THIS IS HUGE! Researchers at Stanford University have developed a dual-antibody treatment that remains effective against ALL SARS-CoV-2 variants by targeting a less-mutable part of the virus. This breakthrough could lead to longer-lasting therapies that OUTPACE viral evolution. 🧪🧵⬇️

We are hiring on the Generative Media team in London: boards.greenhouse.io/deepmind/job... We work on Imagen, Veo, Lyria and all that good stuff. Come work with us! If you're interested, apply before Feb 28.

This paper is wild - a Stanford team shows the simplest way to make an open LLM into a reasoning model They used just 1,000 carefully curated reasoning examples & a trick where if the model tries to stop thinking, they append "Wait" to force it to continue. Near o1 at math. arxiv.org/pdf/2501.19393

More Veo 2 examples Prompt: "A snail wearing pants. The snail is riding a bicycle, and has a large moustache." #veo2 Following a comment by thelokasiffers in Twitter

More Veo 2 examples Prompt: "A cat wearing a suit and a top hat, while driving a tractor. The tractor has lots of hay on top. Cinematic.!"

More Veo physics #Veo2 Prompt: "A large iron ball falls on top of a cardboard box full of coins."

Action scenes (with realistic physics!) with Veo! #veo2 Prompt: "Car going at top speed through a road, until reaching a waterfall. It gets into the waterfall and jumps off a mountain. Cinematic, 35mm film."

Physics with Veo! #veo2 Prompt: One ball is in the floor. Another ball comes rolling.

Soccer from the future, according to Veo 2 #veo2

Example video generated with Veo #Veo2 Prompt: The camera follows an octopus flying through a city with very high skyscrapers. A sign says "Veo". Cinematic, 35mm film released in 2024.

Veo v2 generates a meeting of animals #Veo2 Prompt: A meeting of a lion, a bear and a giraffe, all of them wearing suits. Photorealistic, cinematic.

Proud to see the release of Veo V2! deepmind.google/technologies... "Veo has achieved state of the art results in head-to-head comparisons of outputs by human raters over top video generation models"

At Neurips!

Introducing 🧞Genie 2 🧞 - our most capable large-scale foundation world model, which can generate a diverse array of consistent worlds, playable for up to a minute. We believe Genie 2 could unlock the next wave of capabilities for embodied agents 🧠.

Fascinating account of the history of attention, as told by Dzmitry Bahdanau to Andrej Karpathy x.com/karpathy/sta...

I love the notations and methods people in different fields invent for their work. For example the Megadeth drummer: "I developed a method for, when listening to a song for the first time, just count, name the parts, count the bars, and (...) have the structure" (youtu.be/tbUYVcaF_l0?... at 2:35)

.kkrieger is one of my favorite feats of engineering. A game released in the same year than Doom 3, with graphics of similar quality, that fits fully into 96kb. youtu.be/DN4Dtio8zKA?...

I've often used both print-based debugging, and debugger tools in my work. A great middle ground option is printed-out stack traces at key points of a program

I always liked Named Tensor Notation, it's a bit unfortunate it isn't used more namedtensor.github.io

It was often concluded from emergent behaviors like that of Flocking / Boids (en.m.wikipedia.org/wiki/Boids) that real creatures probably followed similarly simple algorithms too. But given enough brain complexity and an RL-like learning process, they couldn't learn such cleanly disentangled rules.

Stop watching videos, start interacting with worlds. Stoked to share CAT4D, our new method for turning videos into dynamic 3D scenes that you can move through in real-time! cat-4d.github.io arxiv.org/abs/2411.18613

The article asks "Will AI lead to useful predictions at the expense of deeper scientific understanding?" Worth thinking about the same question for engineering (e.g. code gen). How do you trade off the short term productivity gains vs. the long term ones given by increased understanding?

New essay by DeepMind about AI for scientific discovery, there's a lot of interesting ideas and citations to others's work here deepmind.google/public-polic...

I was pretty obsessed with the concept of high amounts of dimensions when I was very young. People from a math club teased me a bit about it... "who cares about that? It's pointless" Fun to think how in my job now, I'm pretty often thinking in the behavior of very high dimensional objects.

The starter packs are a brilliant solution to the bootstrapping problem in social networks. I don't think I ever saw a similar approach before.

Machine learning and game development communities have some big differences; but they do share lots of passion and curiosity.

One of my main motivations has always been making things that seemed impossible. Generative models are a perfect match: what we see them do right now looked definitely impossible just a few years back.

For no particular reason, I really like these starter packs: Google DeepMind: go.bsky.app/GZ4hZzu Theoretical CS: go.bsky.app/6A6GRSi ML Theory: go.bsky.app/21nFz12 Differential Privacy: go.bsky.app/A7ABG83

I once attended a lecture by John Conway, who was asked about how to handle being scooped He answered something like: his first "discovery" (as a child maybe? IIRC, this was long ago) was scooped by Pythagoras, the second one by Gauss, and so on...

Surprised to discover that Quake-like text consoles still exist in Minecraft (with pretty complex commands!). Maybe the only non-PC game that still has that?

Great list!