Agile was invented to solve a specific problem: software development was slow, expensive, and wrong. The waterfall process — full requirements, full design, full implementation, then test and discover you built the wrong thing — routinely produced systems that were over budget, late, and didn't match what users actually needed. The Agile Manifesto in 2001 was a corrective: working software…
In 2024 Apple released an ad for the iPad Pro . A hydraulic press slowly destroys a grand piano, paint cans, cameras, books, guitars, trumpets, record players — physical manifestations of human creative expression — as they are compressed into a thin slab of aluminum and glass. They intended to communicate that all you need is an iPad to access all these art forms, that the iPad was the ultimate…
Oh look another nerd deciding to waste more internet bandwidth on all the wonderful virtues of Jujutsu VCS . I'm not the first and I won't be the last. Here's a sampling of the discourse: Steve Klabnik's Jujutsu Tutorial — probably the most comprehensive community tutorial Jujutsu for Everyone — structured for people with zero VCS experience, which is very relevant to what…
In AI research, scaling laws are gospel. Put in $X$ more compute, get $Y$ improvement in capability. The Chinchilla paper, the GPT-4 technical report, every major lab's internal planning documents—they all revolve around trying to develop these predictable relationships between resources and outcomes. Scaling laws are what is used to justify billion-dollar investments. They're why…
The UI Moment, Again In November 2022, ChatGPT's launch felt like a paradigm shift . Suddenly, everyone was talking about AI. The technology press couldn't stop writing about it. Your relatives asked you about it at Thanksgiving. It seemed to appear overnight. But it didn't appear overnight. Large language models had existed for years. GPT-3 launched in 2020. BERT came out in 2018.…
The Vibe Coding Automated Programming Revolution Software engineering is fundamentally changing. AI coding agents like Claude Code have put the vision of just telling a computer to do something in natural language and it just working out of the box in sight - the kind of natural language computer interface that Star Trek: The Next Generation and other science fictionpromised decades ago. This…
Waypoint Transformer The Waypoint Transformer was published at NeuRIPS 2023 by Badrinath et. al. and is a new method for Reinforcement learning via supervised learning (RvS) that builds on the prior Decision Transformer and is conditioned on automatically generated waypoints. Online vs Offline Reinforcement Learning In reinforcement learning, the agent learns to interact with an environment by…
Building CLIP from Scratch to Classify Pokemon Prior deep learning approaches to performing image classification frame the problem as given a dataset of image, label pairs, learn a mapping between the image space and the label space using a neural network (MLP, CNN, etc.). Here the image space is typically a CxHxW dimensional space where C is the number of channels, H and W are the height and…
This week has been weird to say the least. We have fringe figures in the artificial intelligence community publishing opinions in Time Magazine on the need to take violent action to stop the propagation of large language models , but I thought the most important story was Elon Musk and Twitter releasing the source code for their Tweet recommendation system . I'm going to keep referring to…
Sections 1 and 2 are motivation and background if you care, Section 3 is the real meat of the essay. ChatGPT is more Information Retrieval UI than Artificial Intelligence (In the AGI sense) Generative AI is having its moment right now and OpenAI's ChatGPT is at the center of it. The "humanness" of the conversations one has with these chatbots have many pundits busy wondering if this is the…
A Procedurally Generated Mountain Town I spent the fall learning about computer graphics and OpenGL, a topic which I didn't get to cover in my undergrad. OpenGL is now a 30 year old standard which is primarily written in C/C++. It is a nice way to start since a lot of the common components needed in a graphics stack are baked in. But its usage patterns and APIs are showing their age so I…
In the practice of developing machine learning models, there are few tools as approachable as PyTorch for developing and experimenting in designing machine learning models. The power of PyTorch comes from its deep integration into Python, its flexibility and its approach to automatic differentiation and execution (eager execution). However, when moving from research into production, the…
Dumping Facebook Sophomore year of college I started becoming more aware of the all the data I was posting online, mostly because I realized how much time I was spending on Facebook in the first place. Before it was on the first page of my home screen on my phone and always opened in a browser tab. Whenever I had a free moment on the bus or an empty URL bar in the browser I would find myself…
I wrote a post for the NVIDIA Developer Blog on how we implemented real-time autonomous vehicle perception using the specialized hardware available on the NVIDIA DRIVE AGX Self Driving Computer. The AGX is powered by Xavier, NVIDIAs in-house ARM SoC which includes a Volta-based GPU, two programmable vision accelerators (used for tasks like edge detection, epipolar geometry) and two deep learning…
A portion of the Twilight Installation in the office of ACM@UIUC Twilight is a concept for self organizing office lighting based on similar ideas to my work on the Amorphous Facade . I did this work for my Undergraduate Senior Project with Rauhul Varma and advised by Rebecca Chen and Prof. Can Bayram. Phillips Hue and similar lighting systems are really popular and they actually can have health…
ACM@UIUC GPU Cluster In order to effectively work with deep learning a GPU is required. When I was an intern at NVIDIA, this wasn't a problem. NVIDIA maintains a massive GPU compute pool which is accessible to all its engineers ( https://www.nvidia.com/en-us/data-center/dgx-saturnv/ ). It's harder when you don't work for the company who makes all the…
When one thinks of neural networks, probably the first thing they think of is a deep learning framework like Tensorflow or PyTorch. The creation of deep learning frameworks were crucial to the adoption of deep learning in the products we use every day. Instead of having to write your own CUDA kernels to leverage the parallelization power of a GPU, you can easily structure components together into…
Thoughts after running Experience for HackIllinois 2016 about growth and direction of HackIllinois and Hackathons in general. This was written in May of 2016 with Andrew Kuznetsov as we were trying to assemble a vision of what HackIllinois could become. Background HackIllinois was founded in 2014, just one year after the first MHacks, and 5 years after PennApps. For the past 3 years, HackIllinois…
In 2016 interestingly enough Anheuser Busch sponsored a hackathon in the ECE Department. At the time I started working in Prof. Tim Bretl's Lab on robotic grasping research and our lab had a Baxter from Rethink Robotics. A few labmates and I thought it would be fun to turn it into a bartender.
The impetus to create such a framework stems from the reality that the majority of developers who work in our development group, will not be there in 4 years (and most will leave earlier). We also see new developers walk in oblivious of the frameworks we have used in the past. Just in the past four years, Rails has been supplanted by Node.js and Angular/React, and PHP has gone from a “goto”…
Weave The potential of IoT in the mid 2010s was really embodied by Nest, a company founded by the same people who developed the iPod and ultimately was bought by Google. Nest had focus on elegant design and intelligent behavior and this was embodied best by the Nest Thermostat, a small puck which at the time was completely different than anything on the market. It not only looked different, it was…
This work was done during the Greylock Hackfest 2016 We entered the 2016 Greylock Hackfest and for the first couple hours didn't know what to work on. We actually left Facebook HQ where the hackathon was being held to make a run to Fry's to see what hardware they had available. During this time not having an idea we were browsing Facebook and one of us remarked on how bad the image…
From video to 3D model automatically During my time working in the UIUC Vision Group with Prof. Hoiem I worked on 3DFS (Deformable Dense Depth Fusion and Segmentation). 3DFS is a 3D reconstruction and segmentation system from standard video. The processing stages of 3DFS in order to go from video to object reconstruction and segmentation masks Users can take a video filming different angles of an…
This work was done during the 2015 Greylock Hackfest Coming to a hackathon with no idea what you want to work on is always leads to the same thing. Once you have your team, the HQ for the next 36ish hours and unpacked the two backpacks worth of computer hardware, maybe an extra keyboard and mouse and some pillows you then preceed to spend the next 2-3 hours throwing around ideas, most of which are…
Something unique about UIUC is that they let a bunch of undergrads try to and teach a course in the form of the CS Department's Freshman Projects Course. They took this course as an oppurtunity for undergrads to decide what they thought was important to learn when you are an ambitious student just entering college and want to get up to speed quickly. Everyone has heard the complaints about…
I gave a TEDx Talk on my work at NOAA on climate modeling, CU on swarm robotics and Design and Developement work for my High School and how you can use design thinking as a framework to approach for interdisiplinary studies.
For 2013 and 2014, I spent a good deal of my free time and all of my summer working on research in Swarm Robotics at the University of Colorado Boulder in Professor Nikolaus Correll's Lab. My main research interest was in exploring ways swarm robotics could be deployed to create new interesting interaction models. Concept art for the Amorphous Facade My work centered on a platform developed…
Transcription of my IB Math HL Paper from 2013 Google being a leading company in the technology industry in order to remain competitive needs to find the best talent in computer science, mathematics and the hard sciences. However pure recruitment does not always find the best and the brightest. In an effort to attract driven and curious people to work for Google, the company released this…