type
Post
status
Published
date
Apr 13, 2026
slug
writing-proposal-lessons
summary
tags
thoughts
PhD
academic
reflection
communication
category
生活随笔
icon
password
On Friday, my advisor called and asked me to help him write a proposal.
His idea was to let me try — to see whether I could produce something similar. I don't know physics, but I know large language models. After he walked me through his idea, I ran it through an LLM and wrote a first draft. It looked polished, but a lot of it didn't hold together. Then he sent me his own version, and I rewrote mine using his template. He said:
This one actually works.
Three lessons, which he laid out for me on the spot.
1. Decide who your reader is. Write close to them, not close to yourself.
2. Write to the criteria, point by point. Every review item should map to something in the text — nothing missing, nothing extra.
3. Be thorough, but don't over-explain. Say what needs to be said; don't do the reader's thinking for them.
They sound like platitudes, but I only really heard them after a rewrite. My first draft failed on all three: I wrote to some imagined "general reader," I didn't map to criteria, and I stuffed in my own process of understanding the idea.
But what actually stopped me wasn't those three.
It was realizing that what I did in this whole exchange was, essentially, "use an LLM." The physics wasn't mine. The idea wasn't mine. The template wasn't mine. What I brought was the ability to translate a domain expert's thinking into instructions the LLM could follow, and then to clean up what it gave back.
Anyone who has seriously learned how to use LLMs can do this.
So I've been forming a stronger and stronger subjective take: the CS major, as a major, is basically dead. The skills you used to spend four years acquiring — writing algorithms, wrestling with frameworks, gluing CRUD together — almost no one needs them anymore. What we still need are people with a CS background standing behind the AI as conductors: telling it what to do, and judging whether it did it right.
Which creates a tension: if the skills themselves are depreciating, where does the value of a CS person come from?
Probably from addition. AI plus anything — AI for education, AI for physics, AI for biology. "Knowing how to write code" alone isn't worth much anymore; "knowing how to write code + knowing a field" is. I probably shouldn't be aiming for the lecturer track either. I should find a domain to graft onto, grow a second root.
These thoughts aren't fully worked out yet. I need to talk more with my advisor, and I need to go read — industry reports, demographic reports — to see where the times are actually heading.
Things change too fast. There isn't much we can do except go with them and do our own work well.
Having written all this, I still want to thank my mentors. Being able to meet someone who will walk through it with you, point by point, over the phone — that's a lucky thing. Especially in a time when even "CS as a major" is shifting under our feet.
Loading...
