spacecraft engineer | ai researcher | systems thinker
A neverending holy war rages on Twitter over the future of the robotic form factor; legged or wheeled, highly general or hyperspecialized, Figure or ABB. Overlooked in these debates, like water to a fish, is the most common robotic form factor in our society - the drone. And as the fiber-optic-laden fields of Ukraine have shown us, drones are here to stay. Drones will be one of the first vectors…
In which I make the argument that writing effective simulations will become vitally important to shipping code safely and efficiently in a world with coding agents. Read it on Substack: https://joshholder.substack.com/p/simulation-is-the-new-high-ground
In which I describe Waymo’s safety strategy, and why I think it could be good news for alignment and safety in domains like LLMs and humanoid robotics. Read it on Substack: https://joshholder.substack.com/p/safety-without-understanding-lessons
In which I draw lessons from classical control theory to argue that LLM attractor states are more dangerous, and less blissful, than they seem. Read it on Substack: https://joshholder.substack.com/p/llms-spiritual-bliss-and-the-boeing
It’s May 2024: Falcon 9 is roughly on-target for a mind-boggling 144 launches this year, Starship continues its march to full capability, and thanks to various articles (most prominently Starship is Still Not Understood by Casey Handmer), many in the tech community have built an intuition for what this might mean for the space industry as a whole - cheaper mass to orbit leads to more lenient mass…
With terminal-phase Guidance Navigation and Control in the news lately with the (qualified) success of Intuitive Machines and JAXA’s lunar landers, I wanted to provide an approachable look at the mathematics of rocket landing, and the ways it’s both easier and harder than you might expect. To me, there’s nothing more awe-inspiring than watching rockets land autonomously, and in fact watching this…
Leveraging reinforcement, imitation, and supervised learning to beat humans at Judgement, a playing card game which I’ve played far, far too much of (repository here - try your hand at beating the bot!), with applications to general trick-taking games. 1. Background: The League of Judgement Over my college career, I played a LOT of the trick-taking card game called Judgement/Kachufool (rules…
This piece aims to clearly and concisely lay out the bull case for a post-Starship space industry. While I wasn’t breaking new ground with this analysis (a lot of the core ideas are inspired by Casey Handmer’s excellent blog post), it was a fantastic exercise to distill this core thesis into a two page brief aimed at those with little previous exposure to the space industry. I also dipped my toes…
An obsession with finding the optimal strategy for a simplistic puzzle game leads me through an exploration of various fields of computer science. After countless hours and applying knowledge of parallelization, optimization, human-AI collaboration, and even reinforcement learning, I managed to wrest the world record (1196 splits, 1208 w/ human assistance) first from Craig of flashingleds.net, and…
To address the ever-increasing problem of space debris, I propose a solution involving electrically-propulsed Small-Sats. The focus of this investigation is on the big-picture feasibility of the mission from an orbital mechanics perspective, but everything from the economics of the mission to debris rendezvous and capture is touched upon in the PDF report contained within this post. Mission…