AI Daily News Rundown April 25th 2025: 👨‍💻AI Now Writing Over 30% of Google's Code 🧠Anthropic Launches AI Welfare Research Program 🕵️‍♂️Perplexity's Upcoming Browser 🎵Google DeepMind Expands Music AI Sandbox with New Features
13:54 Share

Welcome to a new deep dive from AI Unraveled. This is created and produced by Etienne Newman, who's a senior software engineer and I believe a passionate soccer dad up in Canada. That's right. And hey, if you're enjoying what we do here, exploring AI, please do take a second to like and subscribe on Apple Podcasts. It really helps us out. It absolutely does. So today we're jumping into some really interesting

significant AI stuff that surfaced around April 25th, 2025. Yeah, it was a busy day, seems like. The scope is just huge, isn't it? We're seeing impacts from basic web browsing all the way to how code gets written. Exactly. We've sifted through a bunch of news and announcements from that day.

Our mission really is to pull out the key developments, the things you actually need to know without, you know, drowning you in jargon. Kind of get you up to speed, maybe spark a few aha moments along the way. Hopefully. Okay, so let's dive right in. First topic, perplexity. They've announced they're building their own browser called Comet.

That's well, that's a big step. It really is. And what's grabbing headlines is Comet's plan to monitor user activity well beyond just the perplexity app itself. Oh, beyond the app? Like what? Well, things like your general browsing habits, what you buy online, even your location data. Wow. OK, so it's not just search queries anymore. They're looking at a much wider picture of your online behavior. Pretty much. And the stated goal, according to the CEO, Eravind Srinivas, is to

to deliver incredibly personalized ads right within the perplexity got form hmm personalized as he thinks users will be okay with this level of tracking he seems to believe so the argument is that the ads especially in things like their discover feed will be so much more relevant that users will accept the trade-off it definitely sounds familiar doesn't it echoes of Google and meta strategies yeah relying heavily on user data for ad revenue it's a classic model yeah but

But it does bring up those big questions about privacy again. How much are you, the listener, comfortable sharing, even if the ads are supposedly better? Right. And with Comet slated for release in May 2025, so pretty soon. Yeah. It's definitely something to think about. Absolutely. A developing story there. Okay. Moving on. Next up, Apple.

There's news about them restructuring their internal robotics team. Ah, yes. This is interesting. They're apparently shifting the unit out from under the main AI division, which John Giandrea heads up. And moving it into the hardware division, right, under John Ternus. Exactly. And, you know, combined with some recent leadership changes around Siri, it hints at a bigger strategic realignment at Apple, maybe trying to get hardware and AI working more closely together on certain projects. And this robotics group.

It's been kept pretty quiet, hasn't it? But the reports mentioned some, well, pretty futuristic concepts they've been working on. Like what? Things like expressive AI lamps and a sort of tabletop home device with a robotic arm and a screen. Wow. Okay. That's not just iterative updates. That's ambitious stuff. Definitely. Yeah. And moving them to hardware kind of signals that maybe these aren't just lab experiments anymore, right? That's right.

That's the suggestion. Yeah. It could mean they're moving from pure research towards, you know, actual product development, getting serious about potentially bringing some kind of consumer robotics to market. Apple entering consumer robotics. That would be massive. It would. And knowing Apple, they'd likely focus on that tight integration, the user experience, making complex robotics feel, well, Apple-like. This move could be about achieving that synergy. Fascinating.

Okay, let's switch gears a bit to something more philosophical. Anthropic is launching what they're calling an AI welfare research program. Yeah, this one's really thinking ahead. Anthropic is basically setting up a program to study the ethical side of advanced AI, specifically looking into whether these systems could potentially develop consciousness or some kind of moral status. AI welfare, like the well-being of the AI itself. That's

quite a concept. It is. They want to develop ways to actually assess if an AI model might be showing signs of, say, distress or even having preferences. How would they even do that? Well, that's the research goal, isn't it? They're looking into criteria for consciousness, studying potential indicators in AI behavior, and even thinking about interventions if needed. It ties into the whole AI safety and ethics discussion. And they hired someone for this specifically, didn't they?

They did. Kyle Fish back in 2024. He's their first welfare researcher. And interestingly, he's put the odds of current models already being conscious at around 15 percent. 15 percent. Wow. That's not zero. And there was a recent report he co-authored suggesting consciousness could be a near term thing. That's right.

Though it's crucial to add, anthropic themselves are very clear. This is highly uncertain territory. There's absolutely no scientific consensus on AI consciousness. It's still very speculative. But what it shows you, the listener, is that leading AI labs are starting to proactively grapple with these really deep issues.

ethical questions that come with building more and more powerful AI. It's good they're thinking about it now, I suppose. Okay, let's talk creative tools. Adobe made some big announcements at Adobe Max London. Yes, especially around their Firefly AI. They launched Firefly Image Model 4 and also an Ultra version. The promises, well, better image generation, more realism, more control for the user, and higher resolution support up to 2K. Better images are always good for creatives. But

But the really big news seemed to be about integrating other AI models into Firefly. That's a major move. Yeah. They're opening up Firefly to include third party models. We're talking OpenAI's GPT ImageGen, Google's Imogen3 and their video model VO2. Even Black Forest Labs Flex 1.1 Pro. Firefly becomes more of a hub.

accessing lots of different AI tools in one place. Exactly. It gives users a much wider palette of AI capabilities without leaving the Adobe ecosystem potentially. And there was more, right? Video features, vector graphics. Uh-huh. Firefly's text-to-video is now officially out of beta. Same for their text-to-vector model. Plus, they launched Firefly boards in data that's for collaborative AI mood boarding, and a new Firefly mobile app is coming too. Lots happening. And Adobe's big selling point is often commercial safety, right?

Absolutely. They stress that their own Firefly models are trained on commercially safe data and designed to be IP friendly. And they've added this new content authenticity feature so you can embed metadata saying AI was used. So for creatives listening, this means more

more powerful, more versatile tools, potentially easier access to different AI engines, and some assurances about using the output commercially. That's the gist of it. Pretty significant updates for that space. Now, speaking of tools, let's shift to developers. There's a new thing called AIDR, an AI coding assistant that runs right in your terminal. Yeah, this is quite handy for coders. It's based on OpenAI's Codex CLI.

And it lets you basically chat with an AI about your code directly from the command line. So you can just type commands in natural language, like explain this code or write a function to do X. Exactly. Explain code, modify it, generate new stuff. It aims to be like having an AI pair programmer right there in your terminal window. Okay. How do you get started with it? Is it complicated?

Seems pretty standard if you're used to Node.js and NPM, you need those installed, then it's just npm install dash e at open a codex. And then set your API key. Export open IP key, your key here.

After that, you can just type codex to start an interactive session or run direct commands like codex refactor this class. Does it just automatically change your code? It has different modes. It can just suggest changes or you can let it auto edit or even go full auto. So you have control. For developers listening, this could be a real boost to productivity, maybe even code quality. Streamlining the coding process.

Makes sense. OK, how about AI and music? Google DeepMind updated their Music AI Sandbox. Right. This is their suite of experimental tools for musicians. It's designed to help with things like generating instrumental riffs, figuring out vocal harmonies, just exploring new musical directions. And what's new there? They've added specific features called Create

Extend and edit. So create lets you generate tracks from text prompts. Extend takes a musical idea you have and continues it. And edit lets you transform audio clips using text descriptions. Powered by a new model, I assume? Yes, the upgraded Lyria 2 model. They say it offers much higher fidelity sort of professional grade audio.

And they also introduced something called Lyria Real Time for, well, real time interactive music generation. Sounds like it could be great for getting past creative blocks or just experimenting. Is it widely available? They're expanding access. More musicians, songwriters, producers in the U.S. are getting invited to try the experimental sandbox now. So wider testing. Interesting stuff for the music creators out there. Now, here's a statistic that really jumped out.

Over 30% of new code at Google is now being generated by AI tools. Yeah, that's a huge number, over 30%. It's not just assisting anymore. AI is writing a significant chunk of their new software. That feels like a major shift in how development is actually done at scale. Absolutely. It points to massively accelerated development cycles, potentially. But, you know, it also raises new questions. Like what? Quality control.

Oversight. Exactly. How do you manage quality assurance when so much code is AI generated? What's the long-term maintenance look like? It's a big change for the industry to digest. Speaking of potential issues, there's this report about science papers. Researchers found hundreds of papers that seem to have used AI-generated text without saying so. Oof, yeah. That's definitely ringing alarm bells in the academic world.

transparency and just the basic integrity of published research are being questioned. So the main point there is? We urgently need clear rules, clear guidelines about how and when AI use needs to be disclosed in scientific publications. It's becoming critical. Makes sense. Okay. On a more positive note, MIT researchers have created something they're calling a periodic table of machine learning. Ah.

I like that idea. It's clever. The goal is to organize all the different machine learning techniques in a structured way, almost like the chemical element. To make it easier for scientists to find the right tool for their job. Precisely. Instead of getting lost in the sea of algorithms, they can use this table to quickly identify promising AI methods for specific scientific problems. It could really help make AI research more intuitive, maybe even speed up discoveries. It could be incredibly useful. Now, this next one caused a bit of a stir.

California's bar examiners admitted using AI to help draft questions for the bar exam. Right. And the controversy comes from the fact that they didn't tell the exam candidates about it beforehand. Yeah. Using AI for a high stakes professional exam without disclosure. That feels problematic. It definitely sparked a debate about transparency and fairness and certification. What does it mean for the validity of the test?

Could there be hidden biases in the AI generated questions? It really shows how AI is creeping into these really critical professional processes. And we haven't quite figured out the ethical guardrails yet. That's a good way to put it. Disclosure, bias, fairness, all big concerns as AI integrates more deeply.

OK, let's look at the hardware side again. Reports from Amazon Web Services and NVIDIA suggest demand for AI data centers is still booming. Yeah. Despite some whispers about maybe an AI investment bubble or slowdown, the big cloud providers and chip makers are saying, nope, the demand for that specialized AI infrastructure is still growing rapidly. Driven by companies actually adopting AI, not just hype.

It seems so. Enterprise adoption, cloud, AI services, they're all fueling this need for more AI-focused data centers. So that part of the tech economy looks

Pretty solid right now. Good to know. The foundations are still strong. Okay, just a few more quick hits from that same day, April 25th. It was packed. Go for it. OpenAI apparently plans to release an open source reasoning model sometime this summer. Okay, that could be significant. Open source reasoning models are always interesting. Tavis launched a new lip sync model, Hummingbird Zero 5.

claiming state-of-the-art results. Lip-sync tech is getting scarily good. There was also an executive order from U.S. President Donald Trump setting up an AI education task force and a presidential AI challenge. Government involvement stepping up. Lovable released version 2.0 of their platform with multiplayer workspaces.

Grammy winner Imogen Heap put out AI style filters on the GenMusic platform. And Hicksfield AI announced a faster, cheaper AI video model called Turbo. Wow. Just a snapshot of one day. And look at the breadth. Reasoning models, lip sync, government initiatives, creative tools, video generation. It really underscores the point, doesn't it? We covered everything from personalized ads and home robots to AI consciousness, creative suites, coding help.

exam questions. The pace is just relentless. And AI is touching almost every sector you can think of. It's becoming pervasive. Absolutely. And look, if this deep dive has got you thinking more about AI and maybe you want to level up your own skills, perhaps get certified in areas like cloud, finance, cybersecurity, healthcare, or business, well, you should definitely check out Etienne Neumann's AI Power Jam Get Tech app.

It's specifically designed to help people master and pass over 50 different professional certifications using AI study tools. Ah, leveraging AI for learning. Makes sense. Yeah, exactly. We'll put the links in the show notes for you to check out the Jenga Tech app. Sounds useful. So wrapping up, considering everything we've just discussed, this incredible pace of change.

It leaves us with a big thought, doesn't it? Which is? For you listening, what parts of your life, your work, your profession, what do you think will be the most fundamentally reshaped by AI in, say, the next few years? That's a heavy question and probably one we'll keep coming back to. No doubt about it. Things are moving fast.