7:58AM: ![[Pasted image 20250111075851.png]] https://x.com/AlexKontorovich/status/1877810433790558243 This is an important point. There are many cases where you will see something, and *think* you know what the answer is but you really don't. To use an analogy with AI: you are *hallucinating an understanding*. The key blocker to this is the cost of acquiring a ground-truth understanding - even if it is only looking up the definition of something. AI is clearly doing a great job here but it requires conscious effort to do so. --- 8:02AM: The above observation reminded me of a [blog post by Cal Newport on how he got the highest marks in his discrete math class](https://calnewport.com/case-study-how-i-got-the-highest-grade-in-my-discrete-math-class/) > **Proof Obsession:** Discrete math is about proofs. In lecture, the professor would write a proposition on the board — e.g., _if n is a perfect square then it’s also odd_ — then walk through a proof. Proposition after proposition, proof after proof. As the class advanced, we learned increasingly advanced techniques for building these proofs. I soon developed a singular obsession: **I wanted to be able to recreate, with pencil and paper, and no helper notes, every single proof presented in class.** No exceptions. Lack of understanding of even one proof wouldn’t be tolerated. I assert that Cal was successful because he avoided hallucinating an understanding through conscious effort to understand how to reproduce the hard to learn. 3:20pm: Reading code is underrated. ![[Pasted image 20250111152143.png]] https://x.com/giffmana/status/1877831103781667146 We forget that we spend more time reading than writing code. This is also a lower-risk scenario than changing your code and a good way for people to build confidence in the AIs. --- 4:23pm: This [video](https://www.youtube.com/watch?v=SQHYj7x-t3A) showed up in my feed. It's about something I had never heard of before - a **linear calendar**. You can see what it looks like in the screenshot below: ![[Pasted image 20250111162343.png]] The simple idea is that each month is a row and the weekends line up vertically, and the calendar starts with Monday as the first column. The idea is that you can zoom out and see the structure of an entire year. In the spirit of Ethan Mollick's excellent [Co-Intelligence book](https://www.penguinrandomhouse.com/books/741805/co-intelligence-by-ethan-mollick/), where he implores you to *Always invite AI to the table* (Principle 1), I had a tag-team of o1 and Sonnet 3.6 write a Python program that would generate a linear calendar PDF. It took a few rounds to perfect, but it worked out quite nicely - here's the [PDF](./linear_calendar_2025.pdf). Here's the [gist](https://gist.github.com/jflam/e01c3c1523121c3263f1259d35d29643). ![[Pasted image 20250111162859.png]]