Google I/O 2024: Models in Every Product, and a Test for the Search Business

Originally published in Chinese on May 17, 2024. Google I/O is Google's developer conference, and this year's edition was about how Google brings its models into its existing products. These are my notes on how to read it. The view underneath the event To look at Google's event, I break it down along a few dimensions. Opportunities that come from the industry's direction: the event was not mainly a response to a challenge. It rests more on new opportunities that industry development has opened. Problems inside the core business: the core of what Google released addresses problems that need solving in its existing business system. So the event can be read as a gradual progression: model upgrades, the evolution of models in products and in service to users, and the near-term and medium-term product goals together with the software and hardware reserves built for them. The points that matter Keeping the models improving Google keeps accumulating data and refining algorithms to raise the underlying capability of its models, across voice, images and video. The final goal may be to solve multimodality inside the foundation model. The model systems released today are meant to let users and developers build better products on top of them, for more scenarios. Google will keep investing in foundation model capability without counting the cost, because it treats this as infrastructure, and in the end the cost is spread thin until it is negligible in the cost of service. Improving service inside its own products With the foundation model logic understood, the next element is improving how model capability serves inside Google's own product system. Google serves a huge base of consumer and enterprise users, and its product system is already strong and forms a closed loop. So Google only needs to raise AI or model capability inside that system to improve the user experience and the feedback it collects.