The history and future of AI at Google, with Sundar Pichai

Sundar Pichai clarifies the origins of Transformers at Google, noting they were developed to solve specific product needs like translation and inference for speech recognition. He highlights their immediate application in Google Search through BERT and MUM for significant quality jumps. Pichai also discusses LaMDA, Google's early conversational AI, which predated ChatGPT but faced internal concerns about toxicity and a higher product quality bar before public release.

Pichai emphasizes speed, or low latency, as a core distinguishing feature of Google's products, from early Search to Gmail and Chrome and now Gemini on TPUs. He explains how Google rigorously manages latency across its teams. Looking ahead, he envisions Search evolving into an 'agent manager,' where it will complete tasks and manage multiple threads for users, moving beyond simple information retrieval to a more agentic and asynchronous experience.

Pichai addresses the negative investor sentiment Google faced a year prior, when many believed Search was under attack. He explains that Google's long-term foresight, including early investments in TPUs (since 2016) and an 'AI-first' company culture, positioned it uniquely for the AI shift. He highlights Google's full-stack approach, from research to infrastructure and platforms, allowing AI to accelerate across all its businesses, from Search to YouTube to Waymo, viewing it as an expansionary, not zero-sum, moment.

Pichai discusses the perception that Google might be less 'AGI-pilled' than other labs, attributing it to semantics and Google's broad product portfolio. He asserts that Google's founders and current leadership have always deeply understood the potential of AGI, pointing to significant CapEx investments as proof. He shares personal 'AGI moments,' from recognizing cats in neural networks in 2012 to experiencing the current speed of AI in coding and agentic workflows, emphasizing the surprising slope of AI's progress.

Pichai describes his methods for staying connected to user experience, including dogfooding internal versions of products and dedicating time to intensive use. He also leverages platforms like X for raw user feedback. On AI's economic impact, he believes it will significantly boost GDP, noting that despite high CapEx, demand for AI remains supply-constrained. He emphasizes AI's potential to expand various markets, such as software engineering, and foresees substantial growth despite societal dampening mechanisms and the need for responsible diffusion.

Pichai details the critical supply constraints hindering AI growth, including wafer capacity, power and energy infrastructure, permitting, and critical components like memory. He notes that while high prices will eventually spur supply, these bottlenecks will shape the market, potentially enforcing an oligopoly. He also touches on other constraints like security, where AI's ability to break software poses new challenges, requiring enhanced coordination and potentially causing system shocks.

Pichai highlights Google's long-term, high-risk projects, comparing them to early Waymo. He mentions exploring data centers in space and advancing quantum computing, believing quantum systems will be crucial for simulating nature. He also expresses excitement about robotics, where AI is now the missing ingredient for significant progress, and projects like Wing (drone delivery) and Isomorphic (AI for drug discovery) as examples of methodical compounding. He emphasizes Google's commitment to these moonshot endeavors.

Pichai describes Google's capital allocation process, which involves comparing highly heterogeneous projects with vastly different payoff curves. He explains that early, smaller funding for deep technology bets allows for long-term commitment. He assesses projects like Quantum based on underlying technological progress rather than immediate returns. This approach involves intuitively weighing the option value and total addressable market (TAM) years down the line, exemplified by continuous investment in TPUs and increased backing for Waymo during periods of market pessimism.

Pichai discusses the shift in R&D budgeting, where compute resources (TPUs and GPUs) now rival headcount in importance, especially for ML. Google implements rigorous compute planning and allocation across projects and teams, with Pichai personally dedicating an hour weekly to granular oversight. He also explains how Google Cloud manages its compute resources, forward-planning to meet customer commitments despite systemic constraints, viewing these as opportunities for innovation and efficiency.

Pichai envisions AI serving as an orchestration layer, making complex functionalities more accessible and enabling stateful AI for consumers. He highlights Google Cloud's programmatic interaction with AI as an example of this, simplifying navigation of vast service offerings. He confirms that Google is actively working on bringing persistent, long-running agentic tasks to mainstream users. Internally, Google is seeing profound shifts in workflows, particularly within Google DeepMind and SWE teams, using tools like 'Jet Ski' (Antigravity) to navigate agent-managed environments.

Pichai identifies several barriers to broader AI diffusion, including the skill required for effective prompting, challenges in sharing AI-generated code, and issues with data access and permissions within large organizations. He emphasizes Google's commitment to solving these problems, particularly in security and robust implementation, to enable a significant leap in AI capabilities. He predicts that 2027 will be a crucial inflection point, with profound shifts in non-engineering processes, acknowledging that startups may have an advantage in adopting AI-native workflows more quickly than large, established companies.

Pichai shares his excitement about small, early-stage initiatives within Google that hold significant long-term potential. He specifically mentions the project to develop data centers in space, which began with a small team and budget, as an example of starting small with a big idea. He also refers to recent breakthroughs in post-training machine learning, highlighting the constant innovation occurring at a granular level within the company.