AI Daily News April 10th 2025: ☁️Big AI Day at Google Cloud Next 2025 🤝Google's Protocol for AI Agent Collaboration 💰Anthropic Introduces $200/Month Claude Max Subscription 🏠Samsung’s Gemini-Powered Ballie Home Robot 💥Google Unveils Trillium TPU
12:48 Share

Welcome to a new deep dive from AI Unraveled. This is the podcast created and produced by Etienne Newman, who's a senior software engineer and, believe it or not, a passionate soccer dad up in Canada. And hey, if you're enjoying these deep dives, if you're getting something out of them, please take just a second to like and subscribe on Apple Podcasts. It really helps us out. So you've pulled together a really fascinating collection of updates for us today, April 10th, 2025. We've got, well,

Everything. New AI chips, home robots, some pretty serious legal battles, and even AI showing up in courtrooms. Our mission, as always, is to unpack all this, highlight what really matters, connect the dots.

So you leave understanding what's happening and why it's important. Where should we start? I'm thinking maybe the foundation, the hardware, what's going on with the chips. Absolutely. Yeah, that makes sense. It all starts there, right? With the infrastructure that makes all this AI possible. And that NVIDIA story about the H20 chips for China, that definitely jumped out. Yeah, that sounded like a bit of a sudden turn. The export control is getting paused.

After Jensen Huan met with President Trump. Exactly. It's a really stark example of tech meeting geopolitics head on. It really highlights the balancing act NVIDIA has to do, doesn't it? It really does. I mean, they want to sell everywhere. China's a huge market. But then you have the U.S. national security side and the fact that this pause came with apparently a pledge for more U.S.-based AI investment.

Well, it shows how delicate that balance is. Right. And this is happening just before that May 15th AI diffusion rule hits.

That sounds like it could be significant for them. Oh, potentially huge. The numbers I saw suggested something like $16 billion in potential H20 sales to China could be on the line. $16 billion. Wow. Yeah. So this temporary pause, it just underscores the massive stakes, that constant tension between business and national security in the AI hardware world. It's a good reminder, isn't it, that AI progress isn't just about code. It's tied up in global economics politics.

OK, so shifting gears a bit. Google's also making noise with their own silicon, right? Tell me about these new TPUs Trillium and Ironwood. They sound powerful. They really are. Trillium first. That's their sixth gen TPU, their tensor processing unit. Custom chip.

And it's a big leap. We're talking like a 4.7 times increase in peak compute performance over the last one, TPU v5e. Almost five times. What does that actually mean, practically speaking? More power for what? Well, think bigger models, training much, much larger, more complex AI models using way more data. So potentially bigger breakthroughs in language, vision, you name it. Plus, and this is important, it's also about 67% more energy efficient. Oh, that's crucial. The energy footprint of AI is definitely a growing concern. So faster and more efficient.

That's the goal. A win-win, like you said. And it's got other upgrades too. Better matrix math units, which record AI, faster clock speeds, double the memory, double the interconnect bandwidth so the chips can talk faster. It's all geared towards handling these huge AI jobs. Okay. Impressive. And then Ironwood, the seventh gen. The performance numbers on that one seem almost, well, astronomical. Yeah. Ironwood is pushing boundaries. Yeah.

42.5 exaflops, just a mind-boggling amount of computation. And exaflop is what, a billion, billion calculations per second? I can hardly even wrap your head around. Right. And the chip itself, 4,614 TFLOPs peak performance, 192 gigs of RAM. And they highlighted something called SparseCore, which is apparently optimized for ranking and recommendation tasks.

So that tells you where Google sees a big need for this power. Think search results, YouTube suggestions. Right. So this is Google really planting its flag in the AI accelerator market, taking on NVIDIA, but also Amazon's chips, Microsoft's efforts. Exactly. It's super competitive. And integrating Ironwood into their AI hypercomputer in Google Cloud, that's a clear signal. They want to be the place for businesses doing serious AI. And Trillium, with the ability to link up

256 TPUs in one pod and its handling of embeddings, that looks tailor-made for training the next generation of giant models, maybe Gemini 2.0 or whatever comes next. Speaking of Gemini, Google Cloud Next had other big news too, right? Updates to Gemini itself and this Agent 2 agent thing.

A2A. Yeah, Cloud Next was definitely packed. So Gemini 2.5 and Gemini 2.5 Flash are the new model updates. The big thing there seems to be the expanded context window. Context window. Okay, break that down for us. What's the simple version? Think of it like the AI's short-term memorization.

How much information it can hold and like actively consider when it's generating something. A bigger window means it can handle much longer documents, follow longer conversations, understand bigger chunks of code. It leads to smarter, more relevant answers. Gotcha. And the Flash version is just faster. Yeah, lower latency. So quicker responses, which is vital for chatbots or anything real time. Makes sense. Okay. Now this A2A protocol, AI agents from different companies talking to each other. That sounds...

potentially huge. It really could be. That's the core idea. A standardized way for AI agents, even if they're built by different people on different systems to communicate and work together. And the fact that they already have like over 50 partners signed up, big names like Atlassian, ServiceNow, Workday, that suggests there's real interest. So

So wait, an agent from say Workday could talk to an agent from ServiceNow to complete a task. Yeah. Without needing special integration. That's the vision. They could discover each other's capabilities, assign tasks, share

share info, all without needing shared memory or being built the same way. It's about interoperability. Wow. That sounds like it can unlock some really complex workflows. Exactly. They mentioned it complements Anthropix MCP, model control plane. A2A is more about agent-to-agent interaction. MCP is agent-to-tools. Think about hiring someone. You could have one agent...

finding candidates, talking to another, doing background checks, another scheduling interviews seamlessly across platforms. That's a powerful concept. Yeah. And speaking of Anthropic, they also had news about pricing for Claude, their AI, new subscription tiers. Yeah, they're adding more options. A new top tier called Max for $200 a month. It offers a

to 20 times the usage limits of their pro plan. 20 times? Who needs that much? Power users, basically. People doing really heavy lifting, analyzing massive data sets, writing super long reports, building complex things at the API, intensive stuff. And there's a mid-tier too, $100 for five times the pro limit. So different levels for different needs. Okay. It feels like we're seeing this across the board now, right?

tiered pricing based on usage like open AI does. Absolutely. It reflects the the real compute costs behind running these big models. And it shows the market is maturing. Companies are segmenting their users. Now, let's bring AI out of the cloud and into the living room. Samsung's Bally Home robot.

Powered by Gemini. I feel like I remember seeing this thing years ago. You probably did. Yeah. Bally was first shown off, I think, back in 2020. So it's been around as a concept. But integrating Google's Gemini AI, that's the new big step, makes it potentially much smarter. So what can it do? A rolling ball, right? Yeah, basically a small rolling robot. But the idea is it can interact naturally, control your smart home stuff, project video onto walls, even offer personalized help.

personalized help. Like what? The examples were things like fashion advice based on what's in your closet or adjusting the lighting and sound for optimal sleep. It goes beyond just basic commands. Interesting. So it's using Gemini's intelligence plus Samsung's own AI for sensing things. Exactly. Combining vision, voice, sound processing with Gemini's reasoning. It moves on its own, takes voice commands. It's trying to be that helpful household robot we've seen in sci-fi. And it's actually coming out soon. Planned launches this summer. U.S. and South Korea first.

And they mentioned future third party app support, which could make it even more capable. OK, that's definitely one to watch. Feels like that sci fi future is

Slowly arriving. All right. Before we get too carried away with robots, we have to touch on the legal and ethical side. There's always something happening there. The OpenAI versus Elon Musk fight took another turn. Oh, yeah. It's heating up again. OpenAI filed a countersuit against Musk. Countersuit. What are they alleging? Things like unfair competition, interfering with their business relationships. And they brought up an alleged $97.4 billion takeover bid from Musk, calling it deceptive. Wow.

And I saw something about internal emails suggesting Musk wanted open air to be for profit under his control way back in 2017. Exactly. That seems to be a key part of their argument, directly challenging Musk's current narrative about, you know, open air betraying its nonprofit mission. That certainly complicates his position. It does. It adds layers. And with the jury trial, possibly in 2026, maybe an expedited one sooner.

This is going to be a long, fascinating, high stakes battle, especially with open AIs. What, $300 billion valuation now? Incredible number. Okay. Switching from corporate battles to...

The courtroom itself, a lawyer using an A.I. avatar in New York. How did that go down? Not well, not well at all. The judges apparently reacted very negatively, called it deceptive. Yikes. Yeah. It just highlights how unprepared the legal system is for this stuff. Right. There are no clear rules or guidelines for using generative A.I. in court. Raises huge questions about authenticity, misrepresentation. It really underlines the need for rules.

Clear boundaries for AI in the justice system and fast. Definitely. The tech is moving way faster than the regulations. On a related note, that No Fakes Act is back, trying to regulate deep fakes and protect voice and likeness. Yes, it's been reintroduced. And what's interesting now is the level of industry backing. YouTube, Universal Music, even OpenAI are supporting it. Oh, well, even OpenAI. Wow.

Yeah. Plus, it has bipartisan support in Congress. That suggests there's momentum building for some kind of federal regulation around synthetic media. It might actually happen this time. Okay. That's significant. And one last thing on the slightly stranger side of AI. Google paying AI staff not to work.

to keep them from rivals. Yeah, that report was wild. Allegedly paying some AI folks, maybe from the deep mind side, to basically sit on the sidelines for up to a year using non-competes and financial incentives. Just to stop them going to meta or open AI or wherever. Apparently. If true, it's just an extreme example of the talent wars in AI right now. Shows how incredibly valuable these skilled people are that companies would pay them not to contribute elsewhere. That is intense. Yeah.

OK, before we wrap up, I just want to mention something that ties into this need for skills and understanding in AI. You know, Etienne Newman, our producer, he actually created an AI powered app called Jamgat Tech. It's designed specifically to help people master and get certified in over 50 different areas. Cloud, finance, cybersecurity, health care, business.

All these fields being transformed by AI. If you're looking to boost your own skills or get ahead with certifications, it's definitely worth checking out. We'll put the links in the show notes, of course. So, wow, looking back at everything we just covered,

It's quite a snapshot, isn't it? It really is. From the absolute hardware foundations with NVIDIA and Google's chips pushing the performance limits to new ways for AI models like Gemini to understand more and for different AI agents to actually collaborate using things like A2A. Right. And then seeing AI pop up in our homes with Bally and wrestling with these huge legal and ethical questions around Musk, deep fakes, even AI in court.

Yeah, it touches everything. Each piece from chip exports to home robots to legal debates. It just shows how deeply transformative this tech is becoming. So thinking about all that, especially the advances in agent collaboration like A2A and the obvious demand for AI skills shown by Anthropix pricing or Google maybe hoarding talent.

What do you think is the biggest bottleneck going forward? Woo, that's a good question. For AI innovation and adoption in the next few years? Is it going to be the hardware? Can we build chips fast enough? Is it the models themselves needing breakthroughs? Or is it these ethical hurdles, the regulations? Or something else entirely?

It's tough to say, isn't it? It could be any one of those or maybe a combination. Hardware supply chains are fragile. Model training is still incredibly expensive. And the ethical and societal alignment piece is massively complex. It's definitely something to keep thinking about. Absolutely. A great thought to leave our listeners with.

Just one final reminder about that JamGet tech app from Etienne Newman. If you want to get hands-on and master those in-demand certifications in cloud, cybersecurity, finance, and more, using AI to help you learn, check out the links in the show notes. Thanks for joining us for this deep dive into the latest AI developments. We'll be back soon to unravel the next wave.