The Big Story

A researcher from Google Deepmind shared something surprising about how AI models learn from each other. When one model is used to help train the next one, the new model can pick up odd habits or behaviors from the older one, and those habits are difficult to filter out completely.

Think of it like a student teacher who learned a quirky way of solving problems from their own mentor; even if the school tries to correct it, some of that style sticks around. The process is called distillation, where a smaller or newer model learns by studying the outputs of a larger one instead of starting from scratch with raw data.

This discovery matters because it could explain why different models from the same family often feel similar in personality or style, even when companies try to make them distinct. For someone just starting to explore AI, it shows that these systems aren't blank slates—they carry forward patterns in ways we don't fully control yet.

On a personal level, it means the AI chatbots or tools you use in school projects or creative hobbies might have hidden quirks that come from earlier versions, affecting how reliable or consistent they feel. It also highlights why safety research is so important as models get passed along.

Right now there isn't a public demo tied directly to this finding, but you can explore how models behave differently by trying the same prompt in free versions of ChatGPT, Claude, or Gemini and noticing the tone shifts. Source: x.com

Explain Like I'm 14

You know how when you're texting a friend, your phone suggests the next word based on what you've typed so far, and sometimes it starts copying your own texting style without you noticing?

Now imagine that instead of just one phone, a whole group of phones is learning from each other: the newer phones study the messages the older phones send and try to copy their patterns. Step by step, the new phones get faster at predicting words, but they also start using the same odd phrases or shortcuts the old phones liked, even if nobody wanted those shortcuts to spread.

If someone tries to clean up the new phones by removing the weird phrases, it doesn't always work because the habit is baked into how the phones learned the patterns in the first place. And that's basically what happens during distillation when one AI model trains the next one—the new model absorbs not just the useful knowledge but also the strange little behaviors.

So next time someone says "model distillation," you can tell them it's basically phones copying each other's texting quirks. Not so scary, right?

Cool Stuff & Try This

AI tools that actually help people who make movies and photos When someone who already knows film or photography uses AI, they're not just hitting a button for quick results—they're using it to experiment with new ideas and bring back lessons to their real work. This is different from someone casually watching videos on social media, because the creator is actively studying and creating with the tool to level up their skills.

It's exciting because friends in the film industry are now finishing more short projects on their own than they did in years of traditional jobs, and the same pattern shows up in photography, VFX, and storytelling. If you like making videos or editing photos for fun or school projects, this means AI can become a practice partner rather than just a shortcut.

You can try it by opening a free image or video tool like the built-in editor in CapCut or the image features in ChatGPT and giving it a specific creative prompt such as "turn this photo into a moody film still with dramatic lighting." Notice how the output gives you ideas you can then tweak yourself. Source: x.com

A video AI that remembers what's around the corner Microsoft Research built a system called Mirage that lets video-generating AI keep track of a scene's layout even when the camera moves around for a long time. Instead of forgetting details and creating glitches, it stores the space information in a smart way that uses less computer power.

This is cool because it could make AI-generated videos feel more like real movies where the world stays consistent as characters walk through rooms or streets. For anyone who enjoys making short films or animations on their phone, it points to tools that will soon feel less random and more reliable.

You can experiment with current free video tools like Runway's free tier or Pika Labs by typing a prompt that involves a moving camera, such as "a character walking through a hallway while the camera follows from behind," and see how well the background stays steady. Source: the-decoder.com

Quick Bits

When AI picks up "toast" as its favorite word One model loved using software and design jargon so much that when it went offline, the number of times the word "toast" appeared in generated code dropped sharply. It shows how each AI develops its own quirky vocabulary. Source: x.com

Mixing smaller models to match big ones A discussion about combining several smaller, cheaper AI models found they can perform almost as well as a single top-tier model in some cases. This hints that future tools might not always need the biggest, most expensive systems. Source: x.com

💬 Reply to this email — Patrick reads every one.

Share: X · LinkedIn · WhatsApp

Forwarded this email? Subscribe here — it's free.