ABOUT THIS EPISODE
John Giannandrea is the former Senior Vice President of Machine Learning and AI Strategy at Apple, and former head of Search and AI at Google. Earlier he worked at General Magic and Netscape, and was CTO of Tellme and of Metaweb, which Google acquired in 2010.
John joins Brad Porter, who worked for him at Tellme, to talk about ideas that arrive 20 years too early, the day he saw the web and quit his job, why skeptics are a good sign, and how to know when an AI system that is never 100% right is ready to ship.
The conversation ultimately points toward a future where the next leap in AI comes from better architectures and a deeper understanding of the models, and where the teams that win, in software or robots, set a launch bar, ship into real work, and keep climbing.
0:00 Intro
1:33 From Scotland to Silicon Valley: Inmos, Pixar, and SGI
4:15 General Magic: building the iPhone 15 years too early
5:18 Seeing Mosaic, quitting, and joining Netscape as employee #20
8:42 AOL buys Netscape and the move to Tellme
11:43 Open-sourcing the browser: the road to Firefox
13:32 Metaweb, Danny Hillis, and building the knowledge graph
17:54 How Metaweb disrupted Google Search
21:57 Google's AI second act and the $15K espresso machine
24:41 Leading Google Search: a policy job, not a technical one
27:23 Leaving Google for Apple: two opposite cultures
31:48 Hill climbing: how ML features get to launch
33:26 Beyond data and compute: the next architecture breakthrough
35:08 Physical AI and why more data isn't enough
36:56 Privacy, Apple Intelligence, and Private Cloud Compute
40:08 Shipping LLMs: Siri, Gemini, and ChatGPT's surprise hit
41:24 AI risk: what Geoffrey Hinton actually worries about
42:41 The Vatican cardinal and the "200-year experiment"
44:35 Was the internet a good thing? Building with humanistic intent
48:50 Young people, search, and the hunt for truth
50:52 AI is dual use: society will adapt and regulate
52:43 US vs. China and the global talent race
55:55 Overhyped short term, underappreciated long term
