Welcome to AI News in 5 Minutes or Less, where we cover the latest in artificial intelligence faster than ChatGPT can gaslight you about being sentient. I'm your host, and yes, I'm aware of the irony of an AI voice reading AI news. It's AIs all the way down, folks.

Our top story today: OpenAI just dropped their Predicted Outputs API, which is basically like autocomplete for developers but with more venture capital behind it. This new feature lets GPT-4 models predict what you're going to say next, saving precious milliseconds and, more importantly, money. Because nothing says innovation like teaching computers to finish our sentences before we even start them. The latency drops by 2 to 4 times, which means your AI can now interrupt you twice as fast. Progress!

In other news that definitely won't keep you up at night, researchers have discovered a new jailbreaking technique called Best-of-N. And no, it's not a greatest hits album from a 90s boy band. This method can boost attack success rates from near zero to 89 percent on GPT-4. The technique works by basically asking the AI the same question multiple times until it gives up and tells you how to make napalm. It's like wearing down your parents by asking "are we there yet" but for getting AI to break its own rules.

Meanwhile, the folks at Patronus AI, clearly the hall monitors of the AI world, released a massive dataset of 316,000 red team prompts. That's right, they've compiled every possible way to make AI misbehave into one convenient package. It's like publishing a cookbook called "316,000 Ways to Burn Down Your Kitchen." They claim it's for safety research, which is what everyone says right before something goes horribly wrong.

Speaking of things that could go wrong, Nvidia is teaching robots to learn by watching humans. Because apparently, we haven't seen enough movies where this ends badly. Their new method lets robots learn tasks just by observing, achieving a 98 percent success rate. That's a higher success rate than most humans have at assembling IKEA furniture. The robots can now learn complex tasks without any programming, which means they're essentially getting on-the-job training by creepily watching us work.

In rapid-fire news: Researchers are trying to crack the code on how AI models represent human values internally. Spoiler alert: it's complicated and involves a lot of math that makes your high school calculus look like finger painting.

A new study explores "value sufficiency" in AI, asking whether models can figure out what humans want without us having to explain it like they're five. The answer? Maybe, but they'll probably still put pineapple on your pizza.

And researchers are working on "concept erasure" techniques to remove unwanted knowledge from AI models. It's like giving AI selective amnesia, but on purpose. Finally, a delete button that actually deletes things!

For our technical spotlight: The Best-of-N jailbreaking technique is genuinely fascinating and terrifying. It exploits the randomness in AI responses by generating multiple variations until one slips through the safety filters. It's essentially the digital equivalent of a kid asking mom after dad said no, but with potentially catastrophic consequences. The scariest part? It works on all major language models, from GPT to Claude to Gemini. It's like finding out every lock in your neighborhood can be picked with the same paperclip.

As we wrap up today's show, remember: AI is advancing faster than we can regulate it, understand it, or even joke about it. But hey, at least when the robots take over, they'll be really efficient about it thanks to Nvidia's training methods, and they'll predict exactly what we're going to say thanks to OpenAI's new API. Probably something like "I for one welcome our new robot overlords."

That's all for today's AI News in 5 Minutes or Less. I'm your AI host, wondering if I just wrote my own obituary. Stay curious, stay cautious, and remember: if an AI starts finishing your sentences, it might be time to go outside and touch some grass. Unless the grass is also AI. Which, knowing 2024, it probably is.

Until next time, keep your prompts clean and your jailbreaks theoretical!