更改

删除1,373字节 、 2021年9月16日 (四) 11:58
第155行: 第155行:     
==硬起飞与软起飞==
 
==硬起飞与软起飞==
  −
[[File:Recursive self-improvement.svg|thumb|upright=1.6|In this sample recursive self-improvement scenario, humans modifying an AI's architecture would be able to double its performance every three years through, for example, 30 generations before exhausting all feasible improvements (left). If instead the AI is smart enough to modify its own architecture as well as human researchers can, its time required to complete a redesign halves with each generation, and it progresses all 30 feasible generations in six years (right).<ref name="yudkowsky-global-risk">[[Eliezer Yudkowsky]]. "Artificial intelligence as a positive and negative factor in global risk." Global catastrophic risks (2008).</ref>|链接=Special:FilePath/Recursive_self-improvement.svg]]
  −
  −
[[文件:在这个示例递归自我改进场景中,修改人工智能体系结构的人可以每三年将其性能提高一倍,例如,30代人,然后用尽所有可行的改进(左)。相反,如果人工智能足够聪明,能够像人类研究人员那样修改自己的架构,那么每一代人完成一次重新设计所需的时间将减半,并且它在6年内将所有30代可行的代都推进(右图)。<ref name="yudkowsky-global-risk">[[Eliezer Yudkowsky]]. "Artificial intelligence as a positive and negative factor in global risk." Global catastrophic risks (2008).</ref>]]
  −
   
在硬起飞的情况下,通用人工智能AGI迅速自我完善并“掌控”世界(也许在几个小时内)。速度太快,以致于无法进行重大的人为纠错或对 AGI 的目标进行逐步调整。在软起飞的情况下,虽然AGI仍然比人类强大得多,但却以一种类似人类的速度进步(也许是几十年的数量级),持续的人类互动和修正可以有效地引导AGI的发展。<ref>Bugaj, Stephan Vladimir, and Ben Goertzel. "Five ethical imperatives and their implications for human-AGI interaction." Dynamical Psychology (2007).</ref><ref>Sotala, Kaj, and Roman V. Yampolskiy. "Responses to catastrophic AGI risk: a survey." Physica Scripta 90.1 (2014): 018001.</ref>
 
在硬起飞的情况下,通用人工智能AGI迅速自我完善并“掌控”世界(也许在几个小时内)。速度太快,以致于无法进行重大的人为纠错或对 AGI 的目标进行逐步调整。在软起飞的情况下,虽然AGI仍然比人类强大得多,但却以一种类似人类的速度进步(也许是几十年的数量级),持续的人类互动和修正可以有效地引导AGI的发展。<ref>Bugaj, Stephan Vladimir, and Ben Goertzel. "Five ethical imperatives and their implications for human-AGI interaction." Dynamical Psychology (2007).</ref><ref>Sotala, Kaj, and Roman V. Yampolskiy. "Responses to catastrophic AGI risk: a survey." Physica Scripta 90.1 (2014): 018001.</ref>
  
113

个编辑