By far the most common work-related question I've been asked over Christmas is how a particular side dish (a porridge) tasted during the Nobel Banquet (it was okay). Not a single discussion about test-time scaling.

Few people use Google search because we hyped it or did fancy commercials. My goal for Gemini is the same.

The pricing for 2.5 Pro is out. Here is the pareto performance of price : lmsys elo visualized as a rainbow.

I hope everyone can forgive us for neglecting the sub 1200 Elo regime in our quest for Pareto optimality.

Introducing DolphinGemma, an LLM fine-tuned on many years of dolphin sound data 🐬 to help advance scientific discovery. We collaborated with Wild Dolphin Project to train a model that learns vocal patterns to predict what sound they might make next. It’s small enough (~400M params)

For instance, on the planet Earth, man had always assumed that he was more intelligent than dolphins because he had achieved so much—the wheel, New York, wars and so on—whilst all the dolphins had ever done was muck about in the water having a good time. But conversely, the

Competing with ourselves is getting a bit boring

Shipping ~1000 tokens per human on earth every day. Still lots of potential upside.

Oklart hur många svenska följare jag har, men likväl. Jag kommer vara i Almedalen 23-26 Juni för Googles räkning och prata om AlphaFold, Gemini och AGI i största allmänhet. Om någon vill ses, säg till! googlesessions.se/program

Pace. Focus. Intensity.

The bitter lesson of this decade seems to be that general methods always win in the end.

Dropping soon.

Well something did drop at least

I'm really excited about accelerating the sciences, but the most impressive part of this post is how much heavy lifting the duality of the word "found" can do.

EU: We've building a new AI Factory with 16 GPUs! Ready in 2030. Google: How 'bout a Dyson sphere?

Reports on the death of pre-training have indeed been greatly exaggerated.

What happened to adding error bars to evals?