AI Weekly News Rundown April 13 to April 20 2025: ⚡️Microsoft Researchers Create Super‑Efficient AI 🤔OpenAI’s New Reasoning AI Models Hallucinate More 💥Chipmakers Fear They Are Ceding China’s AI Market to Huawei 🧑‍💼AI-Powered Fake Job Seekers
19:43 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 a passionate soccer dad up in Canada. Great to be back. And hey, if you're enjoying these deep dives, if you find them valuable, please do take a second to like and subscribe on Apple Podcasts. It really helps us out. It does. And, you know, the pace in AI...

it just doesn't slow down, does it? It feels like every week there's this constant stream of news breakthroughs. Totally overwhelming sometimes. Yeah. Which is exactly why we do this. We try to cut through that noise. Right. We've looked at a whole bunch of sources this week. I mean, everything from super energy efficient models, uh,

robots running races, even AI writing news articles. Yeah, quite the mix. So our mission as always is to pull out the most important bits, connect the dots for you so you feel like you get the bigger picture without drowning in details. Exactly, it's about context. And this week really shows that speed of innovation, but also some really interesting challenges creeping up. Okay, let's dive in then. One of the first things that caught my eye was this Microsoft BitNet B1.58 update.

They keep calling it a one bit AI model. What's the deal with that? Like, what does one bit actually mean here? Well, it's not literally just one bit, but the core idea is radical efficiency. Usually these big AI models, they need tons of power, right? Specialized hardware, GPUs.

Lots of energy. Right, the expensive stuff. Exactly. This one-bit approach, it uses just 1.58 bits for each parameter, which is tiny. It drastically simplifies how the AI stores and works with information. Okay. And the proof is in the pudding. It runs on standard CPUs, even the ones in Apple's M-series laptops. That's how efficient we're talking.

Wow. OK. They mentioned what? Up to 96 percent less energy? Yeah. It's staggering. Yeah. And it's tiny, too, right? 0.4 gigabytes. Fits on a laptop. It still performs well. It has like two billion parameters trained on four trillion tokens. Those still sound like big numbers. They are significant numbers. It shows how clever model design is getting. Researchers are finding ways to get similar performance, but, you know, much leaner, much greener. So the big picture impact here is accessibility.

Democratization. I think so. It really could be. Imagine powerful AI running on everyday devices, not needing giant data centers or gobbling up energy. That changes things. OK, but then there's this other side of the coin. News from OpenAI about their new reasoning models, 03 and 04 Mini. They're supposed to be smarter, better at reasoning. Yeah, that's the goal. But there's a catch. They seem to be making more mistakes.

more hallucinations, as they call them. That's the really interesting, maybe slightly worrying part. OpenAI's own tests showed a pretty noticeable jump in hallucinations. Wow, how much? Well, on one benchmark, person QA, O3 hallucinated 33% of the time. That's double the rate of the previous model, O1. And a 4 mini, even higher, 48%. Wow. So they reason better, but they're less accurate.

that feels backward. It does feel counterintuitive, doesn't it? You'd assume smarter means more reliable. Yeah. And what's fascinating or maybe concerning is that the researchers admit they don't fully get why this is happening. Our understanding of how these models develop reasoning and why it might sometimes trade off with factual accuracy

It's still incomplete. Maybe it's like as they get better at connecting dots, they sometimes connect dots that aren't really there. That's one plausible theory. They might be getting better at inference, but also more likely to stray from the training data. So what's the takeaway for us? We get smarter AI, but maybe we have to double check it more often. It really highlights the need for alignment for safety measures. As AI gets more capable, making sure it's truthful and reliable is even more critical.

It shows progress isn't always linear. You know, pushing one boundary can create challenges elsewhere. OK, let's shift gears. Geopolitics and AI chips. It sounds like U.S. chip makers are getting nervous about China. Yeah, there's definitely concern specifically about losing market share in China's AI sector to Huawei. Because of the new trade restrictions from the U.S. government. Exactly. The U.S. limited sales of advanced AI chips, especially from companies like NVIDIA to China, is

And that, well, it creates an opening. An opening for Huawei to step in. That's the worry. Huawei is huge in China and they could potentially fill that gap left by U.S. companies. And I heard there are even investigations into whether those export rules are being broken.

Sounds tricky. Very delicate. Some analysts think these restrictions might actually push Huawei to develop its own advanced chips faster. It could even spur more global competition in the long run. So global politics is directly shaping the AI hardware market. Absolutely. It's reshaping the whole landscape, potentially creating new leaders in this really critical area. All right. Now for something completely different. Robots.

Humanoid robots running a half marathon in China. That sounds like sci-fi. It really does capture the imagination. Yeah, 21 humanoid robots ran alongside humans in Beijing. And did they? Did they finish? One of them did. Tiangong Ultra completed the whole 21 kilometers, took it two hours and 40 minutes, which is pretty impressive. That's faster than I could do it.

But I guess not all have made it. That's right. Some others had difficulties. It shows the tech is still early. You know, navigating the real world for that long, it's hard for a robot. Sure. But the fact it even happened, it really showcases China's push in robotics and AI. They're not afraid to test these boundaries publicly. OK, let's pivot to the business side.

Johnson & Johnson looked at their AI projects and found something interesting about value. Yeah, this is quite insightful. They found that only about 15% of their AI use cases were actually delivering 80% of the total value. The classic 80-20 rule, but for AI.

Pretty much. It suggests that after maybe an initial phase of broad experimentation, companies are getting more strategic. J&J saw the biggest impact in things like supply chain optimization, manufacturing automation, and R&D. So they're refocusing.

Putting their resources where they get the biggest bang for their buck. Makes sense. It does. And it probably signals a wider trend, right? Moving from just trying stuff out with AI to really targeting specific areas for deployment where you can see a clear return on investment. Less scattergun, more focused strategy. Speaking of trying stuff out, an Italian newspaper went pretty bold. They let

and AI be the editor for an entire edition. Yeah, that was a fascinating experiment. An entire issue written and curated solely by AI. How did it go? Apparently, the human editors were quite impressed. They said the AI showed a surprising grasp of irony, even offered nuanced commentary. Given our chat about hallucinations, were there concerns about accuracy?

misinformation. Absolutely. That warning came alongside the praise. It really highlights this tension, doesn't it? AI is getting incredibly capable, even in creative roles like writing and editing. Yeah. But it also brings that whole debate about authenticity, trust, and the risk of AI spreading false information right back to the forefront, especially in journalism. Okay, let's talk about the job market. Seems like AI is causing headaches there too, but in a different way.

Fake job applicants. Yeah, recruiters are seeing a big uptick in this. People using AI to generate pretty convincing resumes, cover letters. Tailored to the job description, I bet. Exactly. And it goes further. AI voice avatars for interviews, fake work histories, even AI generated portfolios. Whoa.

That makes hiring incredibly difficult. How do you even tell who's real? It really complicates vetting candidates. The implication is, well, employers might need more than just resumes and interviews now. Like AI detection tools for applications. That's likely part of it. Developing new ways to verify skills and experience becomes crucial when AI can fabricate credentials so easily.

It's a new challenge for HR. Okay, quick pause here. If you're finding this deep dive useful, maybe you're thinking about leveling up your own skills in tech or business. Good point. Well, you should check out the JamGatek app created by our producer Etienne. It uses AI to help you study for and pass over 50 different professional certifications. Cloud, cybersecurity, finance, healthcare, you name it. A practical use of AI for learning. Exactly. Links are in the show notes if you want to learn more.

Okay, back to the news. Google's Gemini 2.5 Flash, it has this thing called a thinking budget.

Sounds interesting. It's a neat concept. Basically, it lets developers control how much computational effort the AI puts into reasoning for a specific task. So you can dial it up or down? Precisely. For simple tasks, use a smaller budget, save resources, get faster answers. For complex problems, crank up the budget for deeper thought. It's about optimizing performance versus cost and speed. Smart. Like allocating brain power efficiently.

Kind of, yeah. And apparently, even though it's designed for efficiency, Gemini 2.5 Flash shows big reasoning improvements over the previous version. This kind of granular control could make AI much more adaptable and cost-effective across the board. Now, this next one feels a bit creepy. People using ChatGPT to figure out where photos were taken, even without location data. Yeah, that trend is definitely raising eyebrows. Users upload a photo, ask ChatGPT, where is this? And the AI analyzes visual clues. Like what? Buildings?

Exactly. Architecture, vegetation, street signs, cars.

It pieces together these clues, searches the web, and makes surprisingly accurate guesses about the location. OK, technologically impressive, but yeah, unsettling. Major privacy red flags there. Huge ones. It really shows how much AI can infer from seemingly harmless data. It underscores the need for, well, serious conversations about privacy safeguards and ethical use, doesn't it? Makes you think twice about what photos you share online. Definitely.

OK, let's look at the money side again. Meta reportedly asking Amazon and Microsoft for help funding Llama, their big language model. Yeah, that story really highlights just how incredibly expensive it is to develop these cutting edge large language models. I mean, Meta is huge.

if they're looking for help. It tells you something, right? The computing power, the massive data sets, the specialized engineers, it all adds up to an enormous cost. So meta reaching out to rivals suggests maybe partnerships are becoming necessary. It seems likely. We might see more collaboration, even between competitors, just to share the financial burden of staying at the forefront of AI R&D. Okay, shifting to science. A biotech startup, ProFluent, found scaling laws for AI and protein design.

What does that mean for medicine? This could be really big for drug discovery and synthetic biology. Proflint found that just like with language models, making protein design AI models bigger and training them on more data leads to predictably better results. So bigger is better for designing proteins, too. Seems so. Their latest model, 46 billion parameters trained on 3.4 billion protein sequences, is apparently designing things like antibodies and gene editors really effectively.

And they're sharing some of this open antibodies. Yeah, they're making some designs publicly available. This could dramatically speed up research, making it faster and cheaper to design proteins for specific jobs, new drugs, gene therapies, maybe even new materials. It's potentially transformative.

Wow. And even spreadsheets aren't immune to AI. Google Sheets has an AI formula now. That's right. The AI formula and a help me organize feature. It basically brings AI smarts right into your spreadsheet. What can it do? Things like generating text summaries, analyzing data trends, creating custom outputs, basically automating tasks and helping you manage data more easily. So like a built-in data analyst assistant?

Pretty much. It aims to save time on the tedious stuff, improve accuracy and just make data work easier for more people. OK. Meta's research arm fair has been busy, too.

New stuff in AI perception, understanding the world. Yeah, they're working on improving how AI sees and understands its environment. They have a new perception encoder doing well on visual tasks. And they released open source tools. Right. The meta perception language model, PLM and PLM VideoBench, plus things like Locate3D for better object understanding. It's foundational work, really. Stuff needed for better robotics, better augmented reality. AI.

AI that truly sees and interacts with the physical world. We talked about OpenAI's 03 and 04 mini reasoning, but you mentioned they can also work with images now. Yes, that's a key upgrade. They don't just process text. They can look at images, sketches, diagrams, and incorporate that visual information into their reasoning. And use chat GPT tools too. Exactly. They can browse the web, run code, generate images, all integrated. It makes them much more versatile, more powerful multimodal assistants.

And these assistants are popping up on phones more. Perplexity AI on Motorola, maybe Samsung soon. Yeah, that's a big strategic push for perplexity. Getting pre-installed on phones puts them right up against Google Gemini. So more competition in the mobile AI space. Looks like it. More choice for users, which usually drives innovation. The mobile AI landscape seems to be heating up. And speaking of deals, OpenAI might buy Windsurf.

a coding assistant company for $3 billion. That's the report. If it happens, it'd be OpenAI's biggest acquisition yet. It clearly signals how important AI for coding is becoming a major area of focus and competition.

But not everyone's playing nice. Meta blocking some Apple intelligence features on its iOS apps. Yeah, that's an interesting move. Disabling things like Apple's writing tools and Genmoji within Facebook, Instagram, WhatsApp. Why would they do that? Well, speculation is they want to push their own Meta AI features. Plus, you know, there's always been competitive tension between Meta and Apple. It definitely impacts users on iPhones, though. Microsoft's co-pilot studio has a new computer use feature, AI, that can actually...

Use your computer. Sort of, yeah. It lets AI agents interact with websites and desktop apps by simulating clicks and typing like a human would. Even on older systems without APIs. That's the key benefit. Automating tasks like data entry or invoice processing on systems that aren't easily connected otherwise.

And Microsoft emphasizes the processing is secure and enterprise data isn't used for training. For people worried about privacy, you mentioned running AI locally. Absolutely. Tools like GPT-4 All and Alama let you run AI chatbots right on your own machine offline. So your data stays completely private. Exactly. You download the tool, download a model, and chat away without anything leaving your computer. It's becoming much more feasible for regular users now. And Anthropix Cloud is getting smarter too. Autonomous research powers.

Google Workspace Integration. Yeah, big upgrades for Claude. A new research feature lets it search the web and internal company documents on its own. And pull info from your Gmail, Docs, Sheets. Right. That Google Workspace Integration gives it much more context about what you're working on. It positions Claude as a really powerful context-aware assistant, especially for businesses.

It's rolling out in beta now for certain plans. On the ethical front, Wikipedia is offering AI devs a proper data set via Kaggle to stop scraping. It's a smart, positive move. Instead of bots hammering their servers, the Wikimedia Foundation is providing a high-quality, structured data set. Good for ethical AI training. Good for Wikipedia's resources. Exactly. Encourages responsible development and protects their infrastructure. Win-win. But AI still messes up.

That story about the cursor support agent inventing a fake policy? Oof, yeah. A clear example of AI hallucination causing real problems. The company had to apologize. It's a reminder, isn't it? You can't just let these things run unsupervised, especially talking to customers. Definitely not. It highlights the need for safeguards, human oversight, especially in critical roles. We're not quite at...

set it and forget it yet. Google's making its AI premium plan free for college students. Yeah, quite a move. Students with a .edu email get access to Gemini Advanced, Gemini 1.5 Pro, Workspace Integration, basically the top tier stuff for free. Investing in the next generation of users, I guess. Seems like it. Get them using Google's AI tools early on. It could definitely help students with their work. And finally, MIT researchers found a way to make AI better at writing code.

without retraining. Yeah, a new technique using clever prompting to guide LLMs to follow programming syntax more accurately. So fewer bugs in AI-generated code. That's the goal, making AI coding assistants more reliable, which would be a big help for developers. Wow, we covered so much ground. And that wasn't even everything, was it? There was a ton of other news this week. Absolutely relentless, just rattling some off. OpenAI's O3 acing an IQ test, and Lorena becoming its own company, those perplexity phone deals we mentioned.

XAI's Grok getting memory and workspaces, Alibaba's new open source video model. These are filtering AI songs, OpenAI maybe buying another company, AnySphere maybe building a social network. NVIDIA taking that $5.5 billion hit from China restrictions, Anthropix rumored, VoiceAI, Grok Studio, Kling AI 2.0 for video image generation, templates for building personal data analysts. AI playing detective, an ace attorney, kind of. Mm.

Potential delays for Trump's AI plans. Humans being bad at spotting deep fake voices. Hugging face buying a robotic startup. ChatGPT getting an image library. OpenAI debuting GPT 4.1. Apple's plans for analyzing user data privately. Slop squatting attacks using AI hallucinations. ByteDance's Seaweed7B video model. Google trying to talk to dolphins with AI. Google AI Studios branching feature.

OpenAI updates on safety and libraries. XAI rolling out Grok memory in studio. Cohere releasing embed for Google releasing VO2. NVIDIA's first US AI manufacturing. OpenAI may be targeting O304 mini for science ideas. Yeah. Andy Jassy on GAI, Meta's EU data training plans. Hugging face buying pollen robotics, Ricci2. LM Arena search leaderboard.

Palantir's NATO contract, SSI's massive funding round, AI beating experts at TB diagnosis. XOpenAI staff pushing back, AI lead outreach tools, NVIDIA building supercomputers in the US, more on Google's Dolphin AI. AI action figures trend, Google NVIDIA investing in SSI, DeepSeek v3 deprecated,

high schooler finding space objects with AI, Lama for Maverick ranking well. DeepMind CEO on combining models. Netflix may be using OpenAI for search. OpenAI's verified org status. Musk's XAI rolling out memory. It's a lot.

lot. It truly is. So key takeaways. AI efficiency is improving drastically, like with BitNet. But reasoning advancements like OpenAI's 0304 come with reliability challenges, those hallucinations. Right. Geopolitics is hitting the chip market hard. AI is weaving into everything. Robots, news, jobs, spreadsheets, businesses are getting smarter. Focusing on AI's actual value, it's ROI. And privacy and ethics are constant, crucial threads running through all of it.

Exactly. So for everyone listening, which of these things feels like it could have the biggest impact on your work, your life? Lots to chew on. And maybe a final thought to leave you with. As AI keeps accelerating like this, becoming even more integrated, how do we as individuals, as companies, as society,

How do we best prepare for that future? A big question. Definitely. And hey, one way to prepare is to boost your own skills. Don't forget to check out Etienne's AI-powered Jamgat Tech app. It helps you master certifications in cloud, finance, cybersecurity, healthcare, business, and more. Seriously useful. Links are in the show notes. Good reminder. Thanks so much for diving deep with us today. Keep exploring this fascinating, fast-moving world of AI. Until next time.