In principle, recursive self-improvement is possible indeed! Let me go back a few decades. In the 1970s, my goal became to build a self-improving AI that learns to become much smarter than myself, such that I can retire and watch AIs start to colonize and transform the solar system and the galaxy and the rest of the universe in a way infeasible for humans. So I studied maths and computer science. For the cover of my 1987 diploma thesis, I drew a robot that bootstraps itself in seemingly impossible fashion. This thesis on "learning to learn” or L2L was very ambitious and described first concrete research on a self-rewriting "meta-program" which not only learns to improve its performance in some limited domain but also learns to improve the learning algorithm itself, and the way it meta-learns the way it learns etc. In the thesis, I applied Genetic Programming (GP) to itself, to recursively evolve better GP methods. This was the first in a decades-spanning series of papers on concrete algorithms for recursive self-improvement, with the goal of laying the foundations for super-intelligences. I predicted that in hindsight the ultimate self-improver will seem so simple that high school students will be able understand and implement it. I said it's the last significant thing a man can create, because all else follows from that. I am still saying the same thing. The only difference is that more people are listening.
Now it is important to understand that true L2L is much more than just traditional "transfer learning" from one problem to the next. True L2L means learning the credit assignment method itself through self-inspecting, self-modifying code. True L2L may be the most ambitious but also the most rewarding goal of machine learning. There are few limits to what a good metalearner will learn. Where appropriate it will learn to learn by analogy, by chunking, by planning, by subgoal generation, by combinations thereof - you name it. The challenge is to ensure that the code’s self-modifications are really beneficial to the learning agent in the long run, taking into account that early changes of the learning algorithm are setting the stage for later ones, which makes credit assignment complicated. My Gödel machine of 2003 was the first fully self-referential universal metalearner that was optimal in a certain sense, typically using the somewhat less universal Optimal Ordered Problem Solver for finding provably optimal self-improvements. However, it was optimal only in a theoretical sense that is not very practical. So our work is not finished yet! More information and papers on L2L are here http://people.idsia.ch/~juergen/metalearner.html and here http://people.idsia.ch/~juergen/goedelmachine.html
Juergen Schmidhuber 为机器之心 | GMIS 2017 全球机器智能峰会嘉宾，知乎账号由Juergen Schmidhuber授权，机器之心代为注册和运营，以上为Juergen Schmidhuber的英文版答案，以下为机器之心翻译答案供大家参考。
原则上，可自我改进的人工智能当然是有可能出现的，让我们回顾一下几十年以前。
在20世纪70年代，我的目标就是构建一个自我提升的人工智能，可以通过学习让自己变得甚至比我更加聪明，这样我就可以退休了，看着AI开始殖民外星，将太阳系、银河系和宇宙的其他地方以人类不能企及的方式纳入版图。所以我开始学习数学和计算机科学。在我1987年毕业论文的封面，我描绘了一个以看似不可能方式引导自己的机器人。这篇“学习如何学习”或L2L论文在当时是雄心勃勃的，第一次描述了可自我编写的“元程序”。它不仅可以在某些有限的领域改善自己的表现，也可以改进学习算法本身，改进自己改进算法的方法……这是数十年来首篇针对递归自我完善的具体算法的论文，目的是为超级人工智能奠定基础。
我超前地预测到：可以无限自我改进的程序将会变得非常简单，高中生就能理解并应用它。我曾表示这是人类可以创造的最后一件作品，因为其后的一切都将都根植于此。我还将继续宣扬自己的看法，直到更多的人接受它。
需要注意的是，真正的L2L不仅仅是传统的“迁移学习”，从一个问题到另一个问题。真正的L2L是通过自我检查、自我修改代码的方法来获得收益的方式。真正的L2L或许非常有野心，但也是在追寻机器学习中最有价值的目标。一个好的自学习程序可以发现自己面临的限制。在合适的时候，它将通过类比学习进行学习，通过分块，通过子目标生成，通过组合等你能想到的任何方法来提升自我。我的Gödel是第一个完全自我参考的通用元学习程序，在2003年就被开发出来了，它在某种意义上是最优的，通常使用一些较不普遍的“序列最优问题解决器”来找到自我提升的可靠方式。然而，它只是理论上可行，并不非常可靠。所以我的研究还未完成。有关L2L的论文和其他信息可参阅：http://people.idsia.ch/~juergen/metalearner.html 和：http://people.idsia.ch/~juergen/goedelmachine.html