Creează un cont pentru a te abona la acest blog și a-i urmări articolele pe pagina Lectură.
-
I ran out of time to write a good blog post for September, so I wrote a quick blog post for September.
-
If you live in DC, and any of the events I mention look interesting, please feel free to reach out!
-
Over the past year of writing professionally, I have learned when to write dumber.
-
Everyone needs a ramen food: quick, mindless, cheap, and tasty.
-
I recently went on a ten day trip to China where I didn’t take my computer.1 I still wanted the ability to take notes quickly, so I bought a Bluetooth keyboard to use with my phone.
-
To my knowledge, I've set three concrete goals for myself about how much effort I am going to put into this blog: In A promise to myself, I made an (implicit) promise to write a blog post every month, in perpetuity.
-
Growing up, I was one of those kids who watched every YouTube video I could at double speed.
-
Note: I have a goal of writing a blog post every month, basically for forever.
-
There's an idea that success requires constant effort applied over decades.
-
TLDR: you should leave a comment, either spam or not, so I can test how well this system works!
-
I was a ten year old kid, reading the middle of the fourth book of the second series of Percy Jackson when I abruptly lost interest in the plot.
-
When I started as a full time journalist at Understanding AI, I thought I would want to write less in my free time.
-
Admire the following opening sentence: High, high above the North Pole, on the first day of 1969, two professors of English Literature approached each other at a combined velocity of 1200 miles per hour.
-
Note: I wrote this a couple of days ago.
-
[Note: I got this idea from a Reddit post.
-
I've been researching the privacy implications of AI recently (in a word: scary), and as a result, I've been reading a lot of terms and conditions of different services.
-
The year is half-over.
-
[Note: this is by no means a new idea: I learned it from Rohan Subramani.]
-
2025 has been my best year yet for journaling.
-
Previously: An Introduction to Correspondence Texting In my previous post, I divided texts into three buckets: (1) respond immediately, (2) respond within 24 hours, and (3) respond when convenient.
-
See also: Non-correspondence texting I hit a milestone today.
-
[Note: this is written in a op-ed style to be a writing sample.]
-
I.
-
Continuing the recent spate of AI-related posts ...
-
TL;DR: Download the shortcut here and follow the instructions in the appendix to get yourself set up.
-
In my previous past, I talked about using LLMs to learn more about coding1.
-
N.B.: I wrote this while at the Recurse Center (recurse.com) back in October, but for some reason did not post it.
-
One of the minor marvels of sunsets is that phone cameras cannot capture them well.
-
I often want to change or build a habit.
-
Note: this is most appropriate for the Northeast corridor, as this is what I am used to.
-
It isn't. 1.
-
In my previous post, I talked about using Wikipedia to expand my taste in music by repeatedly using the random page function until I hit a page about music.
-
In the internet age, how does one find new music?
-
Around a week ago, I made a typing test as part of a workshop on bad art.
-
Like many people looking for a CS job, I've had to do LeetCode style assessments, and have practiced by solving problems on leetcode.com.
-
1.
-
There's a tradition at RC that at the end of a batch, one should write a return statement to reflect on their experience.
-
I recently learned about python decorators from Dan Bader's excellent resource Python Tricks the Book.
-
My partner and I both have things we'd like to do every day.
-
One central difficulty with my present journaling habit is that it is brittle.
-
1.
-
Today, I learned about the Boyer–Moore majority vote algorithm - Wikipedia when completing a LeetCode problem about trying to find an element of a list which appears over half the time.
-
0.
-
Measuring Jailbreaks and Adversarial Robustness on a Toy Problem: Fixed Syntax Abstract/tl;dr Jailbreaks occur when Large Language Models are successfully prompted to generate harmful responses to questions, where they otherwise would not.