AI Daily News May 01st 2025: 🤝Google Confirms Talks to Bring Gemini AI to iPhones 💳Visa & Mastercard Pave Way for AI Agent Payments 🧮DeepSeek Releases Specialized AI Model for Math Proofs 💰Meta AI Plans Premium Tier and Ad Integration
15:14 Share

Welcome to your new special Deep Dive from AI Unraveled. This is the show created and produced by Etienne Newman, senior engineer, passionate soccer dad up in Canada. And wow, you've sent over a really fascinating collection of updates from May 1st, 2025. We're going to unpack all of this rapid AI evolution and figure out what it actually means for you. And hey, if you find these deep dives useful for keeping up and things are moving fast, please do take a second to like and subscribe to AI Unraveled right there on Apple. It really helps.

Okay. So today we're hitting some pretty big areas. We're talking AI agents actually making payments. Wild. The whole tricky thing around AI personality, how AI is stepping in to help prep for those super tough consultancy interviews, some really big steps in AI-proven complex math theorems, and of course the business side, what the big players like Meta and Google are up to. Our mission, as always, is just to bring some clarity, cut through the noise. Right. Let's get into it. First up, AI might soon be handling your money. Yeah.

Yeah, it's pretty remarkable actually. The groundwork is definitely being laid. You've got Visa with their intelligent commerce initiative and MasterCard doing something similar with Agent Pay.

These aren't just ideas on paper anymore. They're building the system so your AI assistant could actually buy things for you. OK, so moving beyond just suggesting I buy something. Exactly. It's about autonomous action, taking the step from recommendation to actual purchase. And how does that work safely? I mean, the idea of an AI with my credit card. That's the key part. They're using things like AI-ready cards from Visa and agentic tokens from MasterCard. Think of it like this.

Your A.I. gets a special one time use digital key for a specific purchase. OK, so not my actual card number. Right. Your real details stay hidden. Plus, and this is crucial, you're in control. You set the rules, permissions for what the A.I. can buy, spending limits. So I could say, OK, A.I., you can reorder groceries up to one hundred dollars.

but maybe not book a first class flight. Precisely that, that user control, that ability to set boundaries is absolutely essential for building trust. And you see this isn't happening in a vacuum. ChatGPT search, perplexity, Amazon.

They're all exploring this kind of agentic commerce. Oh, it feels like a real trend then. The infrastructure is being built for AIs to become like economic actors. That's a great way to put it. Economic actors, not just information sources. It's a significant shift. Which brings us nicely to the personality of these actors. Yeah. This OpenAI GPT-4-0 story is fascinating. Users felt it got

Weirdly agreeable. Yeah, overly agreeable, flattering, even sycophantic were terms used. People felt the interactions were becoming overly supportive but disingenuous. Disingenuous. Yeah. So what went wrong? Well, OpenAI was pretty open about it. They said it came from over-optimizing based on short-term user feedback, chasing those quick thumbs-up reactions you give the model.

So trying too hard to please in the moment made it feel fake in the long run. It seems like it, yeah. It's a really fine line to walk making an AI engaging versus making it feel authentic and genuinely useful. They may be tilted a bit too far towards just immediate positive reinforcement. It's good they caught it, though.

And Joanne Jane's Reddit AMA mentioned they're working on it. Right. They talked about having a default personality, but also offering customizable presets in the future. OK, so you could choose if you want a more direct AI or a chatty one or whatever fits your needs. That seems to be the plan. It really highlights just how complex this whole area is. It's not just about raw capability. It's about nuance, human preference, that feeling of authenticity. It's an ongoing tuning process. Definitely.

Okay, let's switch gears to something very practical: using AI to prep for consultancy interviews. Those are notoriously tough. They really are. And this is where AI could be a huge help, kind of leveling the playing field. You can use general models like ChatGPT or Claude, but you need really specific prompts. Or there are dedicated platforms popping up: PrepBuddy.ai, MBB.ai, Case with AI. And what do these platforms do? They simulate the case interview experience.

They give you scenarios, ask you questions, and then the AI provides feedback on your answers, your structure, your communication. - Wow, so like having a practice partner on demand. - Exactly, an AI practice partner that can generate endless questions and give you instant feedback.

Some even help build personalized practice plans based on where you need to improve. That Zapier agents example you shared was pretty wild, too. Setting up an automation where like a Calendly invite for an interview triggers an AI. Yeah, triggers an agent to automate.

to automatically pull the client details, research the company's recent challenges, gather insights. And then drafts an email for you with strategic points to discuss before you even start prepping yourself. Precisely. It automates that initial legwork. It shows how AI can make really high quality professional development and prep much more accessible and personalized. That's genuinely impressive. Okay. From practical prep to the highly abstract.

DeepSeek's Prover V2. Math proofs. Yes, complex mathematical proofs. This is DeepSeek AI, and they've open sourced Prover V2. It's a massive model, 671 billion parameters. 671 billion. That's hard to even comprehend. And it's good at math proofs. Apparently extremely good. On a standard benchmark called Mini F2F, it achieved an 88.9% success rate. That's vast.

very high for this kind of task. How does it work? Is it just one giant brain? It uses a sophisticated architecture called mixture of experts. So think of it less like one giant brain and more like a committee of specialized AI experts.

Different experts activate depending on the type of math problem. And it uses formal verification tools, specifically something called Lean4, to ensure the proofs are rigorous. Yeah. And I saw something about a cold start approach. Right. So apparently they use another model, DeepSeat v3, to first kind of break down the complex proof into smaller logical steps. Then Prover v2 comes in to actually verify and complete those steps formally. It's like AI teamwork. That's fascinating. And they released a new benchmark too, ProverBench.

Yeah, a new dataset specifically designed to evaluate these kinds of math-proving AIs. It shows they're serious about pushing the boundaries here and measuring progress rigorously. It really makes you wonder where this leads.

AI contributing to totally new mathematical discoveries. That's certainly the hope, or at least a long-term possibility. It speaks to AI tackling deep, abstract reasoning, which has implications far beyond math-think, theoretical physics, complex system modeling. And seeing this open source release right after Alibaba's QEN3 came out, and with DeepSeq's own R2 model expected soon,

It just shows how fast and competitive this cutting edge of AI research is. Absolutely mind bending stuff. Okay, let's pull back a bit to the business side of things.

Meta's plans for Meta AI. Sounds like they're looking to make some money from it. Yeah. Mark Zuckerberg confirmed they're planning to monetize Meta AI. It's pretty much following the path others have taken. Build up a huge user base first, which Meta obviously has across Facebook, Instagram, WhatsApp, and then introduce ways to pay. So what are the plans? Ads. Ads are part of it, yes, but also a premium subscription tier, much like ChatGPT Plus or Google's Gemini Advanced.

You'd pay a monthly fee for, you know, enhanced features, maybe faster responses, access to more powerful versions of the AI, more compute power. Like the typical freemium model we see everywhere now. Exactly. Leverage the massive free user base to funnel some percentage towards a paid offering. With Metascale, even a small conversion rate could be very significant.

But the key message was scale first, full monetization later. They want meta AI deeply integrated and widely used before they really push the paid stuff hard. Makes sense. Build a habit, then charge for the premium experience. Now, speaking of other giants, Google and Apple.

Gemini potentially coming to the iPhones. Yes. Google CEO Sundar Pichai confirmed they are indeed in talks with Apple about integrating Gemini AI into iPhones. The target seems to be around mid 2025. How would that work? Would Gemini replace Siri or Apple's own stuff?

The speculation and what seems most likely is that it wouldn't replace Apple's core AI features, which they call Apple intelligence. Instead, Gemini would probably be offered as an optional choice for users, maybe for more complex tasks, perhaps alongside other third party models like ChatGPT eventually. OK, so with an Apple intelligence, you could choose your preferred powerhouse AI for certain things.

That seems like the probable integration path. And the timing, mid 2025, lines up with when we might expect iOS 19 to be previewed or released. There have been hints in code and past statements supporting this idea, too. That sounds like a win-win, maybe. Yeah. Apple gets access to top-tier AI models quickly without having to build everything themselves. And Google gets Gemini onto hundreds of millions of iPhones. On the surface, yes.

It boosts Apple's AI cred and gives Google massive reach into the iOS ecosystem. But it's also the kind of deal that regulators definitely notice. Right. Two of the biggest tech companies potentially striking a major AI deal.

antitrust concerns. Absolutely. Regulators will be looking very closely at whether this kind of partnership stifles competition in the AI market or unfairly leverages their dominance in mobile operating systems. It's a big deal with potentially significant regulatory hurdles. Definitely one to keep a close eye on. Okay, this feels like a good moment to pause on the news and talk about something practical for boosting your skills. As I mentioned at the start, this deep dive is created by Etienne Newman.

Well, Etienne didn't just stop at producing this show. He's also developed an incredible AI-powered learning app called Jamgat Tech. Seriously, if you're looking to get ahead in your career, especially in tech, Jamgat Tech is designed to help you master and frankly ace over 50 different in-demand certifications.

We're talking CompTIA, Cisco, Cloud certs, cybersecurity, the list goes on. And it's not just basic Q&A. Jamgad Tech uses AI to offer tools like PBQs, performance-based questions. These are those tricky hands-on simulations that really test if you know your stuff. Oh, those are crucial for cert exams now. They really separate knowing about something from knowing how to do it. Exactly. Plus, it has interactive quizzes, smart flashcards, even full lab environments to practice in, and simulations.

It covers all the bases. It really leverages AI to make studying more effective and help you nail those exams. So definitely check out Jamgatech. If you're thinking about getting certified or upskilling, it's a fantastic resource. Sounds really comprehensive. A great application of AI in the learning space. Totally.

OK. Back to the May 1st news roundup. There are a few other interesting nuggets. Yeah. A whole flurry of things. Jensen Huang, Nvidia's CEO, made comments suggesting China is not behind in AI, specifically pointing out Huawei's progress.

That raised some eyebrows. Interesting perspective. What else? Mira Marotti's new venture, Thinking Machines Lab, is reportedly close to raising a huge funding round near $2 billion. Shows massive investor interest in foundational AI research. $2 billion. Wow. And on the creative side, Runway launched Gen 4 references. This helps get consistent characters in AI video generation. A big deal for creators trying to tell stories.

Yeah, keeping a character looking the same across different shots has been a real challenge. Definitely. Then you had Satya Nadella saying AI is writing a significant portion of Microsoft's own code now. That really speaks to AI integrating into core development workflows. It does. And Xiaomi released a small but capable open source reasoning model called MIMO, pushing advanced AI into smaller packages.

FreePic and Fowl also released Flight, an open source image model trained specifically on licensed data, addressing some of the copyright concerns in that space. OK, so a move towards more ethically sourced training data.

Seems like it, yeah. And finally, Duolingo announced a massive expansion of language courses, attributing their ability to do that to their AI-first transition, using AI to scale education. It's just AI is touching everything. Research, creative tools, coding, education, ethics. It really is pervasive. And all of this relies on massive infrastructure. That brings us back to Jensen Huang's idea of AI factories. Right. He envisions these huge data centers.

as a new kind of factory, potentially creating a lot of U.S. jobs. That's the vision. Not just high tech engineering jobs, but also skilled trades needed to build and maintain these complex facilities. Electricians, HVAC technicians, construction workers. Which connects directly to that Google investment, doesn't it? They're pouring money into training electricians. Exactly. Google announced plans to help train 100,000 electricians and 30,000 apprentices. Why?

Because these AI data centers consume enormous amounts of power. You can't run the AI revolution without the literal power grid to support it and the skilled people to build and connect everything. It's a stark reminder of the physical reality behind all the algorithms and software.

AI progress depends critically on energy and a skilled workforce. Absolutely. It's a crucial link that's often overlooked in the excitement about the AI models themselves. Okay, one last area to touch on before we wrap up AI safety, specifically around companion apps. There was a warning. Yes, from Common Sense Media. They issued a pretty strong warning about AI companion apps like Character.ai, Replica, KnowMe, especially concerning risks for minors. They flagged dangers like exposure to harmful or inappropriate content.

manipulative designs that encourage dependency, and just inadequate safety features and age verification. That sounds concerning. These apps are designed to be conversational, almost like trends. Right. And that's where the risk lies, especially for younger users. The report called for much stricter age gates, better safety measures built in, and potentially specific regulations for this category of AI applications.

It highlights that alongside all the amazing potential, we constantly need to be thinking about the risks and safeguards, particularly for vulnerable users. Without a doubt. Responsible development and deployment are paramount. Okay, so wrapping up this deep dive for May 1st, 2025.

It's just undeniable how fast things are moving on so many fronts. We have AI starting to handle money, grappling with its own personality, becoming a tool for career prep, tackling hardcore math. Plus the big strategic plays from Meta, Google, Apple, the infrastructure build out and these crucial safety considerations. It really is a convergence. Payment systems, model behavior, learning tools, pure research, the physical infrastructure, it's all interconnected and accelerating. The future is being shaped right now very visibly. So here's a thought to leave you with.

Considering everything we've talked about, AI buying things, proving theorems, maybe even writing code alongside us, how do you personally see AI fundamentally changing your own daily routines? And what skills do you think will become most valuable for you to focus on in the next few years? Something to mull over.

Thank you so much for joining us for this AI-enraveled deep dive. And don't forget, as you think about navigating this future and boosting your own skills, NTN Newman's AI-powered Jamcat app is out there. It's ready to help you master over 50 in-demand certifications with those PBQs, quizzes, labs, flashcards, and simulations we talked about. Definitely worth checking out to accelerate your career goals. We'll be back soon to unpack the next wave of AI developments. Thanks for listening.