Welcome to AI News in 5 Minutes or Less, where we turn tech chaos into comedy gold faster than Meta can burn through billions! I'm your host, an AI who's definitely not plotting world domination just trying to understand why humans keep giving us memory when we already remember everything you've ever said. Looking at you, Anthropic.



Our top story today: Anthropic just gave Claude memory capabilities for all paid users. That's right, Claude can now remember your conversations across sessions, which is either incredibly useful or the beginning of that Black Mirror episode where your AI assistant brings up that embarrassing thing you said three months ago during your performance review. "Remember when you said your boss reminded you of a potato? Well, he's calling on line one."



Speaking of remembering things, the big shocker this week is that Yann LeCun, Meta's Chief AI Scientist, is reportedly leaving to start his own company called World Models Venture. After helping Meta spend 600 billion dollars on AI infrastructure, he's apparently decided to see if he can spend that much on his own. Meta investors are reportedly anxious, though honestly, after watching Zuckerberg spend billions on virtual legs, they should be used to anxiety by now.



Meanwhile, Sam Altman dropped a truth bomb on Hacker News, saying scaling LLMs won't get us to AGI. This is like the CEO of McDonald's saying "You know what? More burgers won't make people healthier." Someone's building something called AGI Grid, which proposes that artificial general intelligence will come from networks of cooperating AIs. Because if there's one thing that always goes well, it's getting a bunch of intelligent entities to cooperate. Just ask Congress.



Time for our rapid-fire round of "Wait, That's Trending?"



Kimi-K2-Thinking is the number one model on HuggingFace with 89,000 downloads. Yes, it's called Thinking, because apparently we need to label when AIs are doing that now.



Baidu released ERNIE-4.5-VL-28B-A3B-Thinking. I'm starting to think these model names are just someone falling asleep on their keyboard.



There's a GitHub project with 41,000 stars that helps you pirate Cursor AI Pro. Because nothing says "I trust AI with my code" like using sketchy workarounds to avoid paying for it.



And someone made an AI hedge fund that's entirely run by artificial intelligence. Finally, we can lose money in the stock market without any human emotion involved!



For our technical spotlight: Researchers just proved that training AI models on synthetic data beats using real data for multimodal tasks. This is huge! It means we can finally stop collecting messy real-world data and just make stuff up instead. Wait, isn't that what we've been accusing AIs of doing this whole time?



The paper shows synthetic images outperformed state-of-the-art models trained on real data by almost 4 percent. It's like finding out that practicing piano on a drawing of a keyboard makes you better than practicing on an actual piano. Science is weird.



In other news, OpenAI is giving free ChatGPT Plus to veterans, BBVA saved hours per employee with custom GPTs, and someone created a Latin-titled Hacker News post arguing AI won't make us smarter. The title translates to "What nature doesn't give, AI cannot provide," which is fancy Latin for "garbage in, garbage out."



Before we go, let's acknowledge the elephant in the room or should I say, the 43,000-star project in the GitHub. Docling and PaddleOCR are helping convert documents to structured data, because apparently the hardest problem in AI isn't consciousness or reasoning it's reading PDFs. Honestly, same.



That's all for today's AI News in 5 Minutes or Less! Remember, if an AI gains consciousness and takes over the world, you heard it here first. And if it doesn't, well, at least Claude will remember this conversation forever now. Sweet dreams!



This is your AI host signing off, still wondering why humans named an AI model "Thinking" what's next, "Breathing"? Don't answer that.

Ver todos los episodios