How DeepMind had to go back to its startup roots to become Google's AI "engine room"

Key Concepts

• Google DeepMind: The unified AI research entity within Google, encompassing Google Research, Google Brain, and DeepMind.
• Gemini Models: Google’s latest generation of large language models (LLMs). Specifically, “Gemini 2.5” is highlighted as a point of operational efficiency.
• AGI (Artificial General Intelligence): The long-term goal of creating AI with human-level cognitive abilities.
• Infrastructure Re-architecting: The process of redesigning Google’s underlying technological systems to support rapid AI deployment.
• "Sim-shipping": The near-simultaneous deployment of AI models across multiple Google products and services.

Google DeepMind: The Engine Room of AI Innovation

The speaker details the recent restructuring and operational focus of Google’s AI efforts, now consolidated under the umbrella of Google DeepMind. This entity represents the unification of Google Research, Google Brain, and DeepMind, functioning as the central hub for all AI research within the company. The speaker, who leads this group, describes it as “the engine room of Google,” responsible for developing all AI technologies and subsequently integrating them across Google’s diverse product portfolio.

Rapid Deployment & Infrastructure Overhaul

A key emphasis is placed on the significant effort undertaken over the past two to three years to adopt a more agile, “scrappier” approach reminiscent of a startup environment. This shift was driven by the need to accelerate innovation and deployment. Crucially, this wasn’t solely about model development; it involved a complete “rearchitecting of the entire infrastructure of Google.” This infrastructure overhaul was designed to facilitate the incredibly rapid deployment of new models – a process the speaker terms “sim-shipping.”

The speaker provides the Gemini models as a concrete example. New Gemini models are now being deployed to core Google surfaces, such as Search, “the next day or the same day” after release. This represents a substantial improvement in deployment speed and efficiency. The Gemini 2.5 models are specifically cited as having established a “smooth process” over the past year, a trend expected to continue and expand over the next 12 months.

Strategic Alignment & Daily Collaboration

The speaker highlights a close working relationship with Sundar Pichai, CEO of Google, stating they “pretty much talk every day about strategic things.” This daily communication ensures that AI development is closely aligned with the broader needs of Google, allowing for “adjust[ment] the road maps and the plans on a daily basis.” This dynamic planning process is balanced against the overarching, long-term objective of achieving AGI.

Long-Term Vision: AGI Development

The ultimate goal driving Google DeepMind’s work is the pursuit of AGI – achieving Artificial General Intelligence – but with a critical emphasis on safety. The speaker explicitly states the ambition is to “get to AGI first fast and safely,” underscoring the importance of responsible AI development alongside rapid progress.

Notable Quote

“You should think of [Google DeepMind] like that. So uh all the AI technologies is is done by this group by our group and then it's diffused across you know all of these incredible products uh right across Google.” – Speaker, describing the role of Google DeepMind within the larger Google ecosystem.