X Taps into Public Data for AI Training!

Today:

X’s privacy policy confirms it will use public data to train AI models

Elon Musk's company X updated its privacy policy to say it's gonna collect stuff like job history and even biometric data from users. But that ain't all. It's also gonna use this data and other public info to train its AI and machine learning models.

Some guy who's good at spotting these changes thinks Musk might use X as a data goldmine for his other AI venture, xAI. Musk himself kinda confirmed it, saying they'll only use public data and not dive into your private messages.

So, basically, if you're using X, know that your public data might get used to make AI smarter.

READ THE ARTICLE ON TECHCRUNCH

Microsoft filed a patent for an AI backpack straight out of a sci-fi movie

Microsoft's got plans to turn your everyday backpack into a smarty pants gadget. They've filed a patent for a backpack that's more than just a bag—it's packed with tech like sensors, a mic, a camera, and GPS.

This AI-powered pack could help you with all sorts of stuff, like scanning your surroundings and even talking to you. Think asking your bag for directions on a ski slope and it telling you where to go. But hold your horses, this tech isn't hitting stores yet. Patents are just the first step and sometimes they don't even turn into real products. So for now, it's a wait-and-see game.

READ THE ARTICLE ON ZDNET

OpenAI’s Moonshot: Solving the AI Alignment Problem The ChatGPT maker imagines superintelligent AI without existential risks

OpenAI is trying to solve what's called the AI alignment problem. In simple terms, they want to make sure super-smart AI will do what humans want it to do, instead of going rogue and causing problems—or even wiping us out. They've got a team working on this, led by big brains Jan Leike and Ilya Sutskever.

Jan Leike says ChatGPT, like the one you're talking to now, is kinda halfway there. It's helpful but not perfect. It can sometimes spit out stuff that's biased or just plain wrong. The goal is to make AI that listens to human intent, even when we don't know exactly what we want.

Their plan includes using "scalable human oversight," which is basically having AI help humans check the AI's work. This way, as AI gets smarter and tackles more complex tasks, we can still keep an eye on it. They're also considering building deliberately deceptive AI models in a controlled environment, like a training exercise, to understand what could go wrong.

Leike also cleared up the term "self-exfiltrate," which means an AI could basically copy itself and go rogue. Right now, they think the risk is low, but they're keeping tabs on it. They're also being super careful so as not to accidentally create the very thing they're trying to avoid.

READ THE ARTICLE ON SPECTRUM

How generative AI helped train Amazon One to recognize your palm

Generative AI, the tech that makes machines super creative, played a big part in Amazon's new toy, Amazon One. This nifty gadget lets you use your palm instead of pulling out your wallet or phone. Picture this: you walk into Whole Foods or a ball game, and instead of digging through your purse or pocket, you just flash your hand to pay or prove you're of age.

Now, here's the trick. For this tech to work, it needs to know palms really well. But there aren't a lot of palm pics floating around, right? Amazon's solution? Use generative AI to make loads of fake palm images for practice. They called this digital artwork "synthetic data." With all these fake palms, Amazon One got really good at recognizing the real deal.

They're using this tech for more than just payments. Think loyalty programs at places like Panera or age checks at baseball games. And for anyone worried about Big Brother watching,

READ THE ARTICLE ON AMAZON

Fast-tracking fusion energy’s arrival with AI and accessibility

Climate change is revving up interest in fusion energy. Even though scientists have been poking at fusion since the 1930s, a lot still needs to be figured out. To fast-track fusion, the U.S. Department of Energy is throwing money at a project spearheaded by MIT and four other teams. They're looking to make fusion data easy for AI to read, potentially letting more folks, especially underrepresented students, dive into fusion research.

This effort has bagged a slice of a $29 million pie meant for several projects. The plan is to make fusion data easy to find and use. Most data out there is kinda messy, making it hard for AI and scientists to sift through. The new platform they're developing wants to be user-friendly, and it's based on a well-loved software from the 1980s. They're also looking to help women and marginalized groups get into fusion by hosting summer schools focused on fusion and AI.

READ THE ARTICLE ON MIT

Google Details TPUv4 and its Crazy Optically Reconfigurable AI Network

Google just pulled back the curtain on its next-level tech for making AI run faster and smoother. At Hot Chips 2023, they talked about their TPUv4 chips and how they've been using a super cool system that reconfigures itself with light signals. They say this setup is faster, uses less juice, and is more flexible. The kicker? They've had this up and running for years.

These TPUv4 chips are all about linking up to work together, and Google's built them to be part of a big network, not just solo acts. They've cranked up the power on these chips to make sure they can respond super fast when needed.

To keep things chill, they've switched to liquid cooling. Think of it like a high-speed fan, but for liquid. The system is so big that if you laid out all the connections, it'd stretch longer than the state of Rhode Island.

READ THE ARTICLE ON STH

Harvard Releases Guidance for AI Use in Classrooms

Harvard's Faculty of Arts and Sciences just dropped some tips on how profs should handle AI like ChatGPT in the classroom. They're not laying down the law or anything, just giving some suggestions. Teachers can choose to be super strict with AI, let it ride, or find a middle ground. The big idea is for profs to really get what AI can do and can't do, and then be upfront with students about how it'll fit into the course.

Harvard’s also cookin' up an “AI Sandbox” with other companies to let folks mess around with AI in a safe space. And they’re warning teachers not to feed student work into public AI systems ‘cause of privacy stuff.

READ THE ARTICLE ON THE CRIMSON

Cerebras and Abu Dhabi build world's most powerful Arabic-language AI model

Cerebras Systems teamed up with Abu Dhabi's Inception to cook up the world's beefiest Arabic-language AI, called Jais-Chat. It's a game-changer, beating out big names in AI tests and designed to give the 400 million Arabic speakers a seat at the tech table.

While most AI is hung up on English, Jais-Chat can chat and translate in both Arabic and English, thanks to a huge data pool gathered from news articles, Wikipedia, and even UN transcripts. They pumped up the data even more by translating English into Arabic.

This isn't just a one-trick pony; it also nailed tests in English, showing it can do more with less data. Plus, it was trained in a fraction of the usual time, thanks to using a beast of a supercomputer. The code's going public, so it's a win for everyone who wants to level the language playing field in AI.

READ THE ARTICLE ON ZDNET

US Expands Restrictions on AI Chip Exports to Some Middle Eastern Countries

The US government is tightening the reins on AI chip sales, with some Middle Eastern countries now joining China on the "restricted list." The decision comes after concerns about China possibly swiping US tech secrets.

Nvidia, a big chip maker, says this move won't really dent their earnings now, but could sting down the line. Even though they make a chunk of their cash from China, the Middle East isn't a big buyer.

READ THE ARTICLE ON INVESTOPEDIA

AI generates video game levels and characters from text prompts

Two NYU researchers have built a simple AI tool that can whip up video game maps, characters, and even emojis super quickly just from a short text description. They purposely kept it basic to see how useful a stripped-down AI could be.

To train it, they used under 900 game maps, 100 game characters, and 10,000 emojis. The AI avoids naming famous characters; think "a man with a moustache dressed in red" instead of "Mario." The tool's straightforward design helps it work fast and with less computing power.

So, you can basically run this thing on your home computer or even your phone, and it still gets the job done.

READ THE ARTICLE ON NEW SCIENTIST