Happy Sunday,

It’s time to unplug, sip something decent, and catch up on what the robots did while you were trying to live your life. This is ‘AI Weekly Insights’ #78, your weekly download of everything that actually matters in AI. This week, OpenAI gives ChatGPT a coding agent that ships PRs while you eat lunch, Microsoft teaches Windows to listen for “Hey Copilot,” Fortnite lets you talk to Darth Vader (and yes, he talks back), and Stability brings AI-generated audio to your pocket.

From voice assistants getting smarter to dev workflows getting delegated, it’s clear: the line between tool and teammate is starting to blur.

Ready? Let’s dive in!

The Insights

For the Week of 05/11/25 - 05/17/25 (P.S. Click the story’s title for more information 😊):

OpenAI’s Codex Arrives to Help You Write and Fix Code

• What’s New: OpenAI has released "Codex", a new AI coding agent that can independently handle programming tasks, fix bugs, and suggest improvements directly in a codebase.

• How Codex Works: Codex runs each task in its own private, secure environment that is connected to a copy of your code. You can ask it to do something like “refactor the login page” or “add tests for checkout,” and it gets to work on its own. It can read and edit files, run commands, and check its own output. Once a task is finished, Codex shows you what it changed and any logs that explain what happened. You can review everything and accept the changes or request edits as needed. You can also give Codex extra guidance by adding a special file called AGENTS.md to your project. This tells it how to test, what rules to follow, and how your team likes to work. OpenAI also updated its Codex CLI tool so developers can get quick help in the terminal using a smaller, faster version of the same model.

• Why It Matters: Codex shifts the idea of AI from a helpful tool to a true coding teammate. It does more than fill in code; it handles entire tasks, shows its work, and fits into your workflow like a junior developer you don't have to micromanage. That makes it easier for engineers to stay focused and hand off the repetitive parts of the job. It also sets a new bar for transparency, with logs and test results built into every action. As developers get more comfortable assigning real work to AI, the idea of asynchronous, multi-agent collaboration starts to feel less like a research paper and more like the new normal. If this sticks, the job of software engineering may shift toward defining tasks clearly and reviewing output, instead of writing every line by hand.

You Can Talk to Darth Vader in Fortnite, and He’ll Talk Back

• What's New: Players can now speak to Darth Vader in Fortnite using their microphones, and he responds in real time with a voice that sounds just like James Earl Jones, thanks to AI.

• Behind the Feature: During Fortnite’s latest Star Wars event, players who defeat Vader on the map can recruit him as a teammate. Once he’s on your team, you can talk to him using voice chat, and he’ll answer back using an AI-generated version of James Earl Jones’ voice. The voice tech is powered by Google’s Gemini model and ElevenLabs' Flash voice system. Epic says it does not store player audio or use conversations to train AI. However, it does log Vader’s replies in case players report anything inappropriate. Players under 13, or under their country’s age of consent, need parental permission to use the feature. On launch day, Vader was briefly caught using strong language on Twitch, which led to a quick patch that cleaned up what he could say.

• Why it Matters: Letting players talk directly to Darth Vader is more than a cool moment, it is a glimpse of how AI could reshape character interaction in games. This isn’t a pre-written line or a triggered animation. It is AI-generated dialogue from one of the most famous voices in media history. That creates exciting new possibilities for immersion but also introduces new risks. Game developers are now moderating what AI might say in real time, something that’s never been done at this scale. With support from James Earl Jones’ family, this experiment shows what is possible when legacy voices are treated with care. But it also pushes developers to rethink how characters are written, monitored, and experienced. The idea of talking with your favorite characters may be thrilling now, but soon it could become the standard.

You Can Now Say “Hey Copilot” to Start Chatting on Windows 11

• What's New: Microsoft is testing a new feature that lets Windows 11 users launch Copilot Voice by saying “Hey Copilot.” It is currently rolling out to Windows Insiders.

• Hands-Free Access: Once enabled, saying “Hey Copilot” brings up a floating microphone on your screen and starts a voice conversation. You can ask questions, get answers, and stay in your workflow without clicking or typing. The wake word detection works offline and uses an on-device audio buffer that does not store or upload your voice. To respond to your question, Copilot still needs an internet connection for cloud processing. The feature is optional and turned off by default, so users will need to go into the Copilot app settings to switch it on. Right now, it only supports English and is being rolled out gradually to testers worldwide.

• Why it Matters: Voice assistants have been around for years, but Microsoft bringing a hands-free wake word to Windows gives Copilot a much stronger presence on devices people use every day. Saying “Hey Copilot” turns the AI into something ambient, something that is just there when you need it. That starts to blur the line between an app and an assistant. With this update, Copilot feels less like a separate tool and more like a baked-in part of the OS. For Microsoft, it is a quiet but important shift. As Copilot becomes easier to use without clicking or typing, it could start changing how people expect to interact with their computer. And for other voice tools on desktop, the race to feel invisible just got more urgent.

Stability AI Launches Audio-Generating Model That Works on Phones

• What's New: Stability AI has released Stable Audio Open Small, a lightweight AI model that can generate short audio clips directly on smartphones. It’s the first of its kind to run locally without needing the cloud.

• Small, Fast, and Royalty-Free: Built in collaboration with chipmaker Arm, this model can create short sound clips like drum loops and instrument riffs in under 8 seconds. It runs on devices powered by Arm CPUs, which includes most smartphones and tablets. Unlike most music AIs that need internet access and use copyrighted training data, Stable Audio Open Small is trained entirely on royalty-free content from Free Music Archive and Freesound. That means developers and hobbyists can use it more freely, though the model has some limits. It only supports English prompts, does not handle vocals or full songs well, and performs better with Western-style music. The model is free for small teams and individuals, but larger companies will need an enterprise license to use.

• Why it Matters: Most AI music tools today live in the cloud, which means they are slow to use, need constant internet access, and come with privacy tradeoffs. Putting audio generation directly on smartphones flips that model completely. It makes creative tools faster, more accessible, and more useful in places where internet access is spotty or expensive. For developers, it is a step toward real-time music features that work inside mobile games, apps, or DAWs without needing a server. The fact that this model only uses royalty-free sounds also sets it apart from competitors like Suno and Udio, which face questions about training on copyrighted music. That helps avoid legal headaches and gives small creators more freedom to experiment. It is not perfect, but it hints at a future where AI music becomes a normal part of everyday tools. The next time someone wants to add a beat or sound effect, they might not open a file library. They might just ask their phone.

That’s it for this week. The future of AI isn’t just showing up in lab papers, it’s landing in game characters, desktop assistants, and codebases near you. Whether you're leading the change or just trying to keep up, one thing is clear: things aren’t slowing down.

Stay curious, stay skeptical, and don’t forget to review your AI’s pull requests.

Catch you next Sunday.

Warm regards,
Kharee