How Google Rebuilt for the AI Era
Google faces the greatest innovator's dilemma in history. They invented the Transformer — the breakthrough technology powering every modern AI system from ChatGPT to Claude (and, of course, Gemini). They employed nearly all the top AI talent: Ilya Sutskever, Geoff Hinton, Demis Hassabis, Dario Amodei — more or less everyone who leads modern AI worked at Google circa 2014. They built the best dedicated AI infrastructure (TPUs!) and deployed AI at massive scale years before anyone else. And yet... the launch of ChatGPT in November 2022 caught them completely flat-footed. How on earth did the greatest business in history wind up playing catch-up to a nonprofit-turned-startup
Today we tell the complete story of Google's 20+ year AI journey: from their first tiny language model in 2001 through the creation Google Brain, the birth of the transformer, the talent exodus to OpenAI (sparked by Elon Musk's fury over Google’s DeepMind acquisition), and their current all-hands-on-deck response with Gemini. And oh yeah — a little business called Waymo that went from crazy moonshot idea to doing more rides than Lyft in San Francisco, potentially building another Google-sized business within Google. This is the story of how the world's greatest business faces its greatest test: can they disrupt themselves without losing their $140B annual profit-generating machine in Search?
More on the episode
Acquired’s 10th Anniversary Celebration!
- When: October 20th, 4:00 PM PT
- Who: All of you!
- Where: https://us02web.zoom.us/j/84061500817?pwd=opmlJrbtOAen4YOTGmPlNbrOMLI8oo.1
Links
- Sign up for email updates and vote on future episodes!
- Geoff Hinton’s 2007 Tech Talk at Google
- Our recent ACQ2 episode with Tobi Lutke
- Worldly Partners’ Multi-Decade Alphabet Study
- In the Plex
- Supremecy
- Genius Makers
- All episode sources
Carve Outs
corrections
- In talking about Google Cloud, we said that Kubernetes launched in 2018… it was actually 2014!
- We stated that AlphaGo ran on 4 TPUs — the actual AlphaGo match ran on 48 TPUs across 2 racks (24 each). Later on they ran AlphaGo Zero on 4 TPUs, but that was after the Lee Sedol match.
- We actually significantly understated Google’s TPU cost advantage vs GPU-based AI infrastructure! We correctly noted that Google’s chip hardware partner Broadcom only charges a ~2x markup to Google on TPUs vs Nvidia's ~80% margin (i.e. ~5x markup) on GPU systems. However the chip itself is only one component (roughly ~25%) of an overall TPU/GPU system cost… and Nvidia’s 80% margin applies to the ENTIRE system! (Nvidia does not sell standalone GPU chips, only full solutions.) So not only is Google getting TPU chips much cheaper than GPUs, they’re also getting huge savings across the board on their AI systems as compared to the rest of the industry.