Sitting in the front. I usually do this when I'm riding alone (more comfortable).

Get well soon!

“The Party told you to reject the evidence of your eyes and ears. It was their final, most essential command.”

shapingai.com points to the Arxiv version, the CACM version, as well as a short op-ed from The Economist by our co-author Dave Patterson about the work, and NotebookLM-generated audio podcasts from the Arxiv version.

PDF link for CACM version is a bit hard to find:
dl.acm.org/doi/pdf/10.1...

It was a delight to work on this with an awesome set of co-authors: Mariano-Florentino Cuéllar, Finale Doshi-Velez, John Hennessy, Andy Konwinski, Sanmi Koyejo, Pelonomi Moiloa, Emma Pierson, and David Patterson.

The CACM version is shorter version of Arxiv version. We cover AI's potential impact on employment, education, healthcare, (mis)information, media, national security, and science, and propose 18 potential moonshot research directions in these areas to maximize positive impact and minimize downsides.

Good catch. Fixing and will roll out with next update. Thanks.

The internal version also has additional examples in code that is more sensitive (e.g. changes to the TPU compiler backend) that we aren't able to include in the external version.

The public version is a somewhat sanitized version of the internal version. People at Google can find the internal version at go/performance-hints, which has links to the actual changelist in our source code repository system.

Sanjay and I also had fun in writing this, plumbing the depths of changelists over multiple decades looking for interesting techniques we'd used and trying to group them together into higher level themes.

Many techniques are probably known to readers, but we felt it was nice to have a collection of tips and tricks, ranging from high-level algorithmic improvements to low-level performancer optimizations collected together.

Others are specific small snippets of code that demonstrate a technique.

We've tried to give concrete examples of the various techniques. Some are high level descriptions of a set of performance improvements, like this set of changes from working on Google's search system in 2001.

The Gemini 3 Flash and Pro models are a huge team effort. Thank you to everyone who worked very hard over the past months to train and get these models ready for release! 🎉

We’re all very excited to see what people do with the new capabilities offered by the Gemini 3 models!

Gemini 3 Flash is available now in the API, and rolling out today as the default model in AI Mode in Search and Gemini app globally.

Not only is Gemini 3 Flash higher quality than 2.5 Pro it is also 3x faster (based on Artificial Analysis benchmarking) at a fraction of the cost. Here’s a side-by-side demo.