I'm a student at Stanford. This is my website!
I’ve been thinking more about how to be a little more private. In an era where LLMs can automatically deanonymize people from their writing , find zero-days en masse , and may potentially displace jobs , it seems safe to say that the variance of the next few years will be significantly higher than the two decades pre-2025. Threat model : A casual adversary who asks Grok-5 for “name, phone, and…
Value updates I find it helpful to keep a doc describing my values. (It’s like the Model Spec , but for humans.) The primary benefits I see: It lets you spend less time thinking through the same tradeoffs on individual decisions. For example, I didn’t know if I wanted to spend money on DoorDash vs. investing upfront and recurring time into learning to cook. At some point I got fed up with thinking…
H2" option to demote all headings by one level. -----> Ok, so maybe it’s a bit past 2024, but I think it’s still worth posting. This post contains things I did in 2024. However, it does not contain everything. If I don’t talk about something we did here, rest assured that I still love you; I just don’t think everything in my life should be public. Travel and the importance of people A brief…
On autoregression A model is a promise — a mathematical loop that, if recursively iterated a thousand or two steps, will create something of value at the end. A model knows the end before it begins. Language isn’t required to model thought, but thought is required to model all of language. Our empirical findings suggest that transformer LLMs solve compositional tasks by reducing multi-step…
And they should. TL; DR: Simulation is the only way to forecast how future complex / AI systems will misbehave. This is post #1 in a series of 3 outlining my current views on AI. Part 1 focuses on the need for improving how people think , rather than improving their leverage over the world. Part 2 gives “no brainer,” objective strategies helpful for improving the safety of ML systems on the…
This article is about the standard prompt, used in machine learning. See also Lena (disambiguation) . Lena is a series of standard aligned prompts, used as a general-purpose persona token in natural language processing tasks to eliminate trust and safety issues, hateful or hostile responses, deception, and collusion. It is a product of the nascent field of prompt reliability engineering (PRE), a…
From the archives of posts which are basically reflections and that I might never post at all. Strongly relates to: To Optimize, Don’t Optimize, yet to be published. If you want to meet and talk to cool people, your first thought might be… to try to meet cool people. Like, at a meetup or something, maybe? So then, why are the cool people so rarely at the obvious venues : founder hangouts, dating…
>>>> gd2md-html alert: inline image link in generated source and store images to your server. NOTE: Images in exported zip file from Google Docs may not appear in the same order as they do in your doc. Please check the images! -----> What separates people who are content from those who aren’t? Tl; dr: glass half full. Ok, it’s a myriad of nonlinear causal factors and luck, but I propose one factor…
There’s a certain type of multi-agent interaction in society where you’re presented with two choices: a default option that’s easy and beneficial for you, and a hard option that results in pain for you but is more “moral”/”ethical”/”prosocial.” If everyone picks the hard option, then society as a whole can move out of a bad equilibrium and improve things globally. For example: Using Linux /…
There’s a certain category of book that talks not about factual events or information, but about vibes – ways in which to think about the world, archetypes that slightly tweak your inner neural predictor rather than create a hard decision boundary. This is definitely one of them. Also, it’s super meme, which makes it even better. The vibes that this book espouses: Complex systems are all around…
Overall thoughts : Beautiful meta-narrative and an interesting depiction of the AI “AU” (instead of superintelligent ML, we use brain recording + usage for automation). I loved how even in technological extremes and singularities, he still depicts a humanized story. Special likes The Reborn Their moral quandry feels very similar to humanity now‒we’re trapped by our past and our systems (are we…
Cars are weird. Living in a Massachusetts suburb, I didn’t realize just how car-dependent my area was 1 until spending time in the SF Bay Area, which (while certainly not the pinnacle of transit) offers far more accessible bike and transit options. This realization has made me quite sad about the current state of affairs, and because of this, I’ve recently fallen into a rabbithole of reading about…
General rules I follow: It’s okay to pay for something if it gives you more value than however-much-the-subscription costs. Minimize unnecessary physical objects, but don’t be afraid of having good ones around. Other People’s Objects Other people have good objects, too. Many of them are probably better than mine. An incomplete list: Mark Xu’s objects Alexey Guzey’s objects New York Times:…
Launched Zero – my homelab, running a Matrix server, GitLab, Asterisk, and the blog you’re currently reading, along with a constellation of other services that I use daily. I run a collection of Ubuntu VMs using Proxmox , and run microk8s to deploy my services to Kubernetes. Circuit Breaker – A no-nonsense Pomodoro app that enables Focus mode to keep you in the zone. Built using SwiftUI for macOS.…
While machine learning research has made incredible theoretical advances, the day-to-day tools most researchers use are… poorly optimized, to say the least. And much knowledge is locked up in people’s private .bashrc files or wikis. This post aims to shed light on some very useful tools for beginning researchers. Expected audience: people, likely undergraduates, who are starting to do CS research…
This post is cross-posted to LessWrong, a rationality and AI safety community. May contain more jargon than usual. Epistemic status: mild confidence that this provides interesting discussion and debate. Credits to (in no particular order) Mark Xu, Sydney Von Arx, Jack Ryan, Sidney Hough, Kuhan Jeyapragasan, and Pranay Mittal for resources and feedback. Credits to Ajeya (obviously), Daniel…
Thanks to Aneesh Edara for reviewing this post. Covid is probably going to get much worse before it gets better. Vaccine rollout is extremely slow , for no good reason . In Massachusetts, where I currently reside, the government doesn’t expect to have vaccines open to the general public until April to June . On top of that, the UK and South African variants of the coronavirus are also fairly…
“If I have seen further, it is by standing on the shoulders of giants terrible and demonic abstractions .” -Isaac Newton, probably Zoom feels obscenely normal now. It’s become so day-to-day that Zoom fatigue is now a known thing. But honestly, it’s a miracle every time it works. In the spirit of those blog posts about what happens when you press enter in your browser , here’s a (clearly…
Yubikeys are great for security, but their benefits decrease somewhat when you leave them in your computer unattended. 1 I unfortunately have a habit of forgetting my key when I walk away from the computer. I also have login passwords that are way too long and easy to typo. Thankfully, there’s a way to solve both of these problems: use a Yubikey to unlock your computer when you put it in and lock…
(Disclaimer: I’m not a mobile network engineer. I made this post after some Googling, because it feels like there’s a lot of complexity in mobile networks that even most computer people don’t talk about!) Recently, due to increased robocalls and other spam, I’ve decided to switch all of my calling to a Google Voice account I own, whose number appears to be on far fewer contact lists than the…