[0:00] Yesterday, Google I/O wrapped, and I was
[0:02] able to watch in person as Sundar and
[0:04] Demis laid out an ambitious vision for
[0:06] the future of software. And apparently,
[0:08] that future is Gemini hiding inside of
[0:10] every product like the microplastics in
[0:12] your bloodstream. But the road map is
[0:14] basically take Gemini, append a noun to
[0:16] it, and ship it. Gemini Spark, Gemini
[0:18] Omni, Gemini Flow, and the list goes on.
[0:21] But they're calling it the agentic
[0:23] Gemini era. The search is now an AI
[0:25] agent, Gmail is an AI agent, Android is
[0:27] an AI agent, your glasses are an AI
[0:29] agent. And as I watched the keynote, I
[0:31] realized something. That Google is no
[0:33] longer trying to organize the world's
[0:34] information with blue hyperlinks,
[0:36] because search engines are now an
[0:38] archaic technology. Instead, Google is
[0:40] trying to become the interface to
[0:42] reality itself before Anthropic and
[0:44] OpenAI create better realities. But
[0:46] luckily, Google I/O wasn't all about AI.
[0:49] I didn't see any updates about Angular,
[0:50] but I did come across a new awesome web
[0:52] API that every web developer should know
[0:54] about. In today's video, we'll break
[0:56] down everything you missed at Google
[0:57] I/O. It is May 22nd, 2026, and you're
[1:00] watching The Code Report. Whether you
[1:02] love it or hate it, one thing is
[1:03] undeniably impressive about Google, and
[1:05] that's its ability to scale. Not only is
[1:07] it serving its core products to billions
[1:09] of daily active users, but in the last 2
[1:11] years, they've gone from serving 9.7
[1:14] trillion tokens per month to a
[1:15] staggering 3.2 quadrillion tokens per
[1:18] month. And that number is going to
[1:19] continue accelerating. In addition,
[1:21] Alphabet's capital expenditures have
[1:23] exploded, building new infrastructure to
[1:25] support all these stupid AI images you
[1:27] guys create with nano banana. You ever
[1:30] see a pug dressed like an accountant?
[1:32] No.
[1:33] You want to? Uh
[1:35] One thing that makes this massive scale
[1:37] possible is Google's TPU chip, or Tensor
[1:39] Processing Unit. I remember being amazed
[1:42] seeing a TPU at my first Google I/O back
[1:44] in 2018. But this week, they announced
[1:46] they're splitting these chips into two
[1:48] distinct jobs, the training and
[1:50] inference with the TPU-T and TPU-I. In
[1:53] other words, Google now has one chip
[1:55] that's optimized to teach a robot how to
[1:57] think, and another chip that's optimized
[1:59] for it to hallucinate search results on
[2:00] a global scale. The headline
[2:02] announcement at Google I/O though was
[2:04] Gemini Omni, a model that takes any
[2:06] input like text, video, and sound and
[2:08] produces any output. Demis Hassabis, who
[2:11] might be the smartest guy at Google,
[2:12] appears to be fully world model pilled
[2:14] because models like this don't just
[2:16] generate pixels anymore. They understand
[2:18] language physics motion and
[2:20] everything else in your world just well
[2:22] enough to simulate reality on demand.
[2:24] But along with this new model comes an
[2:26] entirely new design system for the
[2:27] Gemini app called Neural Expressive. At
[2:30] first glance, the UI looks like a simple
[2:32] glow up with new icons and better
[2:33] gradients. But what's unique about it is
[2:35] that it's optimized for generating UI
[2:37] elements on demand, like diagrams,
[2:40] timelines, and even mini apps that
[2:41] didn't exist before your prompt. Now,
[2:43] when it comes to Google's core large
[2:45] language models, they released Gemini
[2:47] Flash 3.5, which is not the big brain
[2:49] model, but the fast model. According to
[2:51] the trust me bro benchmarks, it performs
[2:53] nearly on par with Opus 4.7 and GPT-5.5,
[2:57] but runs at a much faster speed. Like if
[2:59] we look at this trust me bro diagram, we
[3:01] see that Flash is entirely in a quadrant
[3:03] of its own in terms of speed and
[3:05] intelligence. However, it's important to
[3:07] remember that this is not their top-tier
[3:09] model. The Gemini 3.5 Pro is still under
[3:11] wraps and not expected to release until
[3:13] later this summer, which was very
[3:15] disappointing to a lot of people on the
[3:16] internet. Speaking of disappointment
[3:18] though, not everybody was happy with the
[3:20] new direction of Google's anti-gravity
[3:22] IDE. Anti-gravity was formerly known as
[3:24] Windserve and was code for AI coding
[3:27] just like Cursor. And once again,
[3:29] following in the footsteps of Cursor,
[3:30] its latest version looks like an OpenAI
[3:32] Codex clone that's more focused on
[3:34] managing agents than writing code. Old
[3:36] school programmers might not be happy
[3:38] with this change, but the live demo was
[3:40] pretty badass. They used the tool to
[3:41] build a complete operating system from
[3:43] scratch, which took like 12 hours and
[3:45] billions of tokens. But then, they tried
[3:47] to play Doom on it and it failed due to
[3:49] missing drivers. However, live on stage,
[3:52] they had Gemini code up those drivers
[3:54] and within a few seconds, Doom was up
[3:55] and running. The most impressive part
[3:57] was just the sheer speed at which this
[3:59] thing could spit out tokens. But, the
[4:00] speed is not the only thing increasing.
[4:02] But, the price of Gemini 3.5 Flash is
[4:04] three times more than the previous
[4:06] version and 30 times more than Gemini
[4:08] 1.5 Flash. It's still a lot cheaper than
[4:10] Claude, but not nearly as cheap as it
[4:12] used to be. Almost everything at IO
[4:13] involved AI in one way or another. But,
[4:15] if you're a web developer, one cool
[4:17] thing you should know about in Chrome is
[4:19] the HTML on Canvas API, which as the
[4:21] name implies, allows you to use HTML
[4:24] elements directly in a canvas now.
[4:26] >> Awesome. Native HTML elements rendered
[4:29] into the canvas.
[4:31] Woo!
[4:33] That means you can build highly
[4:34] interactive UIs where you control every
[4:36] pixel with tools like WebGL and WebGPU,
[4:39] while simultaneously using HTML for your
[4:42] more basic UI elements. The only
[4:44] question is which AI coding model should
[4:46] you use to work with this API? Well,
[4:48] that's why you need to know about
[4:49] Emergent, the sponsor of today's video.
[4:51] Everyone's switching between five
[4:52] different coding models these days, but
[4:54] we still need something to help us ship
[4:56] full stack applications that actually
[4:58] work. And that's exactly where Emergent
[4:59] can help. Right now, I'm using it to
[5:01] build a pull request review dashboard
[5:04] where I can paste in a GitHub link and
[5:06] get an AI summary of all the changes and
[5:08] risks per repo. You still start with a
[5:09] prompt, but instead of one LLM guessing
[5:12] how to build everything, Emergent spins
[5:14] up specialized agents to work on your
[5:16] app's front end, back end, database,
[5:18] testing, and deployment all in parallel.
[5:20] You also don't need to mess with any
[5:22] Superbase wiring or Express boilerplate,
[5:24] because that one prompt sets up your
[5:26] app's database, auth, and APIs. If
[5:28] you're really into self-torture, feel
[5:30] free to keep scaffolding this stuff by
[5:31] hand, or you could just describe the
[5:33] tool you want and let Emergent's agents
[5:35] swarm build it all for you. You try it
[5:37] out for free at the link below. This has
[5:39] been the Code Report. Thanks for
[5:40] watching and I will see you in the next
[5:42] one.