AI
Embedding Models as Generative AI’s Companion
Gemini Embeddings 2 is fully multimodal, derived from Gemini itself. A single vector encodes up to 8,000 tokens of text, 128 seconds of video, 80 seconds of audio, and a full PDF. That end-to-end representation avoids information loss from stitching modalities after the fact. The model uses Matryoshka Representation Learning, so retrieval can start at 256 dimensions and expand for more precision. Hadsell traces the design intuition to the Jennifer Aniston neuron: a handful of cells that fire for one person regardless of whether you hear a name, see a photo, or hear a voice.
“generate sometimes we want to retrieve.” — Raia Hadsell
Neural Networks Beat Physics-Based Weather Forecasting
GraphCast, a spherical graph neural network trained on 40 years of global data, predicted Hurricane Lee’s Nova Scotia landfall 9 days out. The best physics-based models topped out at 6. Three extra days of lead time before a major hurricane makes landfall is operationally significant. GenCast, the follow-on probabilistic model, beat 97% of 1,300 gold-standard weather benchmarks and produces a 15-day forecast
“forecast in 8 minutes on a single chip” — Raia Hadsell
instead of hours on a supercomputer. FGN, the latest release, folds cyclone trajectory, wind speed, and eye formation directly into the network. The National Hurricane Center now uses it.
Interactive World Models That Remember
Genie 1 held a 2D platformer world alive for seconds. Genie 2 ran in 3D but not in real time. Genie 3 runs interactively, holds spatial memory (leave a spot, return a minute later, everything is where it was), and accepts mid-session prompts that shift the entire world around you. A demo shows a Kent lane, a ski slope, and an origami lizard environment, all from a text prompt or a short video fragment. An artist contributed a few seconds of hand-crafted animation; Genie 3 turned that fragment into a flyable world.
“I love the idea of a new form of gaming” — Raia Hadsell
Notable Quotes
generate sometimes we want to retrieve. Raia Hadsell · ▶ 7:32
forecast in 8 minutes on a single chip Raia Hadsell · ▶ 13:29
I love the idea of a new form of gaming Raia Hadsell · ▶ 19:33
Key Takeaways
- Gemini Embeddings 2 encodes text, video, audio, and PDFs into one vector using Matryoshka Representation Learning.
- GenCast outperformed 97% of 1,300 gold-standard weather forecasts and produces a 15-day forecast in 8 minutes on one chip.
- Genie 3 generates interactive 3D worlds from a text prompt and maintains spatial memory across extended sessions.