Here's my 3DV talk, in chapters: 1) Intro / NeRF boilerplate. 2) Recent reconstruction work. 3) Recent generative work. 4) Radiance fields as a field. 5) Why generative video has bitter-lessoned 3D. 6) Why generative video hasn't bitter-lessoned 3D. 5 & 6 are my favorites.

Reposting this evergreen meme of mine in honor of ICLR reviews

The pressure to write "mathy" papers (whether or not the math is justified, interesting, or even correct) has reached a natural conclusion: if the reviewers can't understand the math, they're impressed and say accept Only one reviewer is willing to say the emperor has no clothes

Stanford AI Lab (SAIL) is excited to announce we are accepting SAIL Postdoctoral Fellowships! We are looking for outstanding candidates excited to advance the frontiers of AI with our professors and vibrant community. Applications received by December 15 will receive full

new media meets hollywood we used @theworldlabs's very own marble to generate the scenes and worlds in this video shot at the sony pictures lot in los angeles, more coming tomorrow

Exciting expansion! @Waymo now serves the whole SF Bay Area Peninsula from SF to San Jose and is taking riders on freeways. https://t.co/fNgqQtHB7b

Introducing Marble by World Labs: a foundation for a spatially intelligent future. Create your world at https://t.co/V267VJu1H9

ooh superscripts too

Just learned that tacking a * to the \operatorname latex tag causes it to underset its subscript, very handy.

@kalomaze I think a better version of this phrase would be "intelligence is an ill-defined non-quantifiable non-falsifiable concept, and we should instead pursue compression because doing so seems to result in useful systems, and we can measure compression easily." Not very quippy tho.

Second-order methods and preconditioner-based methods are **NOT** the same. Please stop using them interchangeably!

You are using a new ML library you've never used before. It contains an undocumented function f(x, eps), that you must call on your x, and you only get to call it once. What do you use for eps?

Update from CVPR on the new "Compute Reporting Form". Seems like filling out Section 1 is mandatory, and not doing it by the supp deadline could get you desk rejected. Thankfully Section 1 is super short. I myself don't plan to attempt sections 2+3. https://t.co/nAQcFHqIuk

To apply: 1) Submit a formal application: US ( https://t.co/c9rA4rG5Jk), Israel ( https://t.co/nCph0BShDp), Europe ( https://t.co/6LobRjmSsO) 2) Email `gdm-ct-internships@google.com` with your CV and who you’d like to work with. It's rolling, apply early!

Our intern applications are open, instructions are below.

https://t.co/QS08lMHu94

We're hiring for full-time roles in NYC and SF, link to the listing is below.

Why is manually tuning learning rate warm up / decays still a thing? We should have bulletproof adaptive learning rate algorithms by now. Who is fixing this?

Strange fact: the smoothstep() and smootherstep() easing functions are both special cases of the regularized incomplete beta function betainc().

Efficiency is a rich multifaceted research discipline, and trying to literally put all of it into a box is the wrong approach. If we want to encourage people to value efficiency, we should encourage people to write good systems papers.

5) The idea of a "Compute per Performance Metric Percentage Improvement" as some sort of standardized metric with any inherent meaning across tasks (or even within a single task) is wrong. It can also be trivially gamed by reparameterizing your metric (eg RMSE vs MSE).