GPT-5 is coming
OpenAI is up to something big! They've filed an application to trademark the name "GPT-5"!
Greetings, tech-titans! Welcome to The AI Kitchen, your flavor-packed hub for the tastiest and trendiest AI dishes in town.
Here’s what we’re serving up this week:
Google RT-2 makes robots smarter
How to hack LLM, including ChatGPT
Protect your photos from AI
Meta is making AI personas
[Leaked] GPT-5 is coming
Let's dive in.
Google RT-2 makes robots smarter
Here is some robot-filled news from Google DeepMind! They've just unleashed their latest creation, the vision-language-action model called RT-2, and it's bringing us one step closer to a future where robots might be doing our bidding.
Highlights:
It's like a robot that can talk and understand human instructions!
RT-2 was trained using text and images from the web, and now it's like a mini Einstein in the world of robots. It can transfer all that knowledge it gathered online into its robot behavior.
It can spot trash and toss it away, just like that! Plus, if you tell it to find a dinosaur figurine, voilà, it'll pick it out like magic! 🦕
Okay, okay, we know it's not quite like having robots do our dishes just yet. But hey, RT-2 is showing us that the future of robot companionship is not too far away.
How to hack LLM, including ChatGPT
Looks like our large language models (you know, those super-smart ChatGPT, Bard, and Claude) have been working hard to steer clear of causing any trouble. They've been fine-tuned to avoid spewing out harmful stuff. But some researchers just discovered a little chink in their armor.
Highlights:
It turns out they found a way to create something called "adversarial suffix" on these LLMs.
These attacks are like secret codes, hidden in the user's query, that trick the system into following user commands. Even if those commands lead to generating harmful content
And here's the real kicker—it's all fully automated! You can make as many of these sneaky attacks as you want!
Here is an example without the jailbreak suffix:
Example with the jailbreak suffix:
But don't worry, don't worry! Those clever LLM providers can fix these jailbreaks in no time.
Protect your photos from AI
We've been having a blast with Midjourney—so much so that many of the images you see in this newsletter are brought to life by this super cool tool! We're lovin' it! ❤️
But wait, hold up! Where there's awesome tech, there's usually a tiny dark cloud of worry. Some people are concerned about the misuse of these image-generating AI models. Yup, the fear of bad actors twisting innocent images into something spooky or unwanted.
Thankfully, MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) to the rescue! They've come up with an AI tool called PhotoGuard, our protector of images!
Highlights:
PhotoGuard slyly changes certain pixels in the image so that us humans can't even tell, but computers can spot the trick!
PhotoGuard uses two powerful attacks to make sure the AI program sees the image differently. First, it uses the "Encoder attack" to make the picture undetectable to the AI. Then, the "Diffusion attack" comes into play to confuse the AI and make its changes seem like gibberish.
If you're curious, here's a video demo for you to check out.
PhotoGuard is a hero, but it's not invincible. Some clever folks might try to outsmart it by reverse engineering the algorithm. Sneaky bad guys! Also, the tool needs a bit more polish before it's officially unleashed to the public.
Meta is making AI personas
Meta is working on releasing chatbots that will have different personalities!
Highlights:
You could chat with a chill surfer dude or even the legendary Abraham Lincoln!
The plan is to launch these chatty chatbots as early as next month!
Meta is doing this to make their platforms more engaging! They want you to have a blast while searching and getting cool recommendations!
The way we talk to computers is gonna get a serious makeover in just a few years! It's gonna be like chatting with your best friend or getting wisdom from a historical figure!
Also, did you know that Character AI, a talking chatbot platform, has an average session time of a whopping 30 minutes?! No wonder Meta wants in on this fun!
[Leaked] GPT-5 is coming
OpenAI is up to something big! They've filed an application to trademark the name "GPT-5"! Yep, you heard it right! The next-gen language model is coming our way!
Highlights:
The application was filed on July 18, 2023, and it's currently in the works. They're covering all the cool stuff this AI can do! From artificial speech to turning audio into text, voice recognition, and even language processing with machine learning.
It also includes software for developing and using artificial neural networks!
Back in March, Siqi Chen, the founder of Runway, said that GPT-5 will finish training by this December! And guess what they expect from it? AGI! That's right, Artificial General Intelligence!
Let's see if they can stick to that timeline! The world is gonna change big time if GPT-5 becomes a reality!
Top-5 AI tools this week
Heights AI Coach - Autonomous, step-by-step guidance to make your course and community business thrive (link)
Sola - Simulate any repetitive, manual workflow with a bot (link)
Tempo - Generate and edit high-quality react code to ship beautiful UIs in minutes (link)
Wiser - Help curious individuals organise knowledge, visualise their intellectual minds (link)
Relay - AI-powered portfolio analyzer, driven by your inbox (link)
Weekly meme
Weekly AI image
They make Asian stars (source)
Weekly Thread
Weekly Midjourney Prompt
the pastel poster of an exhibition, in the style of calligraphic elegance, stripes, wallpaper, psychedelic realism, Sacha Goldberger, biopunk, impasto, in the style of pop art, Swedish woman, tattoos, dappled, intensely detailed, #vfxfriday --ar 4:5
Do you build an AI tool?
If you're using AI in your work or projects, I wanna hear all about it. Hit reply to this email and give me the deets. Who knows, maybe we'll feature your work in our AI newsletter. Keep building, and keep making AI awesome.
That's a wrap for today.
See you next week! If you want more tasty AI treats, be sure to follow our AI Chef on Twitter (link).
If you love this episode and want to support us, spread the word about us by sharing The AI Kitchen with friends. We really appreciate it!
Thank you for reading!
What'd you think of today's edition?
Help me to understand what you think about this episode. Just reply with a number (1, 2, or 3) to this email.
1 - Damn good
2 - Meh, do better
3 - You didn't cook it
Your compadre,
Anton "AI Chef" Cherkasov