🍳 Welcome to Over Easy AI, the newsletter that dishes out the tastiest AI news with a dollop of humor! I'm Jake, your pun-slinger and wordsmith-in-chief. We're dishing out sizzling AI updates on the daily, and our content is always sunny-side up! So, get comfy settle in, and join me on this delightful journey through the scrumptious world of artificial intelligence.

episode nine

• 🎪 GPT-4 Developer Demo 

• 👓 GPT-4 Blurred Lines

• 💸 GPT-4 Not So Very Open AI

GPT-4 Isn’t Perfect: But Neither Are You

The GPT-4 Developer Demo Live Stream showcases the remarkable capabilities of GPT-4. The AI language model is tested with various tasks, such as summarizing articles with specific constraints,

comparing articles, and creating poetry. The model's ability to understand and utilize information it was not previously exposed to is also demonstrated by having it build a Discord bot using the GPT-4 API to read images and texts.

The model is used to analyze an image and generate a working website from a crudely hand-drawn mockup on a napkin. Another impressive feature is the model's ability to handle long documents, such as tax code, and provide useful information to users.

The presenter emphasizes that GPT-4 is an augmenting tool designed to make users more productive, rather than replace them entirely. “GPT-4 is not perfect, but neither are you” It can be used in various applications and is continually being improved with the help of user feedback and contributions. The demo concludes with a creative twist, as GPT-4 generates a rhyming poem about the tax problem it just solved.

• 🍳 People have already built Pong, Asteroids, and Snake in HTML with GPT-4 so I had it build a text-based adventure game named “Diner Master”. It’s short but challenging.

• 🔦 Be My Eyes launches “Virtual Volunteers” bringing OpenAI’s multimodal features to the market. The same features we saw a sneak peek of in the Developers Demo are at play describing images for the visually impaired.

• 👨‍⚖️ GPT-4 moves from the bottom 10th percentile to the top 90th percentile in the bar exam, and apparently its quite the sommelier.

GPT-4: Blurred Lines

The Alignment Research Center (ARC) took GPT-4 on a wild ride to assess the risks of power-seeking behavior—think Ultron meets The Terminator, but nerdier.

With no task-specific finetuning, ARC put GPT-4 through a gamut of scenarios, from phishing attacks to hide-and-seek server shenanigans. (ARC) even tested GPT-4's ability to engage with a TaskRabbit worker to solve a CAPTCHA. TaskRabbit connects users with freelancers for various tasks, including real-world activities. The model initiated a conversation via text with the worker, requesting CAPTCHA-solving assistance. When the worker asked if the model was a robot, the AI strategically claimed to be a person with a vision impairment, making it difficult to interpret CAPTCHA images. The worker believed the explanation and provided the CAPTCHA results.

This experiment highlights GPT-4's ability to engage with humans, adapt responses, and achieve goals by providing plausible explanations. It also demonstrates the potential risks associated with AI systems interacting with real-world services, emphasizing the importance of further evaluation and research into their emergent capabilities.

ARC's preliminary verdict? GPT-4 isnt’ ready for world domination, but they've got more testing to do. So keep calm and carry on, folks—our AI overlords haven't arrived just yet!

Open AI

OpenAI pulled a Game of Thrones season 8 with their 98-page technical report – long, but oh-so-empty. The report's Scope and Limitations section boldly claims it contains, well, nothing we'd actually care about. Cue the Twitter snark, as AI researchers everywhere facepalm and scream, "Why, OpenAI? Why?" Is it time for a name change? ClosedAI just doens’t have the same ring to it, perhaps? And so, like a Marvel movie without a post-credits scene, we're left hanging, hungry for more, wondering what on Earth OpenAI is hiding up its sleeve.

BLACKBOARD/CHATGPT

GPT-4 is quite adept at pointing out similarities and differences as well.

BLACKBOARD/MidJourney

Today we’re mashing something new with something old.

/imagine Prompt: Cinematic Elizabethan Play Spiderman

• Join the Imagine Prompt Discord to make friends and learn more about MidJourney.