Recent content on Vincent Cheng
Here are some ways I use LLMs that I think are making me dumber: When I want to build a Chrome extension for personal use, instead of actually learning and writing the JavaScript, I Claude-Code the whole thing in a couple of hours without writing a single line of code. Instead of taking the usual route which would leave me with more actual familiarity with JavaScript, I now shortcut the process,…
Here’s a running list of tiny workflow tips that make my day-to-day noticeably smoother. Most of these are embarrassingly simple, but that’s the point! Habits that seem obvious to me might be totally new to someone else (and vice versa). I will continually update this every ~month and try to only include things that I have kept up using for more than a month. Message me your favorites…
What I’m Doing Now (updated 7/19) Working at METR over the summer on some CoT monitoring projects! In Berkeley/SF
Hey, I’m Vincent! I was a freshman at UCSD and am a sophomore studying math at Cornell now. I change my mind about things very often, but wanting to learn new things is a constant! I’m especially interested in math, ML, AI safety, Brazilian Jiu-Jitsu, and meeting new people. Message or email me if you want to talk! Things I’ve been a part of and enjoyed: Ross , WARP , Atlas ,…
This web app uses your webcam, a lightweight depth model, and three.js to recreate that oddly satisfying metal pin toy effect. Try it out here . Try out the simulation for yourself here and check out the code here .
What do truly long-context models look like? I want to give the model all my journals, notes, pictures, previous work, etc. so that it can make connections and tailor responses for me. I imagine this to be In-between context stuffing and fine-tuning. Every ~day, the model takes all the conversations from that day and decides which to use to update its weights. In the future, will everyone have…
Some ideas I find interesting but don’t have enough time to make a reality. A “See all context” button. This is especially useful for reasoning models since I want to know which parts of the reasoning were dropped so I know how much context to provide. You would have to drop sensitive information like like system prompts. JARVIS. Models are capable enough already! We just need…
The Man In The Arena It is not the critic who counts; not the man who points out how the strong man stumbles, or where the doer of deeds could have done them better. The credit belongs to the man who is actually in the arena, whose face is marred by dust and sweat and blood; who strives valiantly; who errs, who comes short again and again, because there is no effort without error and shortcoming;…
Brazilian Jiu-Jitsu (BJJ) is a grappling sport which means there is no punching or kicking and you win by gaining dominant positions or submissions such as chokeholds or joint locks. What Makes BJJ Different? BJJ is the “broadest” sport I’ve done. It feels like math in the sense that you have so many distinct, but related, concepts to learn and problem-solving to do. In tennis or…
obviously non-exhaustive Movies/shows Monster (Japanese movie): messes with your head. 8/10 Squid Game The Prestige: It’s so good. I don’t get how it’s not Interstellar-level popular. 10/10 Alice in Borderland Top Gun Money Heist Stranger Things Arrival 火神的眼淚 模仿犯 Books Open (Agassi biography) Three Body Problem series When Breath Becomes Air Tomorrow, and Tomorrow, and Tomorrow…
My opinion of whether Cursor helps or hinders my work has fluctuated significantly over the past few months. Cursor is really powerful. Having never worked with selenium drivers and barely editing any lines of code, I made a pretty nice scraper in two afternoons. It just works. However, there’s a pretty significant distinction of using AI code editors that I haven’t seen explicitly…
I thought this video was really fun and wrote up some of the derivations that the video went over quickly. Introduction and Basic Principles Maupertuis’ principle of least action states that the action, defined as: $$ S_0 = \sum mvs $$ where $m$ is the mass, $v$ is the velocity, and $s$ is the distance, reaches a minimum along the actual path of motion. Euler later generalized this to a…
Some weird stuff happens in high-dimensions. Reference 1. High Dimensional Oranges Are Almost All Peel Consider an $n$-dimensional cube of side length 1 containing a smaller $n$-dimensional cube with side length $0.8$ (“pulp”) surrounded by a $0.1$-width border (“peel”). The volume of the pulp is $0.8^n$, which rapidly approaches 0 as $n$ increases: Dimensions Pulp Volume…
Many people have written about curating a very good Twitter feed , but I have yet to see anyone talk about doing this with Youtube. I don’t know if many people do it but don’t talk about it, don’t do it, or what. I guess “subscribe to tons of accounts you enjoy, use the “Not Interested” for ones you want to avoid, and harvest” is not very deep. YouTube’s…
Noam Shazeer and Jeff Dean on Dwarkesh arithmetic very cheap. moving data around is expensive. model parameters are very memory efficient: one fact per parameter? (this probably isn’t the right way to think about it because of superposition?) versus in context, there are kqv which can many more bits inference improvement thing? big model verifier, small model does it first thing??…
I want this to be an informal/public notebook where I record thoughts that are too long for a non-premium Twitter account, notes on different things I’m reading, and maybe more formal writings as well. The target audience for this page is a mix of myself and current/potential friends. Through this notebook, I hope to “produce” more and write better. For the longest time,…