Google I/O: Oops, All Gemini!

Summary of Google I/O: Oops, All Gemini!

by MKBHD

2h 1m•May 22, 2026
Listen to Episode

Overview of Google I/O: Oops, All Gemini!

This episode is a broad, highly opinionated rundown of Google I/O and the week’s other tech news, with the hosts arguing that Google’s event was overwhelmingly packed with Gemini-branded AI features, confusing product names, and a lot of demos that felt more like hype than useful consumer advances. The discussion also covers Sony’s badly received AI photo-editing promo, the real-world limits of smartphone image processing, a surprisingly good Google Messages feature called Magic Cue, and a company announcement about sponsoring Marble League teams.

Main takeaways

  • Google I/O was framed as “all Gemini, all the time.”
    The hosts felt Google split Android into its own earlier event, then filled I/O with a flood of AI demos and products that were difficult to track and often overlapped.

  • The recurring critique was image processing, not raw camera hardware.
    Both Sony and Google showed examples where “AI enhancement” made photos look flatter, brighter, and less natural. The episode argues that many phone makers are still using overly aggressive processing that crushes highlights, lifts shadows, and removes the kind of tonal depth that makes a photo look good.

  • Google’s AI strategy feels increasingly tied to ads and commerce.
    Features like Universal Cart, AP2, and smarter shopping/search flows were presented as convenience tools, but the hosts repeatedly framed them as a way for Google to turn AI usage into new revenue streams.

  • The most useful AI demos were the narrow, practical ones.
    Examples like Magic Cue in Google Messages, better search suggestions, and certain agentic workflows were presented as genuinely helpful. The broad “AI will do everything” demos were viewed much more skeptically.

Google I/O and the Gemini overload

New models and AI infrastructure

  • Gemini 3.5 Flash was positioned as fast, cheap, and strong enough for many real-time use cases.
  • Gemini 3.5 Pro was teased as a later release.
  • Gemini Omni was described as a “world model” aimed at multimodal understanding and video generation, though the hosts questioned why Google needed yet another separate branded model.

Agentic AI products

  • Antigravity 2.0: Google’s coding/agent environment, which the hosts found confusing because the product family includes multiple overlapping tools with the same name.
  • Gemini Spark: A 24/7 background assistant meant to monitor things like calendars, subscriptions, and credit cards, then give you digests or take actions for you.
  • The episode repeatedly returned to the idea that Google is trying to make agents do more of the “repeated checking” that people currently do manually.

Retail, shopping, and payments

  • Universal Cart lets users track products across Google surfaces and potentially buy them when prices drop.
  • AP2 (Agent Payments Protocol) is Google’s attempt to make agent-driven shopping safe enough to work with credit cards.
  • The hosts were very clear that this is also about turning AI behavior into monetizable commerce and making Google more central to purchases.

Search and apps

  • AI Mode/Search upgrades now support more natural-language queries and dynamic, custom-generated layouts.
  • Google showed off search results that could become mini apps or custom visual experiences, like trip planners.
  • AI Studio can generate Android apps from prompts and potentially publish them, which the hosts think could create a flood of low-quality “slop apps.”
  • Google Pics / Photos editing was mocked for being essentially an “auto” button with a fresh AI label.

The photo-processing rant: Sony and Google both got dragged

Sony’s AI camera assistant

  • Sony posted examples of an “AI camera assistant” on an Xperia flagship phone.
  • The examples looked washed out, over-brightened, over-sharpened, and generally worse than the originals.
  • Sony later clarified that the feature does not auto-edit after capture; it suggests creative settings instead.
  • The hosts still thought the promo was baffling because the examples made the results look dramatically worse.

Why the hosts hate this style of processing

  • They argued that smartphone makers keep making photos:
    • brighter,
    • flatter,
    • more contrast-limited,
    • more aggressively sharpened,
    • and less expressive.
  • The episode spends a lot of time on:
    • histograms
    • tone curves
    • shadow/highlight clipping
    • why good photos need tonal range
  • Their core argument: a good photo is not just “more visible.”
    Many consumers prefer brighter images because they feel clearer, but that often destroys the depth and realism that makes a photo actually look good.

The bigger philosophical point

  • The hosts argued that smartphone companies often optimize for utility photos:
    • signs,
    • menus,
    • receipts,
    • kid photos,
    • quick reference shots.
  • That’s why a lot of consumer camera processing trends toward “make everything readable,” even when it hurts artistic quality.

Useful versus annoying AI

What they thought was genuinely good

  • Magic Cue in Google Messages
    A text from a mutual friend triggered a Google Photos suggestion that found the exact old pictures being discussed. The hosts called this a “hell yeah, they tested this” kind of success.
  • Agentic reminders / background tasks
    Things like monitoring a price drop or checking a long-running recovery job make sense for agents.
  • Search improvements for natural language
    Because people increasingly search the way they actually talk, not like SEO robots.

What they thought was overhyped or weird

  • Rebranding ordinary improvements as brand-new AI breakthroughs.
  • Building mini apps inside search instead of just returning clear information.
  • AI-generated or AI-enhanced image tools that make simple tasks more confusing.
  • The sheer number of duplicate, overlapping product names.

AI watermarking and authenticity

  • Google’s SynthID watermarking and C2PA/content credentials got praise as necessary tools for identifying AI-generated content.
  • The hosts pointed out the irony: Google is making the tools that can generate misleading content and the tools that can help detect it.
  • They also stressed that detection only works if platforms actually surface the metadata in a way people see during normal browsing.

Smart glasses and the privacy problem

  • Google’s new audio glasses were basically framed as Google’s answer to Meta Ray-Bans.
  • The hosts liked some of the demos:
    • asking the glasses where a place was,
    • pulling in location-aware suggestions,
    • ordering a drink ahead of time.
  • But they were worried about:
    • unwanted recording
    • lack of clear recording indicators
    • people being filmed without consent
    • the broader creepiness of always-on wearable AI
  • They also predicted that once smart glasses spread across the Android ecosystem, public backlash around recording could get much worse.

Broad critique of Google’s AI strategy

What the hosts think Google is doing

  • Trying to create usage for AI where demand doesn’t naturally exist.
  • Turning AI into a way to replace search behavior, not just improve it.
  • Using AI to gather more data, reduce friction, and ultimately help advertisers sell more effectively.

Their biggest concerns

  • Naming chaos: too many Gemini-branded products, too many overlapping tools.
  • Trust issues: the more Google knows, the better the results can be, but the more intrusive it becomes.
  • AI fatigue: even when a tool is useful, it gets mentally lumped in with all the slop.
  • Opportunity cost: useful work like weather models or AlphaFold gets buried under consumer-facing hype.

Other highlights

“Did they even test this”

  • A standout positive example: Google Messages surfaced a useful Google Photos result based on a text conversation.
  • The hosts treated this as a rare example of AI doing exactly what it was supposed to do.

Wear OS 7

  • Touched on in the second half of the episode, though it was only announced in a developer blog.
  • Main themes:
    • better widgets and card layouts,
    • live updates,
    • task automation,
    • richer workout tracking,
    • improved media controls,
    • better watch face tooling.

New compact camera mention

  • A quick aside about a new Panasonic Lumix compact camera with a clever tri-flap lens cover.
  • It was mostly a novelty mention, not a deep review.

Company announcement: Marble League sponsorship

  • The episode ends with a fun reveal: the company is sponsoring multiple Marble League / Jelles Marble Runs teams across channels.
  • Teams mentioned:
    • Waveform → Solar Flares
    • Autofocus → Plasma
    • Studio → Primary
    • MKBHD → Black Jacks
  • The hosts pitched it as one of the coolest company things they’ve done and encouraged listeners to follow the marble season.

Bottom line

This episode is equal parts tech news recap and long-form critique of how AI is being positioned by Google and other phone companies. The hosts are not anti-AI; they’re skeptical of AI as a label for mediocre product design, worse photos, and aggressive monetization. The strongest praise goes to narrow, useful features that solve real problems. The strongest criticism is reserved for demos that feel like they’re trying to manufacture demand rather than earn trust.