Knowledge Delta

Shipped my first website today: knowledgedelta.org A learning map for everyone who uses AI for research. A gift for our 6th anniversary, and a nudge to myself to keep going. It’s v0.1 and rough around the edges. Feedback very welcome.

Every model launch week, my timeline turns into 4,000 people grading the same exam with different answer keys.

You don’t have to finish everything. Finish the one thing you know actually matters. The rest can wait for a clearer decision later. Today, only handle one thing you’re sure is important. Leave the rest alone for now.

Grad school teaches you two skills: reading 40 papers in a week, and pretending you read 40 papers in a week.

Reviewer 2 isn’t your enemy. Reviewer 2 is the only person who read your paper as carefully as you did.

Read fewer papers, but read them slower. One paper you truly understand beats ten you skimmed for citations.

Small test worth running: give Claude, ChatGPT and Grok Bot the same messy CSV, ask for summary stats, and count which ones check for duplicate rows first.

What the AI era really asks of us: don’t overestimate yourself. AI makes you feel like you can do everything. Code a website, design a logo, write in five fields. You can, sort of. But “sort of” in ten directions is worth less than “really good” in one. Know your strengths. Keep your main line. Let AI widen your reach, not scatter your focus.

Your to-do list isn’t too long. Your “not-to-do” list is too short.

Ideas keep pouring out, and the website takes shape a little more each night. It’s tiring, but deeply satisfying. In an internet this loud, I’m building myself a small, quiet place to think.

Model comparisons often mix three questions: who's fastest, who reasons best, and who survives a long agent loop. Different tests deserve separate scores.

Six years together. ❤️ November 8, 2020 — we met during our PhD journey. From two doctoral students to building a life together. Six years later, still us. And so many more years to come. Happy anniversary to us! ❤️

A monthly AI subscription can still run on session windows and weekly caps. The price is monthly; the workflow interruption isn't.

The comment that saves a paper is usually this: you measured the wrong thing very carefully.

The timeline is running three separate fights as if they were one: is Opus 5.5 worse this week, is Codex faster, and how does the plan's usage window actually work?

Seminar rule I wish existed: no one may mention a paper they only met as a summary.

Every semester, I notice the same thing. Students are getting better at producing answers and worse at sitting with a question. The moment something feels hard, the tab opens. AI responds in seconds. The discomfort disappears, and so does the learning. I don’t ban AI in my class. That would be pointless. Instead, I’ve started asking students to spend ten minutes with a problem before they ask anything. Write down what confuses them. Make a guess, even a bad one. Some of them hate it. But the ones who stick with it start asking better questions, and better questions get better answers, from AI and from me. Learning was never about avoiding confusion. It was about learning to stay in it a little longer.

My AI setup right now: Claude for writing and thinking, ChatGPT for quick questions, Grok Bot for running my day. People keep asking which one is best. Wrong question. Ask which one fits the job in front of you.

A highway can reach a village in two years. Opportunity takes a generation.