igilitschenski.bsky.social
Assistant Professor in Computer Science at UofT.
32 posts
331 followers
77 following
Regular Contributor
Active Commenter
comment in response to
post
The latter argument, "power reveals", has also famously been made by Robert Caro in his biography of Lyndon B. Johnson. I'd argue that this is as true for power at a "small" scale as it is for world leaders.
comment in response to
post
Prove it. Are you really mastering that skill? ;)
comment in response to
post
Hard to have no politics once high sums of money are involved.
comment in response to
post
I feel that in industry, everyone knows what everyone else is doing: companies hire from each other all the time. And even if employees follow NDAs, they still bring their experience with them when transferring. When Waymo started working on AV, there was barely anyone to hire from.
comment in response to
post
Waymo is special because it suffers from a first-mover disadvantage like no other AV company. Because there was barely anyone to copy, they had to learn most lessons themselves. If Waymo can demonstrate profitability on the operational side alone, it will already be a huge achievement.
comment in response to
post
That is my main concern, too.
comment in response to
post
I am also looking for a new one. I believe Florian Shkurti has been happily using Quarto (quarto.org).
comment in response to
post
hahaha. This is brilliant! It is exactly why we want vacations!
comment in response to
post
lol. okay. It's an unexpected simulator but a pretty cool one.
comment in response to
post
I'd argue Python-wrapped DSLs don't count as "pure"
comment in response to
post
Hitting the sonic barrier counts, I guess.
comment in response to
post
Umangi proposes an approach for segmentation in scenes created with 3D Gaussian Splatting (3DGS). It avoids changing the underlying optimization process, achieving a speedup over existing 3DGS segmentation methods. end/🧵
Poster: East Exhibit Hall A-C 1303
Project: umangi-jain.github.io/gaussiancut/
comment in response to
post
Ziyi's work in image diffusion models proposes to use a set of per-object representations to control the 3D pose of individual objects in a scene. It was done during an internship with Thomas Kipf's amazing team at DeepMind 2/🧵
East Exhibit Hall A-C 1507
Project: neural-assets.github.io
comment in response to
post
But is it really a progression if many of these ideas were developed independently?
comment in response to
post
More recently, CVPR also started having robotics papers. Plus some multi-modal work. I'd be curious to see the numbers for pure vision papers. They may actually be stagnating.
comment in response to
post
Have you found a good tool for cross-posts? I really like @buffer.com, but it is, unfortunately, aimed at users in advertising with very different social media budgets. It does not have any free education version or personal pricing, and the free version is limited to posting across three channels.
comment in response to
post
With Vector, the Schwartz Reisman Institute, the Data Sciences Institute, and the Robotics Institute, there is a vibrant AI research community around UofT, and we have a growing cluster of roboticists in computer science. Check out our website to learn more and apply. https://buff.ly/3OLbqXr
comment in response to
post
To achieve this, we work on some of the foundational underlying problems, such as controllable 3d and generative modeling, physics and dynamics learning, and model-based RL. It takes a village to build a world. 🌎 3/n
comment in response to
post
The theme of our work is broadly inspired by recent advances in understanding human learning. 🧑🏫 We are interested in leveraging real-world data and knowledge encoded in large-scale models to guide the data generation process. 2/n
comment in response to
post
We've been supervising together for a while, and now, finally, we can officially do this! Congrats and welcome!
comment in response to
post
Wow. That's quite a journey. Great collection!