AI Unveils Millions of New Crystals

Today:

Millions of new materials discovered with deep learning

Researchers Amil Merchant and Ekin Dogus Cubuk just hit the jackpot with a smart tool called GNoME. This AI whiz found over 2 million new crystals, and get this – about 380,000 of them are rock-solid stable. That's big news because stable crystals are the backbone of fancy tech like computer chips, batteries, and solar panels.

These crystals are not just any crystals. They're the kind that can push tech into the future, like super-fast computers and better batteries for electric cars. GNoME, the brain behind this, is a deep learning tool that's way faster and smarter at finding these gems.

Labs around the world have already made 736 of these crystals for real, proving GNoME knows its stuff. Plus, GNoME's findings are getting added to a big online database for all scientists to use. This could really speed up research in the crystal world.

GNoME's a fancy type of AI called a graph neural network. It's like it understands how atoms hang out together, which is perfect for finding new crystals. The team trained it with tons of data on crystal structures, then let it do its thing to predict which ones would be stable.

This isn't just a little step forward. We're talking a giant leap – from a 50% success rate in predicting stable materials to a whopping 80%. And it's not just about finding these materials; it's about using less computer power to do it.

READ THE ARTICLE ON ON GOOGLE DEEPMIND

Amazon joins AI image creation fray with new model

Amazon's stepping into the AI picture-making game with its new Titan tool. This thing's not just any app; it's a fancy tool for developers to use for making their own image creators. To get in on it, these techies need to tap into Amazon Bedrock.

The AWS bigwig, Swami Sivasubramanian, gave everyone a peek at what Titan can do. It's not just about making pictures from words; it can also swap backgrounds. But here's the catch: it's mainly for the big business crowd, not your everyday folks like other image generators out there.

Every image made by Titan has a secret watermark. Why? Well, Amazon promised the White House they'd mark AI-made images clearly. This watermark isn't just slapped on the surface; it's baked right into the image, so you can't just snip it out or squish it away.

But how do you spot this invisible mark? Amazon's got a special API for that. Throw your image in there, and it'll tell you if it's a Titan original. This is a big deal 'cause marking AI stuff is high on the government's list.

Besides Titan Image Generator, Amazon's rolling out two other Titan models: Titan Text Lite for small writing jobs and Text Express for bigger tasks like chat apps.

READ THE ARTICLE ON THE VERGE

AI AT META PRESENTS CICERO

Meta's AI team just knocked it out of the park with a new AI named CICERO. This thing's a whiz at Diplomacy, a brainy board game where you gotta buddy up and wheel and deal with other players. It's like playing Risk, poker, and watching "Survivor" all at once, but the key to winning is smooth-talking your way to victory.

CICERO's a big deal because it mixes smarts in strategy and chit-chat. It can figure out what the other players are aiming for and then shoot the breeze to form alliances and hatch plans. This is new territory for AI – before, they could play games but not gab like a pro.

The tech wizards at Meta made CICERO by blending two geeky things: a fancy dialogue system and a brainy planning tool. Together, they let CICERO guess what others might do and then chat away to steer things in its favor. The AI can chat about game moves or just shoot the breeze like any human player would.

CICERO's so good, it scored double what the average Joe does on webDiplomacy.net, ranking in the top 10%. What makes CICERO tick? It's got a language model that's all about strategy. It doesn't just talk; it talks with a plan.

READ THE ARTICLE ON META

Together lands $102.5M investment to grow its cloud for training generative AI

Together, a hotshot AI startup that just scored a whopping $102.5 million in Series A funding. This big pile of cash, led by the folks at Kleiner Perkins and backed up by Nvidia and Emergence Capital, is like rocket fuel for Together's mission to expand their cloud platform. This ain't no small potatoes—it's over five times their last round of funding!

Their platform got datacenters in the U.S. and EU and can pump out 20 exaflops of compute. That's like having a supercomputer in your back pocket! They've got clients like NexusFlow and Voyage AI, and even Pika Labs, who just raised $55 million, is using their tech to churn out millions of videos.

Together is big on open source, too. They're cooking up AI models that are free for anyone to use and tweak. They've already released some neat tools like RedPajama and GPT-JT. The plan is to keep AI open and flexible, so businesses aren't stuck with just one option.

READ THE ARTICLE ON TECHCRUNCH

Microsoft Paint’s OpenAI-powered ‘Cocreator’ image generator is here

Microsoft just rolled out this cool new feature in their Paint app on Windows 11. It's called Cocreator, and it's like having a little artist in your computer. You type in what you want to see, and bam, it pops out three pictures for you to pick from. This magic is powered by OpenAI's DALL-E 3, a brainy AI that turns words into images.

Before, this feature was just for the Windows Insider folks, but now everyone with Windows 11 can use it. Microsoft isn't new to this game though; they've been playing around with this image-making AI in other places, like their Bing chatbot and a bunch of other Microsoft tools.

READ THE ARTICLE ON THE VERGE

AWS Announces More Model Choice and Powerful New Capabilities in Amazon Bedrock to Securely Build and Scale Generative AI Applications

Amazon just rolled out some cool updates to their Amazon Bedrock, making it easier for businesses to create and grow AI stuff. They've got a bunch of new models from different big names in AI, giving folks more options to pick what works best for them.

Now, they've made it simpler to figure out which model to use for a specific job. Plus, they're letting you use your own data to make AI responses more personalized and up-to-date. If you're into customizing, they're also adding ways to tweak models more to your liking.

For businesses that need AI to do complex tasks, Amazon's got this new feature that lets AI applications plan and do a bunch of steps securely. They've also added safety features to keep AI in line with company policies and responsible AI practices.

READ THE ARTICLE ON AMAZON

AI-enabled imaging company Aidoc raises $30M

Aidoc, a company that uses AI to improve medical imaging, just got a $30 million boost to make a top-notch AI imaging model and some new products. They help radiologists spot injuries and health issues in scans, and they've got special software for stroke and heart care, which makes it easier for medical teams to work together and share info.

They're excited about using AI like ChatGPT for medical imaging. Their system, aiOS, is made to be easy for hospital IT teams to use and set up. Michael Braginsky, one of the big shots at Aidoc, says their plan is to make AI in healthcare really take off with their aiOS and new AI model.

Before this, Aidoc raised $110 million and then $66 million to grow their AI imaging tech. They've teamed up with a couple of companies to add more features to their platform, like checking heart scans and X-rays for bone issues. They've also got the green light from the FDA for software that finds brain aneurysms and collapsed lungs in scans, plus other stuff like spine fractures and blood clots.

READ THE ARTICLE ON MOBI HEALTH NEWS

Boehringer Ingelheim and IBM Collaborate to Advance Generative AI and Foundation Models for Therapeutic Antibody Development

Boehringer Ingelheim and IBM are teaming up big time. They're gonna use IBM's fancy AI tech to come up with new antibody drugs. These drugs are super important for tackling tough diseases like cancer. Andrew Nixon from Boehringer is stoked about working with IBM to speed up finding these new treatments. IBM's got this pre-trained AI model that Boehringer will tweak with its own data. They're hoping this will make it way faster to find good antibodies for drugs.

Making these antibody drugs usually takes a long time and is pretty tricky. But with IBM's tech, they're planning to shake things up and make it quicker. They'll use computers to dream up new human antibody sequences. Then, Boehringer will test these ideas in the lab and keep improving the computer methods based on what they find.

READ THE ARTICLE ON IBM

GPT-4’s potential in shaping the future of radiology

GPT-4 is making waves in radiology—that's like, using X-rays and MRI images to figure out what's wrong with people. The folks behind this research are sharing their findings at a top-notch conference on natural language processing and AI.

The team behind the study teamed up with Nuance, a part of Microsoft, and they use this tool called PowerScribe that most U.S. radiologists dig. They're trying to see how GPT-4 can make a radiologist's life easier by handling stuff like understanding and summarizing reports, classifying diseases, and more.

GPT-4's already showing it's a cut above the rest. It's doing better than older models and even matching or beating human radiologists in making summaries. Plus, it's got a knack for structuring reports in a way that's easier for other healthcare pros to get and use.

But it's not just about making reports look pretty. By making this info more standard and searchable, it can help big time with research and clinical trials. And get this – GPT-4 can translate medical jargon into something folks can actually understand, which could be a game-changer for patient care.

READ THE ARTICLE ON MICROSOFT

Amgen expands pact with Amazon to usher drug manufacturing into the AI era

Amgen, a big player in medicine, is ramping up its team-up with Amazon. They've been buddies for 10 years, using Amazon's cloud service AWS. Now, they're bringing in some fancy AI to help make drugs faster. They're setting up a new plant in Ohio that's gonna be full of the latest digital and robot tech, using AWS and a tool called Amazon SageMaker. This setup aims to cut down on the need for people to step in during drug making.

They've been trying out AI and machine learning in Amgen's research and development, and now they're looking into using it in other parts of the business too. But Amgen isn't the only one playing this game. Sanofi's got a new app that gives a full view of their work, aiming to be a top AI user in pharma. They're using AI in research, clinical stuff, and making medicines.

READ THE ARTICLE ON FIERCE PHARMA

AI can be a game changer for neurodivergent employees

Imagine having a tough time getting your thoughts down on paper or chatting with your boss without feeling like you're stepping on toes. That's the daily grind for people with ADHD, autism, or dyslexia. But guess what? AI tools like ChatGPT and Google Bard are changing the game for them.

Take Billy Hendry, a guy from New Zealand with dyslexia, ADHD, and autism. He used to struggle with writing killer cover letters and emails. But this time, when he got laid off, he had a secret weapon: ChatGPT. He'd feed it what he wanted to say, and bam, it spit out words that made him look good without feeling fake.

Thanks to AI, Hendry landed a sweet job by nailing the cover letter and the interview. And he's not alone. More and more folks with these challenges are turning to AI to help them say what they mean without losing their cool or sounding off.

It's not about hiding who they are. It's about communicating in a world that's not always kind to people who are wired differently. Like, Jennifer Cairns uses ChatGPT to keep her emails short and sweet, while Christina Philippe, who has ADHD, uses it to stay focused and break down techy talk into plain English.

READ THE ARTICLE ON FORTUNE

AI can shore up federal cybersecurity overwhelmed by data, GDIT says

General Dynamics Information Technology (GDIT) found that U.S. government agencies, like defense and intelligence, are drowning in data. They need help sorting through all this info. Turns out, about 41% of the big shots in these agencies feel swamped. They're saying they need more brainpower and better tools to manage all this data.

This isn't just about keeping our data safe. It's about staying ahead of hackers, including heavy hitters like China and Russia. The Pentagon and other federal agencies are betting big on AI to sift through mountains of info fast and smart.

The Pentagon's putting $1.8 billion into AI next year and is juggling over 800 AI projects. Other agencies, like NASA and the Department of Veterans Affairs, are also upping their AI game big time.

READ THE ARTICLE ON FEDERAL TIMES

How machine-learning models can amplify inequities in medical diagnosis and treatment

Marzyeh Ghassemi, a computer science whiz from MIT, had a hunch that AI might make healthcare biases even worse before she earned her PhD in 2017. She's been digging into this ever since. In a fresh paper presented at a machine learning conference in Hawaii, Marzyeh and her MIT pals probed how machine learning can mess up medical diagnosis and treatment fairness.

They found four types of troublemakers: spurious correlations (thinking cows only hang out on grass, not sand), attribute imbalance (imagine men getting better pneumonia predictions), class imbalance (way more healthy folks than sick ones), and attribute generalization (predicting for ladies when there are no female pneumonia cases in the data).

They played around with 20 fancy algorithms and learned that tweaking the "classifier" can help with some issues but not all. And when it comes to measuring fairness, they figured out that trying to boost accuracy for the worst-performing group can mess up precision, which is a big deal in medical decisions.

Marzyeh and her gang aren't just theorizing; they're testing their ideas with real medical data. They want machine learning to be fair for everyone, but they know it won't be a walk in the park. So, they're taking on the challenge one step at a time.

READ THE ARTICLE ON MIT