Wednesday, 20 May 2026

Google just showed up at I/O 2026 with what feels like a significant shift in how they're thinking about AI assistance. They're moving Gemini from a chatbot you ask questions to into something closer to an agent that works on your behalf, even when you're not paying attention. That distinction matters more than it might sound.

The new Gemini app now proactively delivers daily briefs and handles tasks without waiting for explicit prompts. They've also introduced Gemini Spark, which is designed to help you get things done rather than just answer them. By the way, this isn't just UI polish — the architecture underneath has to change significantly to support genuine agency, which requires better planning, error recovery, and the ability to make reasonable decisions in ambiguous situations. The question isn't whether Gemini can now do more; it's whether Google has solved the reliability problem that makes agents actually trustworthy in autonomous mode.

Alongside the agent story, Google announced Gemini Omni, a new family of generative media models focused on creating images and video. The first model, Omni Flash, represents a consolidation of capabilities that previously lived in separate tools. They're positioning this as a "create anything" platform, which is the kind of marketing language I'm usually skeptical of, but the underlying idea makes sense — unifying image generation, editing, and restoration under one model family reduces friction for creators. What's genuinely interesting here is the practical implication: millions of people have already been using Nano Banana for photo restoration, so there's an existing user base that Omni can immediately serve with better tools.

Separately, Alibaba is pushing hard on both hardware and models. They've unveiled a more powerful Zhenwu AI chip and a new large language model, continuing their strategy of vertical integration in AI infrastructure. This matters because the chip race isn't just about raw performance anymore — it's about optimizing entire stacks for specific workloads, and Alibaba's willingness to invest in both silicon and software simultaneously gives them leverage that pure software companies don't have.

Here's what I find myself thinking about: the agent conversation is moving faster than the evaluation frameworks around agents. We have benchmarks for language models — they're imperfect, but they exist. Evaluating agents is harder because you're not just measuring capability, you're measuring safety, consistency, and judgment across hundreds of possible decision points. Google and others are shipping agentic features, which is right, but the industry needs to get serious about understanding when agents fail before they become load-bearing in critical workflows. That gap won't stay closed for long.