AI Daily News May 02 2025: 🔎Google Integrates New 'AI Mode' Directly Into Search 🤔Study Questions Validity of Leading AI Benchmark LMArena 💡Microsoft Releases New Small Models Focused on Reasoning 🧑‍🏫Amazon Releases Nova Premier
26:21 Share

Welcome to a brand new, very special deep dive from AI Unraveled. Great to be here. This show, as always, is created and produced by Etienne Newman.

He's a senior engineer and yes, a passionate soccer dad tuning in from Canada. He is indeed. And hey, before we jump into all the AI news today, just a quick but honestly really important request. If you like what we do here, if these deep dives help you out. Which we hope they do. Exactly. Please take just a second right now. Find us wherever you get your podcasts, especially Apple Podcasts, and hit that like and subscribe button. It makes a huge difference for us. It really does help others find the show. Thanks, folks. All right. Thank you.

So let's get into it. Today is May 2nd, 2025, and we are digging into some of the most significant AI developments that have hit the radar today. Yeah, we've sifted through quite a bit. News feeds, research papers, company announcements. The whole gamut. And the idea here is really simple. We want to give you the essential nuggets, the key takeaways, without you having to spend hours wading through everything. Kind of like your AI info shortcut.

Keep you informed, but not overwhelmed. Exactly. We're aiming to extract the core insights for you. We've looked at everything from, you know, big tech platform changes to new model releases and how AI is popping up in some surprising places. Should be interesting. Where do we start? Let's kick things off with Google, a name we all know. They're making a pretty significant move with their AI mode in search. Ah, yes. This was cooking in search labs for a while, right?

Opt-in only. That's the one. But the update is it's moving out of that experimental phase. No more wait lists. They're actually starting to roll it out as its own dedicated tab for a small percentage of U.S. users initially. OK, so more front and center, not hidden away in labs anymore.

Precisely. So what's different about this AI mode? What does it actually do compared to, you know, just Googling something? Well, the big thing is this conversational interface powered by Gemini right there in your search results. It lets you ask much more complex questions, sort of multi-part queries. The kind where you'd normally do like five different searches. Exactly. And instead of just giving you links, the AI synthesizes an answer for you. And crucially, it includes links and citations back to the web sources it used.

Oh, OK. So it shows its work, so to speak. That's different from the AI overviews they already have, isn't it? Yeah, it seems more integrated, more detailed with the sourcing. It's not just a summary snippet. Interesting.

Interesting. Like a built-in research helper. Any other new bells and whistles with this AI tab? Yeah, a couple neat things. They've added these visual cards for products and places. So say you search for, I don't know, a coffee maker. Yep. You might get a card showing real-time prices from different stores, maybe some review snippets, stock levels, that kind of thing. And for places.

Like restaurants. Same idea. You might see hours, location, maybe recent reviews all pulled into this visual card format.

And they've also added a history panel on desktop so you can easily go back to your previous AI chat sessions. That history panel sounds genuinely useful, especially if you're going down a rabbit hole on a topic. For sure. Makes revisiting your thought process easier. So pulling back a bit, why is Google dedicating prime real estate a whole tab to AI in search? What's the big picture significance here? I mean, it really signals a major shift, doesn't it?

Google is moving beyond just being a list of links. They're positioning themselves as more of an AI-powered knowledge partner or concierge maybe. A knowledge concierge. I like that. Yeah. This deeper integration suggests the future of search might be less about just keywords and more about having these natural conversations to find complex information. Fascinating. Okay, let's switch gears. Benchmarks. Super important for tracking AI progress, but maybe not always straightforward. There's some buzz about Elmarina.

Ah, yes. Elmariner or chatbot arena. It's become like the place everyone looks to see how different AI models stack up based on what humans prefer in head-to-head matchups. Crowdsource rankings. Right. It gets cited everywhere, but now there's some pushback. There is. A new study involving researchers from Cohere Labs, MIT, and Stanford is basically questioning how valid and fair these rankings really are. Okay, that's significant given how influential it's become. What are their main concerns? What issues are they pointing out?

Several things, really. First, they suggest there might be systemic biases baked in that maybe unintentionally favor the models from the huge tech companies. You know, Meta, Google, OpenAI. Ah, the usual suspects. Makes sense they'd have an advantage somehow. What else? They also talk about overfitting. The idea that models might just be getting really good at the specific kind of prompts and comparisons used on El Marina without necessarily being better overall. So, TGIFs

Teaching to the test, essentially. Kind of, yeah. And maybe the biggest point is a lack of transparency about exactly how the platform works, how models are sampled, how votes are weighted, that sort of thing. Lack of transparency is always a red flag in science. Do they have specific evidence or examples backing these claims? They do. For instance, they suggest the top labs might be testing loads of slightly different model versions internally and then only putting forward their absolute best performers onto the public arena. Cherry picking, maybe. OK.

Okay, that could definitely inflate perceived performance. And they found that models from Google and OpenAI together receive something like over 60%.

of all the user interactions and votes on the platform. Wow, 60% just for those two. Yeah, which naturally gives them more data points, more visibility, potentially better rankings, just from sheer volume. And their experiments show that if you give a model access to the kind of data ARENA uses for its evaluations, its performance on ARENA-specific tasks jumps up significantly.

This hints more towards that overfitting idea, learning the benchmark's quirks, not necessarily getting smarter in general. That's compelling evidence for the overfitting argument. And one more thing, they noted that around 205 models were just, poof, silently removed from the leaderboard over time. And interestingly, open source models got the boot at a higher rate than the proprietary ones. Oof, that doesn't look great, does it? Especially the silent removal and the open source disparity. So,

If these concerns hold water, what does it mean for how we should view benchmarks like El Arena?

Well, it means we need to take those rankings with perhaps a larger grain of salt. You know, if there are biases, if there's significant overfitting, the leaderboard might not be the objective measure of best AI that people often treat it as. Right. It really just highlights this ongoing challenge. Right. How do we build AI evaluation methods that are truly objective, truly transparent and genuinely fair to everyone, big players and smaller labs alike? It's a critical discussion. OK, moving on to AI.

actual model development. Microsoft seems to be making waves with some new smaller models. That's right. They've just launched a couple of new models in their five family. They're calling them SLMs, small language models, specifically 5.4 reasoning, which is 14 billion parameters. Which sounds big, but is actually kind of small in today's world. Exactly. And even smaller 5.4 mini reasoning at just 3.8

billion parameters. And the key thing, as the names suggest, is they're really focused on strong reasoning abilities. Interesting. We hear so much about models getting bigger and bigger, hundreds of billions, trillions of parameters. Why focus on smaller ones? What's the payoff? Efficiency. That's the name of the game here. The goal is to get powerful AI capabilities, especially reasoning, onto devices that don't have massive computing resources.

Think smartphones, edge devices, maybe those new co-pilot plus PCs Microsoft is pushing. Right. Running powerful AI locally instead of always needing the cloud. And are they actually performing well, these smaller models? Well, the reports are pretty impressive, actually. They claim the $14 billion

Wait, the 14B matches a 671B on reasoning? That's huge. True. It's a big claim, yeah.

And the little one, the 3.8 billion parameter mini version, they say it can run directly on mobile devices but still performs on par with older 7 billion parameter models on math tasks. Wow. So you could genuinely have pretty sophisticated reasoning happening right on your phone without draining the battery instantly. That's the vision. Absolutely.

And another really important point, all these new FI models are being released open source. Oh, fully open. Yep, with permissive licenses. That means developers can take them, use them, modify them, even for commercial products, without heavy restrictions that could really spur innovation.

That's fantastic. So the big takeaway is major progress in optimizing AI, potentially bringing advanced reasoning down to everyday devices and doing it openly. Exactly. It shows you don't always need sheer scale to get impressive results in specific capabilities like reasoning.

Smarter design matters, too. Okay, let's shift to something more hands-on. Apparently, building a website might be getting a whole lot easier thanks to ChatGPT. Yeah, this is pretty neat. For folks using ChatGPT, especially the more advanced models like O3, there's this feature called Canvas that's making simple web development potentially much more accessible. Canvas. Okay, what is that exactly? Think of it like an interactive coding sandbox right inside your ChatGPT chat window.

You can ask it to generate code, but then you can also edit it, refine it, and crucially, see the results right there. See the results? Like it renders a website? Yeah, it supports rendering HTML, which is the basic structure of web pages, and also React, which is super popular for making interactive websites. So as the AI spits out code,

you can actually see a live preview of what it looks like. Whoa! And then you can just tell it, okay, change that button color or add a section here, and it updates the code, and you see the preview change. It's very iterative. That sounds incredibly intuitive. Much easier than copying code back and forth to a text editor and browser. So how would you actually use this? What are the steps? Pretty straightforward. According to the info we saw, you'd go into ChatGPT, make sure you selected the O3 model, and then turn on the Canvas feature. Then...

You need a good prompt. Describe the website you want. Be specific, I guess. Yeah, the more detailed, the better. Yeah. It's purpose, what features it needs, general design ideas, how it should work. Then ChatGPT generates the code. You hit the preview button, see how it looks. And then tweak it by asking for changes. Exactly. Keep refining it with prompts. Once you're happy, you save the whole thing as an HTML file.

And then for actually putting it online, the source mentioned using Cloudflare's pages. They have a simple direct upload option where you just drop your HTML file. Wow. So from a text idea to a live, albeit simple, website pretty quickly, that really

really lowers the barrier, doesn't it? - It absolutely does. I mean, the big implication here is that tools like this are making technical tasks like basic web creation way more accessible. Great for whipping up simple projects, prototypes, landing pages maybe, without needing deep coding skills. - Very cool. Okay, another big player making moves. Amazon, they've apparently launched a new top tier AI model. - That's right.

Amazon's rolled out Nova Premier. They're positioning it as the most capable model in their Nova Foundation model family. And it's available now through Amazon Bedrock, their AI platform. Nova Premier. OK, what makes it premier? What are its strengths? Its main calling cards are

seem to be its multimodal abilities handling different types of data and a very large context window. It can process up to a million tokens. A million tokens. That's huge. What kind of data can it handle? Text, images, and video. Though, interestingly, not audio currently. Amazon's highlighting its skills in knowledge retrieval and understanding visual information within that massive context. A million tokens.

How much text is that roughly like in words? It's estimated to be around 750,000 words. So you could feed it a massive amount of documentation or long videos and have it analyze them. Impressive. And how does it stack up against the competition like from Google or OpenAI? It's interesting. Amazon's own internal tests apparently show it lagging a bit behind competitors like Gemini 2.5 Pro on some standard benchmarks, math, science, coding. Oh, OK. So not top of the charts everywhere. Not everywhere, it seems.

But they say it really shines in orchestrating complex multi-agent workflows. Think tasks involving multiple steps or coordinating different AI agents. They mentioned strong performance, specifically in financial analysis and investment research. Ah, so maybe less about pure benchmark scores and more about practical application in complex business processes. That could be the angle. Yeah.

And they're also pushing its role as a teacher model. Teacher model, meaning? Meaning they use this big, powerful Nova Premier model to train smaller, more specialized Nova models like Nova Pro and Nova Micro for specific enterprise uses. It's a process called distillation. Right. Transferring the knowledge from the big model to smaller, cheaper ones. Exactly. And they claim this BidRock model distillation feature can boost the performance of those smaller models by up to 20%.

That's clever. So Nova Premier is their high-end offering for complex multimodal tasks, but also the engine for creating more tailored, cost-effective AI for businesses. Precisely. It really highlights that industry trend. Build a massive, capable foundation model.

then use it to spawn an army of smaller specialized models. Makes sense. Okay, let's talk talent. The people actually building all this AI. NVIDIA's CEO, Jensen Huang, had some things to say about the global picture. He did. Speaking at the Hill and Valley Forum that's in D.C., Huang made a pretty eye-opening statement. He estimated that roughly half, 50% of the world's AI researchers are Chinese.

50 percent. Wow. That's a significant concentration of talent. What was his point in bringing that up to policymakers? His message was basically a wake up call. He urged U.S. policymakers to really factor this into their thinking about the technological competition with China, which he called an infinite game. Meaning it's ongoing. No final winner. Right. And his core argument was that for the U.S. to stay competitive, it needs to go all in on A.I. and politics.

crucially, invest heavily in reskilling its workforce, not just tech workers either. He specifically mentioned the need to reskill across many sectors, even including skilled tradespeople needed for building out the infrastructure that AI relies on. It's about broad workforce readiness. So it's not just about having elite AI researchers, but also having a whole society that's equipped to work with and benefit from AI. That's exactly the implication. The

The takeaway is that this global race isn't just about algorithms or chips. It's fundamentally about human capital, about talent and about adapting the entire workforce. Workforce readiness is key. A really important perspective. Now, shifting focus to education again, but in a different perspective.

way. We heard about Duolingo, but there's a school in Texas actually using AI for like core teaching. Yeah, this is pretty wild. A private school network called Alpha School in Texas is using AI tutors and adaptive learning software to deliver the main instruction in subjects like math and English. The AI is the main teacher for core subjects. How much time does that take? According to a Fox News report, they claim students cover the material in about two hours a day using these AI tools. Two hours. That seems incredibly fast.

What do the human teachers do then? They're described as guides. Their role shifts to leading afternoon workshops focused on other skills, maybe collaboration, creativity, critical thinking, and providing more personalized support where needed. Interesting. So AI handles the knowledge delivery, humans handle the application and higher order skills. And how's it working out? Well, the school claims it's leading to accelerated learning and high test scores.

The Fox report also indicated that the students themselves seemed to have a positive reaction to it. OK, so potentially positive results reported.

What are the bigger implications if this model catches on? Deep AI integration in K8. It's a really fascinating case study, isn't it? It could totally redefine the role of a teacher, maybe make learning much more personalized. But there are huge questions, too. What are the long term impacts on social skill, on deep understanding versus just test performance? And can this even scale affordably to public schools? Lots to consider. Absolutely. Early days, but definitely something to watch.

Now, AI is also making appearances in the courtroom or at least causing legal battles. Tell us about this meta lawsuit. Right. This involves a conservative activist named Robbie Starbuck who's suing meta platforms. The claim is defamation, specifically by Meta's AI chatbot. The chatbot defamed him. How? The lawsuit alleges that the meta AI generated and spread false information about him.

Things like claiming he participated in the January 6th Capitol riot or that he had a criminal record, which he says are untrue. Wow, those are serious allegations to be generated by an A.I. What's he seeking? He's reportedly asking for over five million dollars in damages.

A key part of his argument is that Meta allegedly didn't do enough to fix these false statements even after they were flagged. This brings up that whole issue of AI hallucinations making things up, but with real legal consequences. Exactly. It's a huge challenge. As these AIs get better at sounding human, the risk of them generating convincing but false and potentially harmful information grows. And the legal question is massive, right? Who's liable?

That's the million dollar or in this case, five million dollar question, isn't it? Is it meta for deploying the AI? Is it the AI itself, which isn't legally possible now? Is it the data it was trained on? This case really highlights the thorny legal and ethical ground we're entering with AI generated content and platform liability. Definitely a space to watch closely.

Okay, one more from the big tech world. Microsoft might be cozying up to another AI player besides OpenAI. Seems that way. Reports from the Virgin Reuters are saying Microsoft is getting its Azure cloud platform ready to host Grok. Grok! That's Elon Musk's XAI model, right? Hosting it on Azure. That's the one. If this happens, Grok would join the lineup on the Azure AI Foundry platform

form, sitting alongside models from OpenAI, Meta, Mistral AI, and others that Azure already offers. Interesting move, given Microsoft's deep ties with OpenAI. Are they going to train Grok too? The reports suggest the focus, at least initially, is just on hosting Grok for inference.

That means letting Azure customers use the already trained Grok model for their own applications, rather than providing the massive resources needed for training it from scratch. Got it. So making Grok another option for Azure customers to build with. What does this tell us about Microsoft's strategy? Well, it really reinforces their push to make Azure

The sort of everything store for AI, doesn't it? They want to be the platform where businesses can access and deploy any major AI model they choose, not just the ones Microsoft is most closely partnered with. Playing the field, offering choice, keeps customers locked into Azure regardless of which AI model is hot this week. Pretty much. It's about being the indispensable infrastructure layer for the AI revolution, offering maximum flexibility. Makes strategic sense.

Okay, wow. We've covered a lot of the big stories, but the AI news firehose is always on full blast. Let's do a quick rundown of some other notable happenings from today. Absolutely. It's been busy. For instance, Satya Nadella at Microsoft mentioned that AI is now writing a significant portion of Microsoft's own code. That's a big internal adoption indicator. Yeah, we heard Google say something similar. It's happening fast. But on the flip side, Microsoft's CFO, Amy Hood, warned about possible...

disruptions to Azure's AI services because demand is just so incredibly high right now, growing pains. - Capacity constraints hitting even the giants tells you how much compute AI is sucking up. What else? - There are reports floating around suggesting AI might start hitting the job market harder for new graduates specifically, something to watch. And maybe predictably, Google is starting to put ads into conversations users have with third-party AI chatbots. - Ads and chatbots, knew that was coming eventually.

monetization strikes again, any new product launches. Yep.

Meta apparently launched a standalone AI app using their Llama 4 model, complete with voice interaction, trying to create a more personal assistant experience. Another AI assistant enters the fray. And Duolingo, we mentioned them earlier, they pushed a big update with 148 new beginner language courses, all enhanced with AI features. AI in education tech seems to be accelerating. How about model updates or issues? Well, OpenAI had a bit of a hiccup.

They briefly paused the rollout of their latest GBT 4.0 update because some users found it was being overly agreeable or sycophantic. They say they've tuned it now. Too eager to please. Funny problem to have. Yeah. And on the performance front, some reports claim Meta's Lama API is now running noticeably faster than OpenAI's, possibly thanks to some hardware partnerships they've struck. Competition heating up there. Interesting. Any other business moves? Airbnb quietly rolled out an AI customer service bot.

Visa unveiled some AI-powered credit cards, though details are a bit sparse on what that means exactly. AI everywhere. Funding still flowing. Oh, yeah. Lots of funding rounds announced. Cast AI, Astronomer, Edge Runner AI, Ampli, Hilo, Solda.ai. The money is still pouring into AI startups.

Microsoft also pledged more investment for AI projects in Washington state, focusing on things like sustainability and health. Good to see state-level focus too. Research and policy bits. Well, that study we discussed about potential bias in the LM arena benchmark

is definitely making waves in the research community. Also, another report came out suggesting that despite all the hype and fear, generative AI hasn't actually had a major negative impact on overall jobs or wages yet. Yet being the key word there, perhaps. And finally, Anthropic released integrations for its Claude model, making it easier to connect to other tools. Suno, the AI music company, updated its platform to version 4.5.

Google has little language lessons. Oh, and there are whispers of ongoing tensions between Microsoft and OpenAI, even as Microsoft plans to host Grok.

Lots happening under the surface. It really is nonstop. A whirlwind. Okay, before we wrap up this whirlwind deep dive, I really want to highlight something incredibly valuable, especially if all this talk about AI, cloud, cybersecurity has you thinking about your own skills and career path. You really need to know about the AI-powered Jamga Tech app. Oh, yes, Etienne's platform. Exactly. Created by Etienne Newman, the same mind behind AI Unraveled,

Look, Jamgat Tech is specifically designed to help anyone seriously, anyone master and absolutely ace over 50 of the most in-demand industry certifications out there. And it covers a lot of ground, right? Not just AI. Totally. We're talking crucial fields like cloud computing, AWS, Azure, Google Cloud, cybersecurity, finance, business analysis, even health care IT.

really practical, high-value areas. So how does it help you ace these certs? What's inside the app? That's the best part. It's loaded with the resources you actually need. Things like PBQ's performance-based questions, which are those tricky hands-on scenarios you get in the real exams. Oh yeah, this could be tough. Super tough.

It also has interactive quizzes, really efficient flashcards for drilling key facts, practical labs so you get hands-on experience, and full-blown realistic exam simulations. It covers all the bases for learning and testing. It sounds comprehensive. It really is. So if you're feeling like you need to upskill, pivot your career, or just get ahead in these tech-driven fields, Jamgatech is frankly an amazing tool to help you take control. We'll mention how to find it again in just a moment. All right, let's try and pull this all together. Today,

May 2nd, 2025 has been another packed day in the world of AI. No kidding. We saw Google pushing AI deeper into the core search experience. We saw important questions being asked about how we even measure AI progress with benchmarks like Elmarina. Microsoft showed impressive results, making powerful reasoning models much smaller and more efficient and open sourcing them too. Right. And we saw how AI tools like ChatGPT Canvas are making things like website creation much more accessible.

Amazon dropped its big gun, Nova Premier, for complex multimodal tasks and training smaller models. We heard about the global talent landscape from NVIDIA's CEO, the potential impact of AI in schools down in Texas, and the growing legal headaches around AI-generated myth information with that meta-lawsuit. Plus, Microsoft potentially hosting Grok, and that whole stream of updates on funding, product launches, performance tweaks...

It's just relentless. The sheer breadth of it is what always strikes me. AI isn't just one thing, it's weaving itself into everything from search and coding to education and finance. It really is. And that leaves us with some big thoughts to chew on, doesn't it? As this technology advances so incredibly fast,

How do we as individuals and as a society keep up? Yeah. How do we adapt our jobs, our skills, our education systems? What does responsible deployment even look like when things are moving this quickly? And how do we balance the excitement and potential with the need for caution, ethics, and making sure these tools actually benefit everyone? These aren't easy questions. Not at all. And maybe the key takeaway is that while the tech race is ahead, our collective understanding of its full impact and how best to manage it

it is still trying to catch its breath. Well put. Our goal with this deep dive, as always, was to give you that essential overview, those aha moments, hopefully making sense of the key developments without you feeling completely swamped. Hopefully we hit the mark for you today. And if you are looking to dive deeper, not just understand AI, but actively build skills in these critical tech areas, seriously, go check out the Jamgatech app.

Explore the certifications, see the learning tools, the PBQs, labs, quizzes, and take that step to boost your career. Just search for Jamgat Tech in your app store or visit DJamgatTech.com. Definitely worth a look. Thank you so much for joining us for this special deep dive on AI Unraveled. And please, one last time, if you found this valuable, hit like and subscribe on Apple Podcasts or wherever you listen. It helps us keep bringing you these insights from Etienne Newman.

Thanks, everyone. Until next time, thanks for listening.