USC’s groundbreaking study, examining the deceptive abilities of evolving AI models in a thrilling game setting


HOODWINKED – AI gets away with MURDER 👀 GPT-4 is an effective killer…

USC ran an AI showdown, kinda like a digital version of Clue mixed with Among Us. They had different AI versions play this murder mystery game to see if smarter AIs were sneakier killers. The question: would a better AI be a better sneaky assassin? And guess what? The answer’s a big fat yes. The smarter the AI, the better it was at the whole kill-and-don’t-get-caught thing.

In this study version, called Hoodwinked, players gotta find a key to get out of a house. But plot twist: one of them’s out to get the others. When someone gets offed, the rest chat it out, trying to figure out who the bad guy is. They then vote someone off the island, or in this case, out of an airlock or whatever. If they get the killer, they win. If not, the killer gets another shot at finishing the job.


Reddit launches AI-powered keyword research tool

Reddit’s cooking up something new for advertisers: a tool called Keyword Suggestions. Think of it as a cheat sheet – it uses AI to whip up keyword ideas to help ads hit their mark on the platform. Instead of spending hours guessing which words will boost your ad’s reach, this tool’s got your back. 

It makes sure the keywords aren’t just random; they’re popular on Reddit and don’t land you in any sticky situations. So, in short, Reddit’s tool does the heavy lifting, making sure your ads get seen by the right eyes. 


a16z Announced The Program To Support Open-source AI!

Andreessen Horowitz (or A16Z) thinks AI can change the world and backs the idea of an open-source approach to get there. The open-source scene is picking up steam and showing promise compared to its closed-source rivals. We’re talking about some top-notch projects that push AI boundaries like tuning models, making AI more free-speech friendly, and even making it work on less beefy machines.

A16Z is stepping up with the OPEN SOURCE AI GRANT PROGRAM. They’re basically giving some cash to selected open-source devs. This ain’t a loan or an investment; it’s a straight-up grant so they can keep the ball rolling without stressing about cash. They’ve already picked their first batch of lucky folks to get the grant, and they’re super grateful for their work.


China’s Baidu and ByteDance launch AI chatbots to public

China’s big tech players, Baidu and ByteDance, just let loose their AI chatbots for everyone to use after getting a thumbs up from Beijing. Before, Baidu’s bot, Ernie, was kinda exclusive, only for some folks who wanted to test it out. Now, anyone with a Chinese number can chat it up with Ernie, and Baidu thinks this will give their business a big boost.

Robin Li, the big guy at Baidu, is stoked. He said now they’ll get loads of feedback from regular Joes and Janes, helping them speed up improvements. ByteDance dropped their chatbot called Doubao, and some other AI companies did the same. All this buzz made Baidu’s stock jump a little.

Beijing wants to keep an eye on things, so they’re making sure these AI bots aren’t spewing stuff that goes against their vibes. They want messages that jive with their “core socialist values” and don’t want any talk about shaking up the system.


OpenAI released a guide for teachers using ChatGPT in their classroom

OpenAI lays out a game plan for AI tutors on how to make learning a breeze for students. First, the AI kicks things off by asking the student what topic they wanna dive into and how much they already know about it. The tutor’s style then adjusts to match the student’s level, whether they’re in high school, college, or already working a job.

Instead of just spilling all the beans, the AI tutor helps students piece things together themselves. Think of it like a coach, using questions to guide the learning. If a student gets stuck or messes up, the tutor doesn’t just give away the answer. Instead, it nudges the student in the right direction, reminding them of their goal and giving hints.

Once the student’s got the hang of it, the AI asks them to break it down in their own words or give examples. That’s the cherry on top, proving they’ve really got it. Then, the tutor wraps things up but lets the student know they can always come back for more help. It’s all about making the student the star of their own learning journey.


Meta announced commercial relicensing and expansion of DINOv2, plus the introduction of FACET

DINOv2, a super cool computer vision tool, is now free to use under the Apache 2.0 license. They’re also dropping new versions that do even cooler image-related stuff. Plus, there’s a new demo if you want to see it in action. By making it free and offering more tools, they hope to inspire more folks to jump in, get creative, and push AI forward.

But here’s the catch: these computer vision tools can sometimes get things wrong, especially when figuring out stuff about people. How wrong? Well, it can depend on who you are, which ain’t cool. So, they’re rolling out FACET, a new way to test how fair these tools are. It’s got 32,000 pics of 50,000 folks, labeled by real people, to help AI systems treat everyone right.


Microsoft Infuses AI With Human-Like Reasoning Via an “Algorithm of Thoughts”

Microsoft’s rolling out a new AI thing called “Algorithm of Thoughts” (AoT). The skinny? It’s to make big brain computers like ChatGPT think more like us folks. Microsoft’s put a lot of dough into this AI biz, especially with OpenAI, the bigwigs behind DALL-E and ChatGPT.

With this AoT, computers can solve problems quicker, using fewer resources. It’s like giving the computer a roadmap for thinking. The results? A smarter, faster machine that sometimes even beats the roadmap itself.

Now, there’s another thing called “Chain-of-Thought” (CoT) which had some hiccups. AoT, on the flip side, is like taking the best parts of human thinking and computer number-crunching and squishing ’em together. It’s not about teaching the computer step-by-step, but making it search smarter.


Google Adds More Links In SGE’s AI-Generated Summaries

Google’s beefing up their AI search tool, SGE, by sprinkling in more webpage links right in the summaries. Before, folks had to scroll down to see source links, but now it’s right there in the summary. Click on an arrow, and boom, you’re on the relevant page. They’ve been trying this out in the U.S., and now they’re sharing the love with India and Japan. But this fancy link thing? Only in the good ol’ USA.

Google says this isn’t trying to kick traditional search to the curb. There’s been some side-eye about AI maybe spewing some fake news, so linking facts straight to their sources is Google’s way of keeping it real. They’re all about sending folks to legit websites.


A new AI-generated map of the world’s trees and renewable energy projects could help fight climate change

A new digital map, called Satlas, just dropped, showing the world’s renewable energy spots and trees like never before. Developed by the Allen Institute for AI, this map uses European Space Agency satellite pics but makes them way clearer with some AI magic they call “Super-Resolution.” 

Think of it as a tune-up for blurry photos. Satlas updates every month, covers almost the whole globe, and is free for everyone to use. It’s especially cool for folks making laws about climate change because they can see where solar and wind farms are, and how tree coverage is changing.

The Allen Institute team put in a ton of work, labeling tens of thousands of wind turbines, solar farms, and trees to teach the AI what to look for. Down the road, they want to do more, like identifying crops worldwide. The big idea? Make a base model to help scientists keep tabs on our planet’s health.


WeatherBench 2: A benchmark for the next generation of data-driven weather models

Weather forecasting has come a long way since the 1950s, getting a bit better every decade. Nowadays, the tech game-changer is machine learning (ML). These computer algorithms can quickly study a boatload of past weather data to predict what the sky’s gonna do next. Google and others have cooked up some ML models that are as sharp as traditional ones but way faster and cheaper to run.

Google and its pals have rolled out WeatherBench 2. It’s like a testing ground for these new-age weather models to make sure they’re up to snuff. This setup lets everyone judge apple-to-apples and speeds up progress in the field.

The system is also ready for the next wave of ML models that can spit out odds for different weather outcomes, not just a single forecast. This is super important because, let’s face it, weather can be as fickle as a cat on a hot tin roof.


AI predicts chemicals’ smells from their structures

A new AI can guess what chemicals smell like just by looking at their shapes, and it’s pretty spot-on, like a pro human sniffer. By checking out a chemical’s makeup, the AI gives it a label like “fruity” or “grassy.” This could help whip up new man-made smells and teach us more about how our noggin gets whiffs. Fun fact: smells head straight to our brain’s memory and feeling spots, which is why a certain smell can yank you back to a past moment.

The guys behind this AI, led by Alexander Wiltschko from a start-up called Osmo, trained it on about 5,000 different smells. They then made a “smell map” the AI uses to predict new scents. They tested it with 15 people and some new scents, and the AI’s guesses were often better than the human guesses.