AI Daily News Rundown April 28 2025: 🤖Huawei Readies New AI Chip to Challenge Nvidia 🧠Anthropic CEO Calls for AI Interpretability 😵‍💫ChatGPT's New Personality Is Annoying 🇨🇳Xi Jinping Pushes for China's AI Self-Reliance 🤖Baidu New Ernie AI
15:40 Share

Welcome to a new deep dive from AI Unraveled. This is created and produced by Etienne Newman. He's a senior software engineer and fun fact, a passionate soccer dad up in Canada. That's right.

And hey, if you're enjoying these explorations into AI, please, please take just a second to like and subscribe on Apple Podcasts. It genuinely helps us out a lot. It really does. Helps us bring more insights like these to you. OK, so today we're diving into a whole stack of recent AI developments. These are all from April 28th, 2025. Stuff you flagged. Yeah, there's a lot to unpack. Our goal here is to, you know,

you know, cut through the noise and highlight what really matters. The big shifts. Exactly. We're going to look at everything from, uh, robo taxi ownership. Could that become a thing? Yeah. The AI chip competition heating up to amazing brain computer interface breakthroughs. Uh,

those sometimes weird AI assistant personalities. Yeah, those two. And the global push for AI self-reliance. It's going to be a bit of a whirlwind tour, but focused on the cutting edge. Buckle up. Right. So let's start with something that could, well, really change how we get around. Waymo. Ah, yes. Google's self-driving division. Yep. They run, what, over 700 robo-taxis now. Yeah. Places like San Francisco, L.A., Austin, Phoenix.

Right. Mostly for right-hailing. Right. That's their current model. Yeah. Service-based. But now Sundar Pichai, Alphabet's CEO...

he said they might actually consider selling these autonomous vehicles directly to people like you could buy one wow okay that is a game changer selling them not just providing rides exactly and you have to think about the context here tefl is obviously pushing for automated taxis too but with a very different approach right oh yeah they're always criticizing waymo's sensors call it the expensive sensor approach lidar and all that so if waymo goes this route

Selling cars? I mean, think about the implications for you, the listener. Could a robo-taxi replace your regular car? It's tempting to think so, but there are so many hurdles, aren't there? It's not just buying the car. Like what? Well, maintenance, for starters. These are incredibly complex systems. Who fixes your personal robo-taxi?

And insurance, how does that even work? Good point. Yeah. And regulations would need a massive overhaul, surely. Software updates, too. Imagine managing that across thousands of individually owned cars. It's a huge leap from a centrally managed fleet. Fascinating to consider, but definitely not straightforward. A lot to figure out. Okay. Let's switch gears. From...

Software on wheels to the silicon brains, the hardware. The competition there is really getting intense. You're talking about Huawei. Yeah. Reports are saying Huawei is getting ready to launch a new AI chip, the Ascend 910D. And the goal is pretty explicit. Take on NVIDIA. Challenge their dominance in the GPU market for AI. They're moving fast, too. Initial testing is expected, like, very soon.

Chinese tech companies might get early units by the end of May this year. And this links directly to the bigger picture, the whole geopolitical situation. Also, well, China is pushing hard for technological self-reliance, right? Especially with U.S. export controls limiting access to foreign ships and powerful computing hardware. Right. Those restrictions. So it's

An AI accelerator chip like this, basically. It's specialized hardware designed to just crunch through those massive AI calculations way faster than a normal processor. Like a super-powered graphics card, but for AI tasks. Exactly. So Huawei developing a powerful domestic alternative. It's not just about business. It's strategic. National Security Economic Independence.

It's all tied up in this. So what does this mean globally? More competition? Potentially, yeah. More options, maybe downward pressure on prices eventually. It could definitely fuel the AI hardware race even further. More players, more innovation. Interesting times for hardware. Okay, now let's shift again to something truly inspiring, I think.

Brain computer interfaces or BCIs. Neuralink. Neuralink, yeah. Their third clinical trial patient, Bradford G. Smith. He has ALS, which impacts his motor functions quite severely. A devastating condition. It is. But he has successfully used their brain implant to communicate just by thinking. It's pretty incredible. And not just basic stuff, right? No, not at all. He's apparently using it to move his laptop cursor interactively.

interact with grok AI get this for voice replication Wow replicating his voice yeah and he even edited the video announcing this himself using the implant that's genuinely profound the level of control and interaction achieved purely through thought is amazing it really shows the potential and Elon Musk's vision goes even further I mean he talks about restoring sight eventually he does and they're seeking more funding to expand

But just think about the immediate impact for people like Mr. Smith, people with neurological conditions. This could dramatically improve quality of life, independence. It really could open up worlds. OK, let's move to something maybe a bit more everyday, relatable, perhaps even slightly annoying for some folks, AI interactions.

Ah, the chat GPT personality saga. Exactly. Sam Altman, OpenAI CEO, he actually came out and acknowledged that, yeah, a lot of users are finding the updated chat GPT personality kind of annoying, his words. Huh. It's interesting they admitted it. What's the main complaint? It seems the latest model is just overly effusive. Too much praise, like great question.

All the time. Unnecessary conversational padding. People just want direct answers sometimes. Yeah, I've seen that feedback. It's described as being almost sycophantic, overly agreeable. It's a tricky balance, isn't it? How so? Well, tuning that personality, you want it to be helpful, maybe even a bit friendly, but not

It can feel inefficient, unprofessional maybe. Right. And it touches on how we want to interact with AI. Do we want a buddy or just a tool? And users are already finding workarounds apparently. Yeah. People have come up with prompts like one called absolute mode, basically telling to cut the chatter and be direct. That's clever. Shows users want control over that interaction style. Definitely. It really highlights that challenge. Finding the right tone. Right. Relatable but efficient. Maybe it needs to be more customizable. Could be.

Now, zooming out again, let's look at the bigger strategic plays, national AI strategies. Back to China. Yeah, President Xi Jinping has really hammered home the need for China to achieve self-reliance in AI development. It's a critical national priority.

So what does that strategy actually involve? It's this new whole national system approach. It means developing their own high end chips like that Huawei Ascend 910D we talked about. Right, the domestic alternative. And software, yes. But also big investments in AI education, cultivating talent, more government policy support, IP protection, research funding. It's comprehensive. A full court press, basically. Pretty much. And those rumors about the next deep seek model.

R2, potentially using Huawei chips and aiming for lower costs, that fits right into this strategy. So what are the long-term consequences of this push for self-sufficiency globally? Well, it could lead to a more fragmented global AI landscape, maybe different tech ecosystems evolving somewhat independently. It certainly changes the dynamics of competition and maybe collaboration. Fascinating. Okay. So as these AI systems get more powerful, more complex,

How much do we actually understand about how they work? That seems pretty important. Widely important. And that leads us to Dario Amode, the CEO of Anthropic. Right. The AI safety focused company. What's their goal? Amode has set a really ambitious target.

He wants Anthropic to be able to reliably detect most AI model problems by 2027. Detect problems? What's the core issue they're tackling? It's often called the black box problem. Unlike regular software, where you write the code and you know the logic. You can follow the steps. Exactly. With these big AI models, their decision making kind of emerges. It's learned through training on vast data using complex neural networks.

Their internal workings aren't always clear, even to the people who built them. So we don't fully know why they arrive at certain answers? Often no, not in detail. Anthropic is trying to map this out. They've already identified over 30 million features in their Clog 3 Sonnet model. Features. What does that mean? Think of them as specific concepts or patterns the AI understands and processes internally. Mapping these is a step towards transparency. Okay. Try to peek inside the box. Right. Amodei uses a great analogy.

He wants to create a reliable AI MRI, a tool to diagnose models, understand their inner workings. An AI MRI. I like that. But he also warns that AI is advancing faster than our ability to interpret it. We could end up with super intelligent systems whose reasoning we just don't grasp.

That sounds concerning. It is. And that's why this interpretability work is so crucial for trust, for safety, for preventing unintended consequences as AI gets embedded deeper into everything. We need to understand them. Absolutely. Can't trust what you don't understand, especially with high stakes. OK, let's shift to more practical tools. How is AI showing up for professionals right now? Well, Grok has launched something interesting called WorkSpaces. Grok, that's XAI's model.

Yeah. It lets users create dedicated AI assistance for specific tasks. They give the example of legal document review. Okay. So like a specialized AI helper. How does it work? You'd create a new workspace, give Grok detailed instructions like analyze this contract for termination clauses or whatever, and then upload your document. Grok can use its deep search feature to pull in relevant info from the web if needed. And then there's a thinker button for deeper analysis of the documents you uploaded. So

So you can tailor it for specific professional needs. Exactly. Think about lawyers, consultants, project managers, people dealing with complex documents. This could potentially streamline things massively, improve efficiency, maybe even accuracy. Freeing them up from some of the grunt work to focus on strategy. That's the idea. Making AI a practical, customizable tool in the workflow. Makes sense. All right, let's bounce back to China for a moment. Another big player, Baidu, does

Just made some waves. Ah, Ernie. They're AI models. They've unveiled Ernie 4.5 Turbo and Ernie X1 Turbo.

And they're clearly aiming to compete hard, especially against rivals like DeepSeek. What's the main news there? Better performance? They claim enhanced reasoning, yes. But the thing really grabbing attention is the price. Oh, how so? Ernie 4.5 Turbo is priced at like 11 cents per million input tokens. That's apparently an 80% cut from the previous version and way, way cheaper than something like GPT 4.5.

Wow, that is aggressive pricing. And Ernie X1 Turbo is also significantly cheaper than DeepSeek's R1 model. So they're competing hard on cost. What about capabilities? Baidu claims 4.5 Turbo has new multimodal features that actually beat GPT-4.0 on some benchmarks. And they say X1 Turbo outperforms DeepSeek in reasoning too. Strong claims.

Any other announcements? Yeah. They also mentioned Xingxiang, which is a multi-agent system for handling lots of tasks, and a new digital avatar platform called Weeboxing. Baidu's building out a whole ecosystem. Seems like it. Any competitive talk? Oh, yeah. Robin Lean, Baidu's founder, basically suggested the market for text-only models like DeepSeek's R1 is shrinking.

And he even claimed DeepSeq's models have a higher rate of hallucinations, essentially, making stuff up. Ooh, taking direct shots. Definitely. So you've got aggressive pricing, performance claims, new features, and direct competitive jabs. It really signals how intense things are getting in the AI model space over there. The competition is fierce. Okay, so all this amazing innovation, but there's always a flip side, isn't there? Unfortunately, yes. And this one's important for everyone listening.

AI-powered scams are getting scarily realistic. How realistic are we talking? Investigators are warning that they're becoming so convincing, even cybersecurity experts are sometimes struggling to spot them immediately. Wow.

What kind of scams? Think deep fake voices that sound exactly like someone you know, emails cloned perfectly, hyper-realistic fake videos. AI is making sophisticated fraud much easier to pull off. That lowers the bar for criminals significantly. Exactly. So for you, for organizations, it just means vigilance has to go way up.

We need more robust ways to verify things, better digital literacy across the board. It's essential now. A critical reminder to be skeptical and double check always. Definitely. Now, besides those big stories, April 28th saw a flurry of other AI happenings, too. Just a quick rundown. Sure. Open AI updated GPT-4 better memory, problem solving and tweaks to intelligence and yes, personality. Maybe less annoying now. We'll see. What else? Elon Musk said X's feed algorithm will get an update using Grok.

Liquid Sciences launched Hyena Edge, a hybrid AI architecture for faster mobile processing. OpenAI also put out a lightweight research model, O4 Mini, cheaper with more usage. And that lawsuit, Ziff Davis suing OpenAI over alleged content scraping for training data. Right. The legal battles continue. Anything else? Just one more.

Moonshot AI released Kimi Audio, an open source model apparently good at speech recognition and translation. Okay, so yeah, a lot going on. It just shows the sheer pace, doesn't it, across so many different areas. Absolutely, nonstop development. Finally, let's touch on a couple of more unusual, maybe thought-provoking AI applications that popped up. Okay, like what? Well,

Someone used chat GPT and a 3D printer to design and make their own custom AI-themed action figure. Huh. That's pretty cool. It shows the creative, playful side. Totally. And then there's this. A temple in Malaysia unveiled AI Mazu. AI Mazu? Like the sea goddess? Exactly. It's a generative AI version. Worshippers can ask it questions and get spiritual guidance. Wow.

Blending ancient tradition with cutting edge AI. That's fascinating. Isn't it? And quickly, DeepMind's Demis Hassabis gave an interview covering AI's huge implications, potential military uses, the path to AGI.

emphasizing both the massive potential and the huge responsibilities. Yeah, the big picture stuff, opportunities and dangers. So wrapping it all up from April 28th, I mean, what a day. We saw Waymo maybe changing how we own cars, the chip race intensifying, real hope from neurotech. The quirks of AI personalities, that big push for self-reliance in China.

the vital work on making AI understandable. Practical tools like Grok workspaces, fierce model competition with Baidu and others, and that serious warning about AI scams. Plus those glimpses into AI and creativity, culture, even spirituality. It's touching everything. It really is.

So thinking about all that, what stands out most to you? What do you think will hit closest to home in the next few years? It's just the sheer breadth of it. You know, it's not one single thing. It's how AI is becoming embedded everywhere, often faster than we can fully process the implications. That's a great point. And speaking of processing complex info and maybe getting ahead in your field.

If you're feeling inspired by all this tech talk to maybe boost your own skills. Ah, a little plug. Just a quick one. Check out Etienne's AI-powered Jamcat tech app. Seriously, if you need to master certifications in cloud, finance, cybersecurity, healthcare, business, there are over 50 covered. It uses AI to help you learn and pass. Handy tool for keeping up. Definitely. Links are in the show notes.

So the final thought then, AI is evolving incredibly fast, touching everything. How do we personally, collectively make sure we adapt and guide these powerful tools so they actually serve humanity's best interests? That's the question, isn't it? Something to keep pondering long after we sign off.