I was in Las Vegas two weeks ago for Google's Cloud Next 25. What few people realize about Next is the sheer caliber of the Googlers walking the floor, staffing the booths and giving the talks: when you walk up to any Google booth on the show floor, chances are the Googlers staffing that booth are the ones who actually BUILT that technology or are the top experts in it. Have a really niche deeply technical question about a Google technology? There is no better place on earth to get it answered than walking into the exhibit hall of Next and talking to the engineers and product teams who build, maintain and push it to its limits every day. Walk the floor of Next and you can pretty much meet at least one key technical engineer from every single one of GCP's major products in just a single afternoon.
While each of its major cloud technologies made an appearance somewhere on the floor or talks, AI was the dominant theme of this year's event, infusing itself into nearly every corner of the cloud stack.
At last year's I/O, a fascinating theme was how technologies that large model AI was supposed to replace, such as knowledge graphs and specialized AI models, were actually becoming newly important as cornerstones of the "large model" AI stack, augmenting rather than being replaced by large models. That theme was front and center at Next 25: large AI models were everywhere, but instead of a singular monolithic nexus into which all input and output flows and which encapsulate the entire pipeline powered by a textual prompt, those AI models were merely building blocks, used in multiple places and surrounded by myriad newly reinvigorated specialized models and tooling. In other words, far from the popular portrayal of building an app by typing up a textual description of what it should do and having the "model be the code", the reality of production AI today is that large models are merely acting as highly specialized software engines in more traditional kinds of processing workflows. This actually makes them far easier to work with and to incorporate into existing and new kinds of applications as opposed to some kind of entirely new processing metaphor. A newfound emphasis on cost and speed is also vastly broadening the universe of potential applications for large AI models, along with the move towards the edge and on-device inference.
The second dominant theme of Next 25 was that the future of AI lies in the applications that use it, rather than the raw technical stats of their size or capabilities. While an obvious observation, it is worth reflecting on just how much of the publicity and hype around the AI landscape over the last few years has focused on the former, rather than the latter, as if the precise number of parameters in a new model or the novelty of its reinforcement algorithm or its leaderboard benchmark score has any impact on how companies and the public adopt that model into their lives. AI only becomes useful when it is used to build killer apps that change how companies do business or improves people's lives. Google showcased a number of fascinating applications at Next built around Gemini and its surrounding technologies that focus on precisely that, from call center automation to "tell me about my day" summaries to website and ecommerce automation to code assistance to machine assisted everything. The common denominator of all of these is that Gemini sat at their core, but it was the user experience, surrounding tooling and vertical customization that transformed them from technical excellence into a potential killer app.