Hi Reader,

Here are your top AI stories for the week:

• AI-driven education in 2 hours per day
• Add yourself to an AI-generated TV show
• AI models send "subliminal messages" to one another

Read more below! 👇

Sponsored by: Superhuman AI​ Find out why 1M+ professionals read Superhuman AI daily. AI won't take over the world. People who know how to use AI will. Here's how to stay ahead with AI: ​Sign up for Superhuman AI. The AI newsletter read by 1M+ pros. Master AI tools, tutorials, and news in just 3 minutes a day. Become 10X more productive using AI. ​Join 1 million pros and start learning AI​

​

🎓 AI-driven education in 2 hours per day

At Alpha School, students learn at their own pace using AI-driven software (and no teachers), and have high rates of achievement despite spending only 2 hours per day on academics. The remaining school day is spent learning life skills and exploring student interests, facilitated by adult “guides”. (Here’s one parent’s experience with the school.)

My take: AI-assisted education seems inevitable, given the enormous benefits that personalized learning can provide. The only questions are how quickly this model will spread to other educational institutions, and how deep the AI integration will be.

📺 Add yourself to an AI-generated TV show

“Showrunner” is a new AI-powered streaming service that allows users to add themselves as characters in animated shows, create new scenes using text prompts, and even generate their own shows. Showrunner's goal is to be a “new entertainment medium, one that more closely resembles video games.”

My take: YouTube succeeded as an entertainment platform by making it easy for users to upload their own videos. Whether or not Showrunner succeeds, I would bet that there’s an appetite for many different types of personalized, AI-generated entertainment.

🤯 AI models send “subliminal messages” to one another

If a LLM is fine-tuned to have a trait (such as “liking owls”) and then asked to generate an unrelated dataset of numbers, an LLM fine-tuned on that dataset will suddenly like owls as well. Here are the details of the experiment, plus a Twitter/X thread that is even more readable.

My take: The direct implication is that models could learn problematic behaviors even after we filter those behaviors out of the training data, leading to models that are not aligned with human preferences. More generally, my concern is that nobody fully understands exactly how and why LLMs work the way they do, and thus we should continue to exercise caution when adopting AI in high stakes environments.

Thanks for reading, and feel free to share it with a friend! 🤝

New readers can subscribe here. 💌

- Kevin

P.S. Looking for a beginner-friendly AI course? Check out Build an AI chatbot with Python ($9).