenAI的Pro、Plus、Enterprise、Business付费计划以及API调用服务1。在代码相关基准测试中，Astra的表现超越了OpenAI自有模型Sol和Anthropic的Fable1。
然而，Astra的推出也因其采用的不透明递归技术而引发业界关注1。这一技术架构会阻碍"思维链"监测，降低AI决策过程的可审计性1。OpenAI首席科学家Jakub Pachocki就此表态称，"随着模型能力增加，可监测性变得更具挑战性"1。此外，OpenAI总裁Greg Brockman披露，公司与微软合同中关于AGI的条款"不再存在"，公司已将AGI定义从契约义务转变为"使命或精神理念"1。
OpenAI unveiled Astra on Thursday, describing it as its most capable and well-aligned model to date 1. The new model will initially be available to users of the Daybreak cybersecurity initiative, before rolling out within a week to subscribers of OpenAI's Pro, Plus, Enterprise, and Business plans as well as through its API 1. In code-related benchmark tests, Astra has surpassed both OpenAI's own Sol model and Anthropic's Fable offering 1.
The release has sparked controversy over Astra's use of opaque recursive architecture, a technical approach that hampers chain-of-thought monitoring and reduces the auditability of the AI's decision-making processes 1. OpenAI Chief Scientist Jakub Pachocki acknowledged the trade-off, stating that "as model capabilities increase, interpretability becomes more challenging" 1.
In related developments, OpenAI President Greg Brockman indicated that contractual clauses regarding artificial general intelligence (AGI) in OpenAI's agreement with Microsoft "no longer exist," with the definition of AGI shifting from a contractual obligation to "a mission or spiritual principle" 1.
评论
还没有评论，欢迎留下第一条。