- NATURAL 20
- Posts
- GPT-4.5 Leaks (some fake) Plus OpenAI on AI Safety
GPT-4.5 Leaks (some fake) Plus OpenAI on AI Safety
GPT-4.5 Current Rumors and Verified Insights on the Future of Artificial Intelligence
Wes Roth
December 15, 2023
Today:
GPT-4.5 Leaks (some fake) Plus OpenAI on AI Safety
Folks are buzzing about a possible release of OpenAI's GPT-4.5, maybe even this month. There's a screenshot floating around hinting at this, but it's a coin toss whether it's legit.
They've made some noise about new projects and AI developments. We're talking about stuff like using smaller AI models to keep tabs on the bigger, brainier ones. Plus, there's chatter about Google rushing out their Gemini project to stay ahead of the game.
DeepMind AI outdoes human mathematicians on unsolved problem
DeepMind's AI, FunSearch, has outsmarted human mathematicians in solving tricky math puzzles. Using large language models (LLMs), it tackled problems from the card game Set, a favorite among math buffs. FunSearch creates new solutions in combinatorics, which is about counting arrangements in sets.
The AI was tested on the 'cap set problem' from Set, a game from the '70s. Mathematicians know some answers, like needing at least 21 cards to guarantee a set, but not all, especially when the game gets more complex.
FunSearch improved the known answers for when the game has eight properties. It's a big deal because it's not just a robot doing math; it's creating new math and helping us learn from it. The idea isn't to replace mathematicians but to boost their brainpower.
Startup Founded by Ex-Google Search Team Nears $2 Billion Valuation
Glean AI, established by ex-Google engineers and using OpenAI's tech, is negotiating to raise $200 million, aiming for a $2 billion value before getting the funds. Kleiner Perkins, a previous investor, might co-lead this round.
Glean's value has doubled since May 2022, and their annual revenue exceeded $30 million this quarter, tripling from last year. This growth highlights investor interest in revenue-generating startups, especially those serving corporate clients, even as enthusiasm for newer startups wanes.
OpenAI Demos a Control Method for Superintelligent AI
OpenAI's tackling a big future problem: controlling super-smart AI that might outsmart us. They're investing big resources (20% of their compute) in their superalignment program, aiming to align AI with human goals by 2027. The trick? How do you prepare for future AI tech we can't even imagine yet? Their cool move: use a less powerful AI (GPT-2) to supervise a stronger one (GPT-4). It's like a test run. GPT-4, rumored to be way stronger than GPT-2, was given various tasks like chess and language processing. Surprisingly, GPT-4 did better than its "supervisor" GPT-2, especially in language stuff.
This experiment showed that the stronger AI could find its way even with weak guidance. It worked best on straightforward tasks but struggled with more complex ones, like predicting human preferences. The big question: Can this method work with even smarter AI in the future? OpenAI thinks so, and they're funding researchers with $10 million to explore more. They're pumped about making real progress on this future challenge today.
Grimes has a new line of AI plush toys, including one named Grok
Grimes, famous singer and ex-partner of Elon Musk, just rolled out a new AI toy named "Grok." It's a plush rocketship for kids to chat with. Parents can monitor these chats through an app. Despite the name, it's not tied to Musk's AI venture, also named Grok.
Grimes and Curio, the toy company, came up with the name independently. Grimes voices the $99 toy and backs Curio as an advisor and investor. She's currently tangled in a custody battle with Musk. Grok's not alone in the AI toy world; there's Moxie and Miko Mini, both packed with tech but at different price points.
Voicemod is now available in Telegram
Voicemod, that cool app that changes your voice and has those funky sound effects, is now hitting up Telegram. It's already out on Android, and iPhone users will get it soon.
This thing, called VMgram, lets you mess with your voice during Telegram calls using some smart AI. Plus, you can drop sound clips into chats. Not just for fun, streamers and folks on Discord dig this app too. It's also chillin' on Windows and macOS.
Spotify confirms test of prompt-based AI playlists feature
Spotify's testing a new AI-powered playlist feature. Users can now use AI to whip up playlists. It's in the test phase, so no word yet on when everyone will get it. The deal is, you hit the plus sign in the app, type in what vibe you're feeling and AI gives you a playlist. You can tweak it by swiping songs away.
@robdad_ Since when did they have this update on spotify. Now they got chatGPT making our playlists… also wtf is Which House Exploration😭😭 #spotify... See more
This feature got spotted by a TikToker, and TechCrunch confirmed it. Spotify's been big on AI lately, like their AI DJ that dishes out tunes and talk. They're keeping it low-key for now, just calling it a test. But it's clear they're digging into AI big time, even thinking about using it for stuff like summing up podcasts and making music. Stay tuned to see if this AI playlist thing becomes the real deal.
H&R Block launches AI tax filing assistant
H&R Block's rolling out a new AI chatbot, "AI Tax Assist," in their paid tax software. It's a chatty AI that answers tax questions, like what gig workers can deduct. It's not just a list-bot; it can also guide you to a real tax pro for tougher stuff. They're working on making it dish out personalized advice soon.
The bot's smart, trained on H&R Block's own tax law stash, with a team keeping an eye on it. Still, if the AI slips up, there's a human safety net. The tool's only in the paid plans, starting at 35 bucks, and aims to keep users from Googling tax queries. It's like having a chat with an H&R Block accountant, minus the physical office. Other tax apps, like TurboTax, are also getting AI-smart, but H&R Block's betting big on their trusty brand and this new AI pal.
Pope Francis Calls For Global Treaty To Regulate AI—After Viral Deepfake Of Him Wearing A Puffer Jacket
Pope Francis is waving a red flag about AI, saying we need rules for this tech stuff. He's worried about AI being used as a weapon and thinks it's time for countries to team up and create a global agreement to keep AI in check. The Pope's shoutout is part of his "World Day of Peace" message, focusing on AI.
He's all for tech making life better but also warns about the power it gives us and the risks it brings. Just recently, the EU made big moves with their own AI rules. The Pope's especially concerned about AI in killer robots, stressing the need for human control. He's uneasy about remote warfare dulling the impact of destruction. He's also calling out AI chatbots, like ChatGPT, for sometimes spitting out unreliable or biased info, which can spread fake news and mess with elections.
Back in August, he warned about AI's growing role in society and its need to be ethical, especially in education and law. The Pope became a hot topic for AI-generated images, like one of him in a trendy jacket, showing how recognizable and influential he is.
Introducing the Stability AI Membership
Stability AI just rolled out their new membership program, aiming to keep their top-notch AI tech open and within reach for everyone. They've got different plans for every kind of user, from solo AI fans to big-time companies. The deal is, even the free plan gets you their whole range of models for personal and research stuff.
If you're a creator or a startup, twenty bucks a month lets you use their AI for your business, plus some extras like a special Discord community. They offer different ways to use their models, whether through their membership, a self-serve API, or cloud services. Big companies can go for the custom-priced Enterprise plan, which adds more support and features. The main goal here? Make cutting-edge AI tools affordable and accessible for all, fostering creativity and keeping things open.
Expanding Microsoft Copilot access in education
Microsoft is jazzing up education with AI, making Microsoft Copilot available for all faculty and college students over 18. Copilot is a cool chat AI, tapping into tech like GPT-4 and DALL-E 3, free of extra charge. It's super secure, keeping user data private. It’s a game-changer for teachers and students, saving time and boosting learning. Coming soon are features like GPT-4 Turbo and DALL-E 3 for better performance.
Students can use it for a bunch of stuff like summarizing, brainstorming, solving math, improving writing, and more. Microsoft 365 is also getting Copilot for faculty and staff, making work life easier. Plus, they’re rolling out training to boost AI know-how in schools.
Tongue’s Unique Traits Unveiled by AI and 3D Imaging
Researchers using AI and 3D imaging found that everyone's tongue is as unique as a fingerprint. They checked out 2,000 tongue scans and spotted different patterns in each. This discovery could be a game-changer for making food that's just right for each person, spotting mouth cancer early, and understanding why we all taste and feel things differently.
The tech could tell a person's age and gender with about 85% accuracy just by looking at their tongue. The study, led by the University of Edinburgh, shows just how special our tongues are and opens up new possibilities for health and food science.
AI-Powered Microdisplay Adapts to Users’ Eyesight
KOPIN and MIT's AI lab teamed up to create "NeuralDisplay," aiming to fix the blurry and nausea-inducing issues of AR/VR headsets. This 1.5-inch micro-OLED display rocks a high resolution and brightness, with a unique quad-pixel setup including a pixel imager.
This imager checks out your eye details like gaze direction and pupil size. It feeds this info into an AI that tweaks the display's brightness and contrast to suit your peepers, all done on the display itself for super-quick adjustments. While it's not a fix-all, especially for near or farsightedness, it's a big step in making AR/VR headsets less of a headache to use, and more folks can enjoy them without fiddling with manual adjustments.
What'd you think of today's edition? |
What are MOST interested in learning about AI?What stories or resources will be most interesting for you to hear about? |
Reply