AcquiredGoogle Part III: The AI Company. Google is amazingly well-positioned... will they win in AI? (Audio)
CHAPTERS
- 0:58 – 6:19
Google’s Innovator’s Dilemma: AI threatens the Search cash machine
Ben frames Google’s core tension: the company invented key AI breakthroughs (including the Transformer) yet risks disrupting its own highly profitable Search business. The hosts set up the strategic question—lean into AI even if it cannibalizes ads, or protect margins and risk losing the future.
- •Innovator’s dilemma applied to a monopoly with massive margins (Search)
- •Transformer paper (2017) as the foundation for the current AI boom
- •Google’s unique asset stack: Gemini + Google Cloud + TPUs + talent
- •Strategic question: protect Search profits vs. win the AI platform shift
- 6:19 – 18:16
PageRank to PHIL: Google’s early language-model roots inside Search and Ads
The story rewinds to Google’s earliest AI mindset: Larry Page’s long-held belief that AI is the ‘ultimate version of Google.’ Early probabilistic language modeling work (Noam Shazeer, Georges Harik) becomes directly monetized through spelling correction and AdSense content understanding.
- •Larry Page’s early AI worldview and PageRank as statistical/AI-adjacent
- •‘Compressing data = understanding it’ as a proto-LLM intuition
- •Noam/Harik build ‘Did you mean…?’ from probabilistic language modeling
- •PHIL language model powers AdSense context matching, expanding ad inventory
- 18:16 – 23:56
Jeff Dean rewires Google Translate: making massive language models production-fast
Google Translate’s early success comes from Franz Och’s large n‑gram models trained on trillions of words—too slow for real-time use. Jeff Dean re-architects translation to run in parallel on Google’s distributed CPU infrastructure, cutting translation time from hours to milliseconds.
- •DARPA translation challenge pushes model scale (two trillion-word corpus)
- •Model initially takes ~12 hours to translate a sentence
- •Jeff Dean parallelizes the algorithm across Google’s distributed systems
- •Translate becomes a proof point that large language models can ship in products
- •Knock-on potential for query suggestions and ad quality prediction
- 23:56 – 37:26
Sebastian Thrun, Geoff Hinton, and the road to Google X
Sebastian Thrun joins Google and proves AI can reshape products via Maps ‘Ground Truth,’ replacing expensive third-party map data. His push to bring academics into Google leads to Geoff Hinton’s pivotal 2007 talk and eventually to the creation of Google X as a moonshot division.
- •Ground Truth rebuilds Maps data using Street View + many sources + human cleanup
- •‘Bring professors part-time’ becomes a talent pipeline strategy
- •Geoff Hinton enters Google orbit despite neural nets being fringe at the time
- •Deep learning becomes plausible as compute improves (Moore’s Law)
- •Google X forms as a home for bigger AI-driven bets
- 37:26 – 46:57
Google Brain and the ‘Cat Paper’: unsupervised learning becomes a business engine
Andrew Ng and Jeff Dean launch Google Brain with DistBelief, betting that deep networks can train at Google scale even asynchronously on distributed CPUs. The ‘cat paper’ proves large-scale unsupervised learning works and becomes a foundational technique for recommendation, ads, and content understanding.
- •DistBelief: distributed + asynchronous training that ‘shouldn’t work’ but does
- •Cat paper trains a 9-layer net on YouTube frames using 16,000 CPU cores
- •Unsupervised feature learning enables YouTube recommendations and classification
- •AI era for feeds effectively begins in 2012 (before ChatGPT)
- •Business impact: improvements compound across massive revenue bases
- 46:57 – 1:24:49
AlexNet + GPUs + DeepMind: the modern deep-learning arms race begins
AlexNet’s 2012 ImageNet leap shows GPUs are the unlock for deep learning, catalyzing NVIDIA’s trajectory and changing the field overnight. Google acquires DNNResearch after a bidding war, then makes the landmark DeepMind acquisition—positioning itself at the frontier of general intelligence research.
- •AlexNet cuts error rate dramatically by training on consumer NVIDIA GPUs
- •GPU parallelism becomes the new default for deep learning scaling
- •DNNResearch acquisition auction: Baidu vs Google (and DeepMind lurking)
- •DeepMind founded on ‘solve intelligence’ ambition and heavy compute needs
- •Google buys DeepMind for ~$550M with an independent oversight structure
- 1:24:49 – 1:30:45
DeepMind’s early wins: data-center cooling and AlphaGo as a public proof of AI power
DeepMind quickly delivers practical ROI inside Google (data-center cooling) while also showcasing breakthrough capability with AlphaGo. The AlphaGo match demonstrates creative, non-brute-force reasoning and resets public expectations about AI’s trajectory.
- •DeepMind cuts data-center cooling energy ~40% (rapid payback)
- •AlphaGo defeats Lee Sedol; ‘Move 37’ symbolizes emergent creativity
- •Go’s combinatorial complexity makes it a perfect non-brute-force benchmark
- •DeepMind’s game DNA connects back to reinforcement learning focus
- •Acquisition becomes an ‘all-time’ strategic asset for Google
- 1:30:45 – 1:40:09
Elon’s backlash and OpenAI’s founding: a new lab forms to counter Google’s gravity
Google’s DeepMind purchase enrages Elon Musk and helps trigger OpenAI’s formation with Sam Altman, framed as an open, nonprofit alternative to Google/Facebook control. Recruiting hinges on Ilya Sutskever’s willingness to leave Google—establishing the team that will later commercialize Transformers at scale.
- •2015 Rosewood dinner: ‘What would it take to leave Google?’ (mostly ‘nothing’)
- •Ilya Sutskever’s openness catalyzes OpenAI’s initial research team
- •OpenAI’s early mission: benefit humanity without financial constraint
- •Early funding pledges vs. actual cash collected; compute costs loom ahead
- •OpenAI initially mirrors DeepMind with game-focused research projects
- 1:40:09 – 1:51:21
Google’s hardware pivot: GPUs at scale, then TPUs to escape the NVIDIA tax
Even after deep learning proves itself, Google initially runs models on CPUs, then rapidly embraces GPUs—placing a huge early order that also signals AI’s future to NVIDIA. Compute costs explode with speech and other inference workloads, pushing Google to build the TPU and the TensorFlow ecosystem to scale efficiently.
- •Alex Krizhevsky finds Google is ‘a CPU shop’ and hacks in a GPU box
- •2014: Google orders ~40,000 GPUs (~$130M), approved by Larry Page
- •Inference demand implies doubling data centers (‘We need another Google’)
- •TPU created as a specialized matrix/tensor multiply chip; deployed in ~15 months
- •Design choices: reduced precision (quantization), hard-drive form factor, secrecy
- 1:51:21 – 2:03:57
‘Attention Is All You Need’: Transformers emerge—and Google underplays the platform shift
Google’s Translate work evolves from RNNs to LSTMs, but scaling and parallelization constraints push researchers toward attention-based architectures. Noam Shazeer’s rewrite unlocks the Transformer’s performance, and the paper becomes one of the most cited of the century—yet Google treats it as incremental rather than existential.
- •LSTMs improve Translate but are compute-heavy and hard to parallelize
- •Attention over full context mirrors how human translators work
- •Noam Shazeer rewrites implementation; Transformer begins to scale extremely well
- •Paper impact: ~173k citations; launches the modern scaling era (‘Bitter Lesson’)
- •Google builds models like BERT but doesn’t fully pivot the product platform
- 2:03:57 – 2:18:50
Elon exits, Microsoft enters: OpenAI finds the money and the cloud to scale GPT
Elon’s departure creates a funding crisis that accelerates OpenAI’s need for a commercial structure and a deep infrastructure partner. Reid Hoffman connects OpenAI to Satya Nadella, leading to Microsoft’s investment and the creation of the OpenAI LP structure—setting up the GPT-2/3/4 era and Azure distribution.
- •2017–2018: Elon ultimatum and exit removes key early funding source
- •OpenAI shifts toward large-scale pretraining (GPT-1 announced)
- •2018 Sun Valley meeting: Microsoft invests $1B (cash + Azure credits)
- •OpenAI LP created with nonprofit control; Microsoft gains exclusive licensing
- •Strategic fit: only a hyperscale cloud can supply the required compute
- 2:18:50 – 2:40:13
The ChatGPT shockwave and Google’s ‘Code Red’: Bard stumbles, urgency spikes
ChatGPT’s rapid adoption and Microsoft’s ‘new Bing’ force Google to view AI as disruptive rather than sustaining. Google rushes Bard to market with visible quality issues, while internal realities (trust, publisher relations, and ad economics) complicate the pivot away from ten blue links.
- •ChatGPT growth: 1M users in <1 week; 100M by Jan 2023
- •Google had internal chat (Meena/LaMDA) but held back on safety/trust risk
- •December 2022 ‘code red’ shifts priorities to shipping native AI products
- •Bard launch problems: weak RLHF/post-training; factual error in promo causes stock drop
- •Core tension: AI answers may cannibalize Search ads and publisher ecosystem
- 2:40:13 – 2:46:33
Unifying Google AI: DeepMind + Brain merge and the single-model bet on Gemini
Facing competitive pressure, Sundar merges Google Brain and DeepMind into Google DeepMind, elevating Demis Hassabis and reorganizing around a single flagship model. Gemini becomes both the internal platform and consumer brand, with rapid iteration, multimodality, and direct integration into Search via AI Overviews and AI Mode.
- •Strategic consolidation: end the ‘two lab’ model and unify culture/roadmap
- •Gemini as the one model for all modalities and all Google products
- •Key technical leadership: Jeff Dean + Noam Shazeer (after Character.ai deal)
- •Rollout milestones: AI Overviews, Gemini 1.5 (1M tokens), Gemini 2.x cadence
- •Distribution leverage: Google can run LLM inference at massive search scale
- 2:46:33 – 3:20:26
Waymo’s long road: from DARPA Challenge to commercial robotaxis
The episode detours to Alphabet’s standout AI-adjacent bet: autonomous driving. Originating in the DARPA Grand Challenge and Sebastian Thrun’s Stanford team, Project Chauffeur becomes Waymo—showing how quickly breakthroughs can happen and how long the final ‘edge cases’ take to productize.
- •DARPA Grand Challenge (2004–2005) as the talent seed for self-driving
- •Stanford’s software-first approach: fuse lidar with camera via ML
- •Project Chauffeur starts 2009; completes ‘Larry 1000’ in ~18 months
- •Commercialization is slow: perception, planning, safety, and ops complexity
- •Waymo’s scale today: multiple cities, millions of rides, large safety improvements
- 3:20:26 – 3:32:28
Google Cloud’s strategic role: distribution for models, leverage for TPUs, enterprise revival
Google Cloud evolves from opinionated App Engine to credible hyperscaler, especially after Thomas Kurian builds enterprise go-to-market. In the AI era, Cloud becomes the distribution channel for Gemini and the external vehicle for TPUs—giving Google an unusually complete ‘model + cloud + chip + apps’ stack.
- •Early misstep: App Engine PaaS vs. market demand for IaaS
- •2018 Kubernetes bet enables multi-cloud posture as the #3 provider
- •Thomas Kurian scales enterprise sales and flips Cloud to profitability
- •Cloud growth to ~$50B run rate; AI workloads amplify demand
- •Cloud makes TPUs externally usable—creating ecosystem and cost advantages
- 3:32:28 – 4:06:37
Bull vs. bear: can Google monetize AI without sacrificing Search?
The hosts weigh Google’s strengths—distribution, infrastructure, TPUs, data, and self-funding—against the central risk that AI may be harder to monetize than Search. They highlight how cost-of-tokens, ad models, and user intent may reshape unit economics and competitive shares in a multi-player market.
- •Bull: Google’s ‘full stack’ (model + chip + cloud + apps) and massive distribution
- •Bull: lower token costs via TPUs vs. paying the ‘NVIDIA tax’
- •Bull: personalized data from Google products could power differentiated AI
- •Bear: AI monetization unclear; high-value search queries may shift to chat
- •Bear: market share likely fragmented vs. Search’s ~90% dominance
