OpenAI在其新型Astra模型中采用了名为"递归深度"的推理技术，该技术让模型在非线性循环中处理查询，但留下的可读追踪明显减少1。这一做法立即引起了AI安全领域的警惕。
Redwood Research首席执行官Buck Shlegeris表示，"如果OpenAI进一步推进这项技术，他们将有选项大幅增加递归并完全破坏思维链可监测性"1。AI安全倡导者Zvi Mowshowitz则认为，可能需要采取法律手段防止AI实验室之间的"竞争下沉"1。安全专家们的核心担忧在于，该技术的推广使用可能会削弱对AI模型行为的可监测性。
对此，OpenAI首席科学家Jakub Pachocki表示该公司致力于保持思维链的可读性1。与此同时，Anthropic和Google DeepMind已开始讨论这一技术1。
OpenAI has implemented a reasoning technique called "recursive depth" in its Astra model that processes queries through non-linear loops, significantly reducing the readability of the model's thought process 1. This approach generates fewer visible traces compared to traditional chain-of-thought reasoning, prompting alarm among AI safety experts who fear diminished transparency in how advanced AI systems operate 1.
Redwood Research CEO Buck Shlegeris warned that if OpenAI continues advancing this technology, the company could substantially increase recursion and "completely destroy chain-of-thought interpretability" 1. AI safety advocate Zvi Mowshowitz has suggested that legal measures may be necessary to prevent a "race to the bottom" among AI laboratories in deploying such techniques 1. In response, OpenAI Chief Scientist Jakub Pachocki has stated the company remains committed to maintaining readable chain-of-thought processes 1. Both Anthropic and Google DeepMind are reportedly engaged in discussions about the technology 1.
评论
还没有评论，欢迎留下第一条。