The Big Story

Nineteen different AI models can now run directly on a regular smartphone and score higher than GPT-4 on standard tests. GPT-4 used to be the single best model anyone could use, and it lived in the cloud on OpenAI’s servers. These new models are small enough to live on your phone yet still smarter at the same tasks. Think of it like moving from a giant shared computer in the library to a super-smart notebook you carry everywhere. The change matters because you no longer need fast internet or a paid subscription just to get top-level help with homework, story writing, or game ideas. A student who wants instant feedback on an essay or a quick coding hint can now get it without waiting for a server or worrying about data costs. For anyone thinking about future jobs, this also means the skills to use and improve these on-device models will be useful sooner than most people expect. You can already test several of these models yourself on Android or iOS through free or low-cost apps that let you chat with them offline. Search your app store for “on-device AI” or “local LLM” demos and try one of the free chat versions for a few minutes. Notice how fast the answers arrive compared with cloud-only tools. The tweet that shared this news comes from Ethan Mollick, who has been tracking model progress for years and noted that GPT-4 held the top spot for a full year before these smaller versions caught up. Because the models run locally, they avoid sending your prompts to distant servers, which can feel more private for everyday school or creative work. This shift also lowers the barrier for people in areas with spotty internet, since the AI keeps working even when the connection drops. Source: x.com

Explain Like I'm 14

You know how when you’re learning to ride a bike, you wobble, fall a little, then adjust your balance the next time you try? The robot hand in today’s news learns the same way, except it practices thousands of times inside a computer game first. Engineers build a pretend version of the hand inside a simulator and let it try spinning a pen or turning a cube over and over. Every time it drops the object, the program gives it a low score; every time it succeeds, it gets a higher score. After millions of these tiny experiments, the hand figures out which finger movements work best. Then the same instructions are copied to the real metal-and-plastic hand, which can now copy the moves without ever having practiced in the physical world. The surprising part is that the real hand keeps making small corrections on its own, just like you lean left or right while riding. That whole process is called reinforcement learning, and it’s why the hand can do delicate tricks instead of just opening and closing. The open-source code for this training setup is available so anyone curious can look at how the simulation connects to the real hardware. The hand has twenty independently moving joints, and the training lets it copy human motions captured by a camera or a special glove. Source: x.com

Cool Stuff & Try This

Meta’s new “Wonder” models just joined the race Meta released two fresh AI models called Muse Wonder and Muse Wonder Full that add something called active memory and active focus. These features let the model remember what you talked about earlier in the same chat and stay on topic even when the conversation gets long. The models are free to try on Meta’s platforms right now. Go to meta.ai in any browser, start a new chat, and ask it to help you plan a weekend trip while changing the destination halfway through. Watch whether it keeps track of your original plans without you repeating everything. The announcement highlights that these models aim to match the strongest current systems while running on Meta’s own services, giving users another free option alongside paid tools. Because the new abilities focus on keeping context across many turns, they work especially well for longer creative projects like writing stories or outlining school reports. Source: theverge.com

An AI agent that fixes problems without being asked Ethan Mollick shared a short video of his AI agent noticing a signup bug on his Substack newsletter and quietly testing it on its own. The agent figured out the mistake and even suggested a fix before Ethan had to do anything. This kind of helpful background work is still new, but it shows where personal AI helpers are heading. You can try a similar free agent today at the Poe.com website. Create a free account, pick the “Claude” or “Gemini” bot, and give it a simple task like “check if this link works and tell me what’s wrong.” The agent will attempt the task step by step and report back, showing how these tools can handle small chores without constant instructions. Source: x.com

Quick Bits

People are starting to like AI music even when they wish they didn’t A Reddit thread is full of listeners who keep discovering songs they enjoy, only to learn later that the music was made by AI. The comments show a mix of curiosity and worry about what this means for human artists. Many users describe the same pattern: they add a track to a playlist, then feel conflicted once they realize no person performed it. Source: reddit.com

💬 Reply to this email — Patrick reads every one.

Share: X · LinkedIn · WhatsApp

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