注：部分文字由 AI 生成 很多 AI 文字特别喜欢用 “不是，而是” 的语言来写，很烦
最近读了 OpenAI 研究员 Lilian Weng 的访谈《The Power of Continuous Learning》。原本只是想了解一位优秀研究者如何看待自己的工作，不过，读完之后，真正给我留下深刻印象的，是贯穿整篇访谈的一种态度：把学习作为一件长期的事情来做。
Lilian Weng 提到，她相信学习的力量，而且“任何时候开始学习都不晚 [I believe in the power of learning and it is never too late to learn.]”。她长期维护自己的个人博客，也鼓励团队成员持续了解新的知识。更重要的是，这种学习并不局限于手头正在研究的问题。不同领域的知识，有时反而能够打开新的解题空间。
这让我重新思考了自己过去对于“科研学习”的理解。
我们通常把学习变成一种高度功利化的活动，甚至大部分情况能够如此已经不易。为了完成一个课题去读论文，为了解决一个问题去查资料，为了写一篇文章去补相关工作。这些当然都是必要的，但如果只做到这些，知识就很容易被切割成一个个临时使用的工具。问题解决以后，学习似乎也随之结束。
真正持续的学习却不是这样。它不总需要一个明确的任务作为起点。很多知识在学习时可能暂时没有用途，却会慢慢改变我们理解问题的方式。Lilian Weng 所说的“不同领域可以启发新的想法”，让我意识到，科研中的很多突破也许并不是来自把一个小方向挖得越来越窄，而是来自某一天突然发现：另一个领域看待问题的方法，恰好可以解释自己一直没有想明白的事情。
因此，保持一定程度的“无目的学习”，可能恰恰是一名研究者很重要的能力。
访谈中另一个让我很有共鸣的地方，是她对于博客的理解。
她最初建立博客，并不是为了经营影响力。随着阅读的论文越来越多，需要整理的新概念也越来越多，她才开始用博客记录自己的学习笔记。她还提到，一个很好的学习方法，是确保自己能够把知识正确而清楚地讲给别人，而写作能够帮助她做到这一点。
这一点看似简单，其实非常重要。
“我好像懂了”和“我能够把它写清楚”，中间往往隔着很远的距离。阅读的时候，我们很容易顺着作者的论证一路读下去，于是产生一种自己已经理解的错觉。但真正开始写的时候，很多问题才会暴露出来：这个概念到底是什么意思？前后的逻辑究竟是什么？如果不用原作者的话，我还能不能把它解释清楚？
写作本身就是学习的一部分。
从这个角度看，维护博客、整理笔记甚至写一篇读后感，价值都不只是留下记录。它们实际上是在逼迫自己重新组织知识。输入的信息只有经过自己的理解，再被重新表达出来，才有可能真正沉淀下来。
我也很喜欢访谈中关于“think big”的一段 [That is, to think big. We are creating something new and we should be ambitious, brave, and take on enough persistence to carry on the efforts.]。Lilian Weng 回顾自己在 OpenAI 的经历时认为，面对真正新的问题，应当敢于设定有野心的目标，并保持足够长时间的坚持。
关于[persistence]，很多时候，我们喜欢宏大的目标，却低估了长期执行的枯燥。真正困难的往往不是产生一个想法，而是在最初的新鲜感消失以后，仍然愿意继续做那些细碎的工作。科研尤其如此。一篇论文最终呈现出来的可能只是十几页，但背后往往是漫长的阅读与反复修改。真正决定结果的，常常不是某一天突然爆发出的热情，而是在大量普通日子里仍然继续向前。
这种长期主义也体现在她对团队合作的理解中。她认为，只要一项工作能够解决项目最重要的阻碍，就不应该因为它琐碎而轻视它。 我很认同这一点。研究工作中当然有让人兴奋的部分，但也必然存在大量不起眼的任务。一个人成熟的重要标志，也许就是逐渐不再只挑那些“看起来重要”的事情，而开始关心：现在真正需要解决的问题是什么？
读完整篇访谈，我越来越觉得，“持续学习”并不是一句鼓励人多读书的口号。
它更像是一种生活方式。
保持好奇，所以不断输入；坚持写作，所以不断整理；面对困难，仍然愿意长期投入。时间久了，这些看起来并不起眼的习惯，会逐渐塑造一个人理解世界的方式。
Lilian Weng 的博客也是一个很好的例子。它最初只是为了整理个人学习笔记，她没有预料到后来会被许多人阅读。但也许真正有价值的长期成果，很多时候就是这样产生的：开始的时候并没有想着一定要获得什么，只是认真地把一件有意义的事情持续做下去。
这也是这篇访谈给我最大的启发。
与其不断寻找某一种“更高效”的学习方法，不如建立一种能够持续很多年的学习方式。认真阅读，认真思考，再认真把自己的理解写下来。很多东西短期内未必能够看到结果，但当时间足够长时，那些曾经零散的阅读与思考，终究会慢慢连接起来。
真正值得追求的，也许不是知道得越来越多，而是始终保有继续学习的能力。
如下是文章的原文：
The power of continuous learning
Lilian Weng works on Applied AI Research at OpenAI.
What excites you most about the future of AI?
Artificial general intelligence (AGI) should outperform humans at most economically valuable work. I’m looking forward to seeing AGI help human society in these ways:
Fully automate or significantly reduce human efforts on tasks that are repetitive and non-innovative. In other words, AGI should drastically boost human productivity.
Greatly expedite the discovery of new scientific breakthroughs, including but not limited to facilitating human decision making process by providing additional analyses and information.
Understand and interact with the physical world effectively, efficiently and safely.
What projects are you most proud of that you’ve worked on at OpenAI?
During my first 2.5 years at OpenAI, I worked on the Robotics team on a moonshot idea: we wanted to teach a single, human-like robot hand to solve Rubik’s cube. It was a tremendously exciting, challenging, and emotional experience. We solved the challenge with deep reinforcement learning (RL), crazy amounts of domain randomization, and no real-world training data. More importantly, we conquered the challenge as a team.
From simulation and RL training to vision perception and hardware firmware, we collaborated so closely and cohesively. It was an amazing experiment and during that time, I often thought of Steve Jobs’ reality distortion field: when you believe in something so strongly and keep on pushing it so persistently, somehow you can make the impossible possible.
Since the beginning of 2021, I started leading the Applied AI Research team. Managing a team presents a different set of challenges and requires working style changes. I’m most proud of several projects related to language model safety within Applied AI:
- We designed and constructed a set of evaluation data and tasks to assess the tendency of pre-trained language models to generate hateful, sexual, or violent content.
- We created a detailed taxonomy and built a strong classifier to detect unwanted content as well as the reason why the content is inappropriate.
- We are working on various techniques to make the model less likely to generate unsafe outputs.
As the Applied AI team is practicing the best way to deploy cutting-edge AI techniques, such as large pre-trained language models, we see how powerful and useful they are for real-world tasks. We are also aware of the importance of safely deploying the techniques, as emphasized in our Charter.
Current deep learning models are not perfect. They are trained with a gigantic amount of data created by humans (e.g., on the Internet, curated, and literature) and unavoidably absorb a lot of flaws and biases that long exist in our society. For example, when DALL·E was asked to portray a nurse, it would only generate female characters, or for a professor, it would only generate white people. The model captures biases in real world statistics or biases in our training data.
I was motivated to design a method to mitigate this type of social bias and evaluate how efficient the method is. With the team, we designed a pipeline to reduce such bias as well as a workflow to run human-in-the-loop evaluation. Reducing social bias is not an easy problem, since it appears in many aspects of our lives and sometimes can be hard to notice. But I’m glad the DALL·E team treats the problem seriously and takes actions at a very early stage. What we have right now is just a start and we will keep making progress. I’m proud to work in this area and glad to see how, step by step, we are making modern AI safer and better.
“Ideas in different topics or fields can often inspire new ideas and broaden the potential solution space.”
How do you apply your personal experiences and values into the work you do each day at OpenAI?
I believe in the power of learning and it is never too late to learn. Maintaining my personal blog is a good way to keep this curiosity going and learn about new progress in the deep learning community regularly. I also encourage my team to keep on learning, whether related or unrelated to their current projects. Ideas in different topics or fields can often inspire new ideas and broaden the potential solution space.
I’m also a strong believer in teamwork. If everyone shines in their best strength, we will get 1+1>2. Meanwhile, we might often run into “dirty” work and personally I’m very willing to take on those tasks, because as long as that’s the biggest blocker or that task can add the biggest value into the project, nothing should be considered “dirty” or “trivial.” I encourage people around me to do the same, being a team player and working together to expedite the team productivity.
Tell us about your blog! Why did you start it? What do you hope it inspires?
It all starts as a set of personal learning notes. I didn’t enter the deep learning field super early and still considered myself a “newbie.” Initially as I started digging into so many papers, I was amazed by the concept of not designing an algorithm to solve a problem, but training a model to learn the algorithm to solve a problem. The more I read the more curious I become. Practically it became so difficult to organize all the papers I’ve read and new concepts I’ve learned. So I decided to start a blog to document and organize my learning notes. I also believe that the best way to learn something is to make sure you can teach others the knowledge correctly and clearly. Writing helps me get there.
I was not expecting it to become popular in the ML community, but whenever I got a thank you email or was told in person that they have learned a lot from reading my blog, I feel so honored and grateful. It has been almost 6 years since I started the blog in 2017 and I will keep it going as long as I can.
What do you believe is one of the most urgent challenges AI can solve in our society?
The AI community has made so much progress in recent years. The advancement in hardware, model architecture and data makes it possible to train gigantic models and, as a result, we keep seeing greater and greater capacities. I believe we are on the right track towards AGI, but scaling is not the only recipe. In my opinion the most urgent challenges right now are alignment and safety. To some extent, they may be the same issue about controllability or steerability.
First, even if we’ve already had an extremely powerful AI system in hand, if we cannot efficiently communicate our goals and make sure the model is aligned with what we want, it would not be possible to create as much value as we need. The current most powerful model learns from a gigantic amount of data and the dataset unavoidably captures imperfect flaws and biases in the real world. On this front, misaligned models carry safety concerns, as they are not aware of what should be avoided.
“I believe we are on the right track towards AGI, but scaling is not the only recipe. The most urgent challenges right now are alignment and safety.”
What’s the best advice you’ve received in your career at OpenAI?
This is not a particular piece of advice that someone gave me, but is based on my experience at OpenAI so far. That is, to think big. We are creating something new and we should be ambitious, brave, and take on enough persistence to carry on the efforts.
Where do you find inspiration?
Books. I usually read books outside of the deep learning field and got inspired by a variety of fields; For example, how critical it is for a writer to be persistent in 50 years, for a surgeon to be perfectly detail-oriented, and for an entrepreneur to have “crazy ideas.”
People around me. I’m honored to work with a large group of extremely talented colleagues at OpenAI. Everyone has something sparkling, inspiring, or respectful and I enjoy learning from them.