Every year Google I/O feels like a pulse check on where AI is heading, and this year was no different.

In 2024, Google I/O was buzzing with AI, AI and AI. This time it was almost the same, but the focus was largely on models. Faster models, better performing models, cost efficient models. You can pull knobs in any direction to suit your need.

The big positive this year was seeing real world applications of the research investments announced in 2024. AI research is now starting to show up in usable products. Still, it feels early days for most consumer adoption. Developer products are where the real disruption has started, with Gemini claiming to be the fastest growing model across coding tools. Jules is there for coding too, but not sure yet what the real play is.

Consumer products are seeing a shift in user behaviour:

  • Google Search: AI overviews are driving more queries. At first that feels counterintuitive. If a search product does its job well, you should only need to search once. But curiosity pushes you deeper, and you end up searching more. I think that is mostly the case, which is why it was seen as a positive signal.
  • AI mode in Search: essentially a chat version of search. Not clear how this will differ from Gemini. The bigger question is how long before it becomes default, since that could hit the bottom line without an ads model on top.

Some of the future products felt almost unreal:

  1. Google Beam for 3D video calling. A hardware product with six cameras stitching videos into a 3D view. Futuristic, but not really AI. Likely an enterprise use-case unless it can be baked into a TV screen.
  2. Personalisation across the platform: smarter email replies using past emails, Gemini app tying into Drive and Calendar for suggestions, and search tuned to your own habits (eg. you like rooftop bars, so the next time you search “restaurants near me”, you get your preferences on top).
  3. Universal AI Assistant - to me, this could be the real game-changer. Powered through “world models”, with the ability to take action on everyday tasks. Summarising a YouTube video, making a table reservation, answering queries. The promise is big, but the risk is - it goes the Siri way, with demos that look impressive but never land in everyday use.
  4. Real time speech translation that mimics voice, tone and expressions.
  5. Gemini Live Search with camera and screen sharing.

In a world where models will eventually feel commoditised, Google’s real play is embedding intelligence into products that already have massive distribution. That is their moat. The big question is whether it’s an effective one or not. Do these demos translate into reliable capabilities that work across edge cases and actually get adopted and retained by users? If I want to book a table at my favourite restaurant and seven times out of ten it needs my intervention, I would rather just do it myself. Models will keep improving, but the bar is whether they can handle enough of these edge cases to feel dependable.

It almost feels like Google was pushed into bringing these products out earlier than planned to keep pace with the market because some of their new products lack that polish. Those initial bumps may be necessary since they create the real-world data that models need to become truly reliable. And if data from those feedback loops is what separates flashy demos from real capabilities, Google is better positioned than most to close that gap given their distribution. That’s the bull case for Google - leverage existing distribution to either launch completely new products or make existing products more useful and reliable. User experience is where the competition will play out.

“Gemini is pretty good at telling you when you are wrong,” Sundar Pichai, taking a jab at ChatGPT.