Kostas Daniilidis

@KostasPenn

Oct 28

Kosta Derpanis

@CSProfKGD

Oct 28

I vividly recall advice from my PhD supervisor: always read the conference version of a paper before the journal version.

Journal versions often become overly formal, losing the intuition of the original paper. This advice has served me well.

Today we have a flood of LLM-generated writing with little intuition, excessive formality, and no regard for succinctness. Some papers (and blogs) go on for pages without saying much or, worse, confusing the reader 🙄

Am I against LLMs for writing? NO! But I worry that the next generation of researchers may never develop the skills and judgment to steer them effectively, instead ceding the entire process to an LLM.

How can you recognize good scientific writing if you’ve never learned to produce it yourself?

Kostas Daniilidis

@KostasPenn

Oct 10

Peyman Milanfar

@docmilanfar

Oct 10

(Retired) Adm. James O. Ellis Jr., Chair of the National Academy of Engineering @theNAEng gave a great speech recently about the significance of Engineering practice and AI : x.com/i/article/2108…

Kostas Daniilidis

@KostasPenn

Sep 18

Angjoo Kanazawa

@akanazawa

Sep 16

A few years ago, it was hard to imagine OpenAI putting serious resources into 3D/4D.

Now large-scale real2sim is within reach!

These systems build on the tools we’ve built and what we’ve learned. Progress!!

So much more to do, and a whole new set of problems to explore.

Jason Ma

@JasonMa2020

Sep 16

Wentao Zhu

@walterzhu8

Sep 16

GPT-6 Astra test 4/n

Prompt: "Implement pen spinning with a dexterous hand. Use Isaac Lab for RL training, use the Sharpa hand, and create the pen mesh yourself. Give me a trained RL policy and a visualization video. You are free to search the web and download papers or anything running autonomously for a day and a half (including training the policy), this is what we got:

Credit: my student Chengyang Li

Gautam Kamath@thegautamkamath

Sep 16

Transactions on Machine Learning Research@TmlrOrg

Sep 16

TMLR has faced a deluge of submissions, necessitating stricter desk rejection policies due to limited reviewer capacity

Co-EiC Nihar Shah reached out to authors of 10 papers slated for desk reject. Could they answer questions about their *own* submission?
medium.com/@TmlrOrg/askin…

Kostas Daniilidis

@KostasPenn

Sep 14

Lerrel Pinto

@LerrelPinto

Sep 12

I’m sensing deep despair in academics over the past week. Astra, Fable, Muse are zero-shotting benchmarks in robotics & world models.

Uneasy pill to swallow, but this is what step jumps in progress looks like.

Kostas Daniilidis

@KostasPenn

Sep 12

Qixing Huang

@qixing_huang

Sep 12

Every field seems to be confronting some form of misalignment with AI; see, for example, mathandai.org.

I am attending ECCV this week, and in my own field of 3D vision, there were serious debates in the workshops about whether explicit 3D representations are still necessary for many 3D-related tasks, such as novel-view synthesis and robotic action. One increasingly influential view is that video generative models have become so powerful that we can simply dispense with explicit 3D.

I would not necessarily disagree if the sole objective is to achieve strong performance on benchmarks or practical tasks. Given the limited availability of explicit 3D training data, approaches built around explicit 3D representations can currently be outperformed by alternatives that exploit much larger sources of 2D and video data.

But, much like the mathematicians above argue, science is not just about achieving better numbers. Understanding is also extremely important. Explicit 3D, for example, gives us a language in which to define and reason about multi-view consistency. If one of our goals is to understand why a method works and what it has actually learned, I do not see how we can simply throw explicit 3D away. Without such understanding, how can we know when a method will fail, or whether it will behave correctly in corner cases that are poorly represented in the training distribution?

Fortunately, I was also very happy to hear several interesting talks at ECCV that emphasized understanding rather than performance alone—for example, Christian Rupprecht's talk at the SFM-DL workshop.

Kostas Daniilidis

@KostasPenn

Sep 12

Phillip Isola@phillip_isola

Sep 11

With all the Astra demos, there might be a feeling in robotics of a bitter lesson taking hold?

I think this moment is rather the opposite.

Astra works *because* it leverages traditional human-engineered methods as its tools.

(...for now)