I now adapt on-the-fly translations to help me distinguish the vocabulary used by the media or other CS papers, and my own interpretation of those words. Cognitive science or psychology people are usually more careful when dealing with these concepts than the CS folks. And anthropomorphizing computer programs, as far as in the media and the papers, is an unfortunate and unproductive namespace pollution/inflation. These words are so ill-defined in the CS context that replacing them actually incurs zero loss of communication. Try:

• Replace “intelligent” with “good” or “smart”
  • This intelligent model boosts your work productivity. → This is a good model that boosts your work productivity.
  • This intelligent coding model gets pretty much everything done for you. → This good coding model gets pretty much everything done for you.
  • Self-reflection is a unique skill core to intelligent systems. → Self-reflection is a unique skill core to smart systems.
• Replace “reasoning” with “do something complex”. We can keep well-defined ones like “deductive reasoning” or “logical reasoning”.
  • A visual reasoning task → do something complex with visual inputs
  • Chain-of-thought reasoning → generating long, somewhat connected, sequences to do something complex
  • A reasoning model → a model that does something complex
• Replace “agent” with “someone that gets things done”. And the “someone” here can refer to a human or a computer program.
  • Our best intelligent multi-agent AI will cater to your business needs 24/7. → Our smart team of people/programs who gets things done will cater to your business needs 24/7.

To go one step further, I didn’t really like “learning” in “machine learning” when I first encounter it: what does it even mean for a machine to “learn”? Is it “explain a bunch of generated/curated data with a model”? Then it should be statistics, not the way kids (or adults) actually go through in an education, by acquiring information (I will constrain myself from using inflated words like “knowledge”) mostly second-hand like how we read from textbook about dinosaurs (no human have ever seen them), Antarctica (few of us have been there), or all kinds of abstract ideologies.

I also have a complaint with “artificial intelligence”: not only is it poorly defined, but also it is egotistic - it sort of implies that, only, we humans have intelligence, and everything else is either not smart enough, or artificial (created by us), but what about the dolphins, the aliens? That said, I have no better alternative right now.

I can empathize with the desire to anthropomorphize non-human activities though - we are a lonely species, blessed and cursed with the motivation to understand itself. Yet messing up with the vocabulary is counterproductive to achieving that goal, however convenient/compelling/attractive that is to an investor, a funding agency, a customer, and the public.

This covers things happened between October 30, 2023 and November 17, 2024. In 2023, I moved from Toronto, ON, to Ithaca, NY, then Palo Alto, CA, and finally settling down in NYC, NY in January 2024. So far I do not regret the decision to quit and move.

Highlights

• Wrapped up months of human-model live deployment on MTurk.
• Completed the first draft 2 weeks ahead.
• Faster at writing. It now takes only one weekend to draft a 2 page proposal.
• Improved engineering skills. Things often turned out way easier than I imagined.
• Occasional dopamine rush from random ideas and a delusion of absolute research freedom.
• Reached the top of Koko Crater Arch Trail in Honolulu.
• Wore the Sorting Hat in Museum of Pop Culture in Seattle.

Lowlights

• Dealt with two annoying MTurk workers.
• More comfortable with (silent) rejections. This is perhaps a highlight.
• Communication failure that led to a good amount of stress.
• Social energy drained in too many occasions, e.g., COLM 2024, NLP retreat, student lunches.
• Did not manage to use all of the TPUs.

Gradients

Should-haves

• Asked for reference letters weeks in advance, not last minute.
• Taken advice from those more experienced with reviewers and ACs.

On research (They are subjective and need more gradient updates of less bias.)

• Academia is more stochastic on a micro level (advice and reviews for specific projects), but way more predictable as collective groups (research paradigm and popular benchmarks follows 80/20 rules).
• I tend to give the higher scores among all reviewers.
• More words than I thought are devoted to drawing, clarifying, and justifying the scope. The trick of many papers is “just” moving around the conceptual or implementation interface across components.
• Communication and writing are more important than I thought. Good writing is audience-aware and information dense.
• Ease-of-use is more important over other virtuous bits than I thought, i.e., pip-installable, tweet-able gifs, catchy titles, first-to-market, extensibility.
• The benchmark methodology is more pervasive than I thought.
• An average ML reviewer
  • puts in less attention than I thought.
  • wants a method diagram, a table, a plot going up. The more benchmark/method names they are familiar with (regardless of relevance), the better.
  • has a broader definition of comparable scope than the authors (which could be a good thing in retrospect).
  • assumes they understand a paper more than they actually does.

What’s next

I think about

• Latent/fluid abstraction synthesis without programs.
• Cooperative learning that seeks to build a common ground.
• What good research is, what valuable research is, and what kind of research I prefer.
• Is everything, even abstract concepts, a remix of a careful selection of everything else? If yes, then the argument against AI for inability to create doesn’t hold.

I want to learn about

• ML theory: learning theory, RL theory, …
• Humanity in general: psychology, sociology, economics, history, …

I want to practice

• Being aware of task rationale, i.e., why am I doing this, and task execution, i.e., what do I need right now.
• Being aware of ego, mine or others.
• Active listening and be kind(er).

这是2018年江苏卷高考作文题，我就是在这一年参加高考的。高中语文我学得很糟糕，一则我懒得背默写古文和写作素材，二则我理解和表达没有适应阅卷人期待的方式。至于高中我能保存些读书的兴趣，那全归功我藏书破万卷的语文老师。本科出国后，我便把这段回忆打包留在了国内床底。

今年秋季我即将读博，而研究的方向是自然语言习得和推理。闲来无事我看认知科学的科普视频，介绍到感质(即主观意识经验的存在性和唯一性)，当头一棒，承载高考语文记忆的匣子突然打开。这好像从哲学层面(笑)解释了我为什么认为阅读理解不应该有标准答案，以及写作不应该单一评分。至于“花解语，鸟自鸣”题干，我现在依然没读懂。2018年的答卷，是要我用博士五年来提交吗？

以前同事聊到我们是不是存在于模拟世界中，坚定无神论的我自然是嗤之以鼻。可是最近生活总是捶我，提醒我冥冥中天意已有安排。高三语文拉垮，然后大学专业选择了擅长的化学；大一等不及动辄上月的化学实验，接着转专业学机器人；大三修不动硬件，又开始捉摸编程语言；大四认识到自然语言兼具表达力和组合性(以及LLM的风浪)，又调整博士申请方向。我不停折腾，但最终好像又回到起始。不说了，我要信宿命论了(笑)。

这好像又是有征兆的。高中读特德·姜讲外星人语言的科幻小说《你一生的故事》，大学的爱好是学习形形色色的编程语言，连最近读的王小波都在讨论说话和人工智能。如果没有天意，大概是我兜兜转转找到了喜欢的领域吧。

Too long.

Before senior high: The mechanics

Learning English was a necessity in China. Parents send their kids to English tutors when they are merely fluent in Mandarin. Part of it is the fear of missing out on the Golden Age for language learning, but I think pragmatically, it is because school admissions assess English skills. My mom used to put sticky notes on furniture and appliances with their English names. “Washing Machine” was my favorite because it read and sounded like two naughty twins - there was something visually and phonetically interesting about languages.

Then started primary school and junior high. Classroom English was about building up the vocabulary and memorizing grammatical rules and exceptions to those rules, of course. Recitation and spelling tests were routines. For example, we were asked to write fourteen ‘put’ phrases and their translations. I doubt I can now, but once upon a time, I could.

In the summer after Grade 7, I went on an excursion to learn French and gave up when I heard about verb conjugations. English rules ain’t that bad… Retrospectively thinking, English is like dynamic-typed programming languages, e.g., Python, while French is like C++. The more conventions and rules, the more precise and less ambiguous. But what about Mandarin? Mandarin is like APL, where visual hints matter and the grammar rules are minimal. Alright, back to English.

I learned the mechanics of the English language in classrooms, but I also wanted to learn English in the wild. China banned Google in the early 2010s, and my rebellious curiosity kicked in. After a winter of trial and error and plenty of help from friends, I managed to bypass the Great Firewall to consume the “forbidden” content. To an unimaginative fourteen-year-old, they were Bob Ross’s painting tutorials and the latest The Big Bang Theory episodes.

Senior high: Chemistry is more straightforward in English

English remained a core subject in senior high (the other two being Chinese literature and mathematics). I was drawn toward natural science subjects like chemistry and physics. My phenomenal chemistry teacher believed one should learn a topic in the language it was developed in, and she recommended Principles of Chemical Science. I completed the collection in two weeks: It was unexpectedly and refreshingly intuitive. Watching MOOC videos grew on me as a hobby (oh yes) till today, but at that time, I was just procrastinating and avoiding the regular school work. English became my native language for STEM subjects such as organic chemistry and calculus. I benefited little from those topics in high school (I did not even attempt to link them to what I learned in the day). On the other hand, though, they did make some first-year undergraduate courses particularly familiar.

In senior high, I was close with our English teacher. She tried her best to neutralize exam-oriented training with immersive learning. Every week, she handed out letters or speeches to recite. She was remarkably tolerant and supportive of me. When others were required to take excerpts from newspapers, I was allowed to read any material I enjoyed, and she lent her book collections to me. I am grateful for her trust.

I became confident that I could succeed academically in an English-speaking institution. So I took the IELTS test for the first time in the summer of 2017. One year later, I flew to Canada and joined the University of Toronto as a first-year undergrad in Chemical Engineering.

Undergrad: A second personality

I will speak English well in 3 months. That was a goal I set in September 2018. Well, that didn’t quite work out - I might be able to describe a chemical reaction, but I had no idea about daily conversations, such as asking for notes or comforting a low-spirited friend. My writing skills were underdeveloped too. Writing a trivial thank-you email took me longer than an hour of editing and researching. Perfectionism did not help, and nor did the global pandemic or remote classes since the Spring of 2020.

I noticed that I acquired a different personality when I spoke English. Good puns and savvy comebacks in Mandarin just don’t translate well. The cultural barrier also manifested inner dialogues such as “Will this come across as being rude?” or “Is this politically right?” I became this dull and dubious person who only talked about work where she knew there was a single truth that everyone agreed. Fortunately, I studied STEM. And unfortunately, STEM demanded teamwork and engineering writing too. I was often the tongue-tied person in any group discussions before the fourth year. Technically, I needed to learn how to convey subtle points. Mentally, I was attached to producing perfect, bulletproof arguments. It blocked doing research, too, because great researchers could comfortably present half-baked ideas and convince others with their novel ideas.

It was only after much deliberate practice that I gained some confidence to focus on my ideas and stop worrying about the delivery. I stole proudly from mentors and collaborators when they used vivid idioms. Talkshows and sit-coms helped, too. Humor was the perfect test ground with slang. Writing guidelines were a game changer. I wish I could send my-2018-self a copy of How to say it.

Learning a second language and my frustration with their subtleties drove me to linguistics and research artificial languages. What if we can purposefully design every aspect of a language? How do we minimize ambiguity? How do we choose the minimal units that are extensible and composable to express arbitrary ideas? Are symbols suitable units? Should they be characters like in natural languages or functions in functional programming? It was incredibly exciting to find people caring about these questions too, so I guess the struggles were a blessing in disguise after all.

Work: A tool, ChatGPT, and beyond

I graduated in 2022 and joined Google as a software engineer. Unlike others, I was genuinely curious about “corp speak” and diplomacy. Sugarcoating and un-sugarcoating are the next skills I want to develop.

Working at Google was a great opportunity to learn about communication styles too. Someone smarter than me identified that cultural differences, namely low-context versus high-context, impede communication. For instance, “not bad” in a low-context culture might translate to “excellent” in a high-context culture. It seems like I transpassed the region where there is a right way to say something to a new region where there is a suitable style to say something.

Recently, I came across the Dictionary of Obscure Sorrows, which invented artificial words to cover real feelings. I have always considered English a static summit that overlooks every possible idea. Yet people deemed the need to invent new words. I get this vague feeling that learning a language is no longer about learning grammatical rules or sounding smart or funny, but about self-expression and connecting with others.

Very recently, OpenAI revealed ChatGPT. Excited about its unprecedented humanness, I also felt embarrassed that I had spent more than twelve years without getting anywhere even close. ChatGPT’s capability raises questions: Is communication irreplaceable by automation? What skills remain valuable in 50 years? In my opinion:

• Deliberate creativity. Not only in the sense of execution, as Imagen does with text prompts, but knowing what to mix. It is the same case in art as in research. Great scientists are visionaries with intrinsic heuristics.
• Trust and empathy. When it comes to dealing with people, humans are at a natural advantage over robots.
• Decision-making. It is not only because of the lack of transparency of ML systems today but also because robots cannot be held legally responsible. (Yet?)

This is a long post. So I will stop here.

– 契科夫 1886年5月27日 致 尼 亚 列伊金

快到六月了，我又想起了契科夫的摸鱼宣言。今天来兴趣想要读一读英文翻译，却没在古登堡计划的《契科夫书信集》找到这封信，只有TO N.A.LEIKIN MOSCOW April 6, 1886。收信人应该对上了，但是时间不对。

我愿意相信契科夫写过这句话。高中好像读过契科夫的短篇小说，但什么也记不得。没想到再次遇到竟还是这跨越一个半世纪却依然让人共鸣的吐槽，可以和胡适的打牌，鲁迅的讽刺，汪曾祺的栀子花媲美。（不过为什么我记得的都是些网红语录呢？大概既想炫耀“有趣”灵魂，又懒得真的去读。）契科夫1860年出生，那时26岁，边在莫斯科当医生，边在文学届摸打滚爬 — 好一个“社畜”。我想读他的其他信件。

以前的作家还可以留下思考的证据，二十年后研究当代作家可能得翻推特微信了。我有些偷窥欲，很喜欢翻看@TechEmails，观察人们的想法是怎么浮现在文字中的，尤其是和信任的人交流没有保留时。ESL的我常常词不达意，或者没有感情干巴巴的，所以只能默默向往拥有传达丰富意象的能力。

他们说Large Language Model实现了这么好的泛化能力，就是在于语言文字的庞大数据库和它丰富的表达力。那只是英语的思维模式，那另外一种语言的呢？我学的这么困难，总因为有语言隔阂。如果我们可以直接用脑电波交流呢？进一步，如果我们可以收集脑电波数据集来训练机器呢？这样我们可以像给照片添加莫奈滤镜，给吐槽文字添加契科夫滤镜，给物理题添加爱因斯坦滤镜。从某种意义上，说不定这就是大脑永生的秘密。马斯科的Neuralink，我指望你呢。