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.

Pros: - This first landing of radiance field tech in a mainstream film (that I'm aware of), amazing - Shipped to Hollywood before generative video, nice - It looks great - The movie is superb Cons: - Its narrative role is depicting a holograph that is broken I'll take it

I had the pleasure to work on this project, and I am so stoked to finally see it in the Sphere instead of in a colab notebook.

Start at https://t.co/tzjYhdeODc. Works great on mobile too, just add the page to the home screen in Chrome.

My preferred way to use nano banana is via Google AI Studio. You want to make sure you're on the "Chat" tab (upper left) and you need the "Gemini 2.5 Flash Image Preview" model selected (upper right). Safety settings (middle right) are probably all off by default.

Gemini 2.5 flash image editing (aka 🤏🍌) does way better at this task than any other model I've tested, and also is able to make unusual stylized diagrams that still hold up. Here's a vanilla generation, and one where the nodes are apples and the edges are worms eating apples.

Try it yourself, just upload a still from a cartoon and say "make it 3D" or "make it look like a still from a stop motion film", works great the first try about 80% of the time.

having a blast over here 3D-ifying Simpsons stills using Gemini 2.5 flash image editing (aka 🤏🍌). Some of these were tricky and required first prompting Gemini to split the image into foreground/background layers, prompt-editing them separately, and then prompt-recompositing.

those coin-op airport massage chairs are roughly as expensive as renting a couple H200s

The perceptron algorithm is nearly 70 years old and still holds up great. All it's missing is some nonlinearities and multiplicative interactions.

https://t.co/uIszQLHIoX

There's a team printing radiance fields and showing them off at SIGGRAPH. Posting this here in the hope that someone can tell me how to buy one.

Another fun party trick: light takes 0.1337 seconds to circumvent the globe (or equivalently, to go from one point to an antipodal server and back). Very elite number, very easy to remember.

A helpful rule of thumb for lower-bounding latency between servers: It takes ~5 milliseconds for light to traverse one time zone (ie, 1/24th of the globe). So SF<->NYC, a 3 hour time difference, can never have a ping less than 2x3x5 ms, etc. Fiber incurs an extra 1.5x.

Seems like we haven't yet figured out the right adjectival form of the word "slop" in the context of AI. "Sloppy" is wrong --- slop is often quite neat and tidy. Sloppish? Slopical?

frantically trying to generate engagement bait to reverse my -25 spots/week declining metis list ranking. Kernel machines are good. Consciousness isn't real. Dwarkesh needs more hair. Matlab was fine.

The optimal web presence for job stuff is *a single-page fast-loading vanilla HTML website*. Resume PDFs are hard to share, and usually too short. LinkedIn pages are behind a walled garden. Fancy websites (sections, javascript, etc) are hard for both LLMs and humans to traverse.

The generative 3D/video corner of Google DeepMind that I run in is now hiring research scientists. If you're on the market for full-time roles in that space, email us! barron@google.com

@kiaran_ritchie We have two simulation models. One that does exactly what you tell it, repeatable. And another that can do anything with infinite variation and creativity but unpredictably. The magic will happen when the two are brought together.

Incredible evolution of "Neural Video Games": from GQN (2018) to Genie3 (2025). The future is exciting! https://t.co/EYygTAa3UF