(4) Compute efficiency breakthroughs at training and inference. Stanford's work on benchmarks for CUDA kernel writing and "Cartridges" for 38x less memory consumption. DeepMind's DiLoCo enables distributed training with 500x less communication overhead - critical as models scale.

I will be presenting our work on decantralized training, looking forward to all the other great talks!

Great opportunity at Google Research for folks interested in AutoML, evolutionary methods, meta-learning, and open-endedness:  London: google.com/about/careers/… Zürich: google.com/about/careers/…

> AI Influencers say OpenAI has a "universal verifier" > ask if its just LLM as a judge > they don't understand > pull out Dec 2023 paper about LLM as a Judge > they laugh and say "its a universal verifier sir" > buy subscription to read article > it's LLM as a Judge

Our work on scaling laws for multimodal models and MoEs got an Oral at ICCV. Check it out !

What if you could not only watch a generated video, but explore it too? 🌐 Genie 3 is our groundbreaking world model that creates interactive, playable environments from a single text prompt. From photorealistic landscapes to fantasy realms, the possibilities are endless. 🧵

Genie3, Opus 4.1, and GPT-OSS released simultaneously. Feeling the acceleration, anon?

we live in the best era, it's just insane

HOLY SHIT. Recursion reached with Google Genie 3 The Matrix is becoming real... Watch this video until the end! 👀

DiLoCo is a distributed-optimization method for training LLMs across slow or geographically separated networks. Each worker runs many local AdamW steps on its own data; only every ~500 steps do the workers send compact “pseudo-gradients” to a global Nesterov-momentum optimizer,

One word: relentless. just in the past two weeks, we’ve shipped: 🌐 Genie 3 - the most advanced world simulator ever 🤔 Gemini 2.5 Pro Deep Think available to Ultra subs 🎓 Gemini Pro free for uni students & $1B for US ed 🌍 AlphaEarth - a geospatial model of the entire planet

AGI won’t exist without the best infra possible. I strongly recommend to apply to PI

The problem with "understanding" in deep learning, is that it actually rarely hold in practice, as it requires way too much simplifying assumptions. I'd rather have a SotA model with some intuitions why it works than a toy model with perfect understanding 🤷

Imagine the rhythmic noise of an entire datacenter performing forward-backward & AR across thousands GPUs. A sight to behold, the datacenter is breathing!

Absolute cinema