Google's Challenges with AI Productization and Internal Bureaucracy

Key Insight

Google, despite its early lead in deep learning and declaration as an 'AI-first company' in 2016, evolved into a 'lumbering giant' that often impeded its own innovation. Its rapid expansion from 3000 employees in 2004 to over 135000 by 2020 fostered a fragmented organizational structure, described as a 'series of fiefdoms' with 'parallel tracks' rather than a unified 'tip of the spear.' This internal disorganization led to widespread duplication, blockages between teams, and a sluggish pace in bringing advanced AI projects to market, frustrating those pushing for faster adoption.

A core challenge was Google's reluctance to integrate cutting-edge AI into its most profitable products, particularly search, which generated over $40 billion in 2020. Despite a team demonstrating that DeepMind's technologies could improve Android battery optimization and the Google Play store's recommendation engine, a similar effort to enhance search results proved a 'dead end.' The search team, intensely conservative and wary of adding 'an undesired element of mystery and hocus pocus' to its algorithm, rejected the AI model after its results reportedly 'deteriorated over time,' prioritizing stability and existing profit streams over deeper AI integration.

This cautious approach actively suppressed Google's advanced conversational AI, Meena (later LaMDA), which Suleyman described as 'ChatGPT before ChatGPT.' Developed internally and trained on 341 gigabytes of social media data (nearly nine times OpenAI's GPT-2), Meena produced 'incredible, jaw-dropping' human-like dialogue and became 'absolutely viral inside Google' by mid-2020. However, management, 'more petrified than excited' by potential PR nightmares reminiscent of Microsoft's 'Nazi-saluting Hitler' chatbot Tay in 2016 and Google Photos' 'gorilla' mislabeling in 2015, deliberately downplayed LaMDA's capabilities at its 2021 developer conference, showcasing 'stupid' sample conversations to avoid making it 'sound like a person'.