Welcome back! Why can’t Big Tech seem to quit on smart glasses? Meta just announced a new pair starting at $299, cheaper than its entry-level Ray-Ban Meta glasses. Mark Zuckerberg keeps trying to make wearables feel like a normal consumer product, despite the fact that they’re hard to build and awkward to wear.
The bet seems to be that glasses are the next great computing interface. That is, if someone can figure out the one feature that makes people actually want a computer on their face.
What do you think about wearables?
The AI Skills Worth Installing This Week
If you've been using Claude Code, Codex, Cowork, OpenClaw, Hermes, or any other coding agent, you've probably noticed something: the people getting incredible results aren't just better prompters. They're using skills to give their agents repeatable, dialed-in behavior.
A skill is basically a pre-written, battle-tested prompt saved as a SKILL.md file. When you invoke it, the model reads both the skill and your prompt together, so you get consistent results without re-explaining context every time. Most of these are universal, meaning they work across just about any agent harness. And installing one is as simple as opening a new chat and saying "install this for me" with the GitHub URL.
Here are my favorites right now:
GStack: Built by Garry Tan, CEO of Y Combinator. A bundle of 23 specialists turns your agent into a virtual engineering team: a CEO who pressure-tests product ideas, a designer who catches AI slop, a security officer who runs OWASP audits, and a release engineer who ships the PR. Run /gstack-office-hours and stress-test any idea before writing a line of code.
Stop Slop: A tiny skill that strips the AI tells out of AI-generated writing. Paste something from ChatGPT, run the skill, get cleaner prose. If you use AI for scripts or blog posts, this earns its keep on day one.
Graphify: Probably my favorite on this list. It turns codebases, docs, or your second brain into a queryable knowledge graph. You get a beautiful visual web of how everything connects, and your agent can query the graph as memory instead of re-reading every file each time. The README claims 71x fewer tokens per session on big projects.
Last 30 Days: Real-time sentiment research across Reddit, X, YouTube, Hacker News, Polymarket, and the open web. Type /last30days research [topic] and get a synthesized read on what people are actually excited about, confused about, and speculating on.
My POV: The magic of skills isn't just that they give your agent more context. It's that they give it repeatable behavior—every time you invoke a skill, the agent acts the same way. That's a fundamentally different relationship than "fresh prompt every time." It's the difference between teaching a new intern from scratch every morning and having a teammate who already knows your standards.
If you've been doing serious work with these tools and haven't installed a single skill yet, you're leaving a lot of value on the table. My suggestion: pick one or two from this list and give them a try next week.
— Matt
OpenAI Unveils First Custom AI Chip
Via The Verge
OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom AI accelerator chip, built specifically for running AI models after they’ve already been trained.
Meet Jalapeño, OpenAI’s first custom in-house AI processor. The codename continues a familiar pattern in AI with highly technical projects often having oddly playful food names (from OpenAI’s Strawberry to Google’s dessert-themed Android releases).
The new chip is an ASIC, or application-specific integrated circuit, designed for inference—the expensive part that happens every time someone uses ChatGPT, Codex, the API, or a future OpenAI agent.
Bloomberg reported that the chip could cut inference costs by about 50%, which matters a lot when you’re serving AI products at such a massive scale.
AI building AI: Jalapeño reportedly went from early schematics to fabrication readiness in about nine months, an unusually fast timeline in the semiconductor world. OpenAI says its own models helped speed up parts of the chip design process. Put another way, the company used AI to help build the chip that will help run future AI … 🤯
Why it matters: OpenAI will still rely on Nvidia and other chip partners, but Jalapeño gives it more control over the cost and performance of its own products—a factor that’s increasingly important in the AI race.
Figma Pushes Design Deeper Into Code
Figma announced a major update that adds code layers, support for motion and shaders, and more AI features directly inside its design platform.
What’s new: The biggest addition is code layers, which bring code directly into Figma’s collaborative canvas. Teams can clone repositories, extract flows from code into design layers, and test ideas without leaving the design environment. Figma says the goal is not to make designers write pristine production code but to help them explore ideas faster in the same multiplayer canvas.
A few other updates:
Support for animations, transitions, and 3D transforms
AI-generated shader effects and fills
New AI assistant skills for repeatable workflows
Integrations with tools like Notion, Granola, Excel, and GitHub
The bigger picture: AI is turning software creation into a more fluid process. Design tools are becoming coding tools. Coding tools are becoming design tools. And the teams that used to hand work back and forth are starting to work in the same canvas. Figma clearly does not want to be just where products are designed—it wants to be where they start getting built.
Stop Losing Money on 5 Different AI Tools. Use One.
Coding? Servers? APIs? All of the sounds like a lot of work for an AI agent that doesn’t even fit your exact needs. That’s where Nexos comes in.
Nexos lets you access all of the top AI models in one place to create your perfect digital assistant that can:
Take all of your messy notes and files summarize it, pull out the action items, draft any follow-up emails, and build out a priority list for the whole week
Transform your rough outline for a presentation into a structured slide deck with talking points and even speaker notes
Build a custom AI agent for whatever you need based on your description alone—no coding needed
Ready for your custom AI agent? Stop losing money on 5 different AI tools—get 50% off Nexos now and save up to $200 a month.
Find out which AI models your computer can run
Via CanIRun.ai
CanIRun.ai is a browser-based hardware checker that scans your GPU, CPU, RAM, and WebGPU support, then compares your setup against a catalog of AI, LLM, and multimodal model requirements. It grades your machine’s AI-readiness and ranks models by expected performance.
How you can use it
Check which local models your computer can actually run
Avoid downloading models your hardware can’t support
See whether a GPU upgrade is worth it
Pricing: Free
AI audio editing and music generation
Audjust AI is a web-based audio editor and music generator that can shorten or lengthen tracks while preserving musical structure. It can find loop points, convert audio to MIDI for editing, and generate songs from text, lyrics, or images.
How you can use it
Extend or shorten music without awkward cuts
Create seamless loops for videos or podcasts
Generate royalty-free songs from prompts
Fix timing problems without opening a full DAW
Pricing: Paid
Jobs, announcements, and big ideas
OpenAI staggers its GPT-5.6 rollout after the US government flags national security concerns.
Google revamps Finance with AI-powered insights and a brand-new standalone app.
Runway launches Agent 2.0, generating marketing briefs and full campaign assets on demand.
Google introduces Computer Use in Gemini 3.5 Flash, letting the model operate interfaces directly.
Perplexity debuts Computer for Counsel, an AI agent built for legal professionals.
May I present: 9 free skills and plug-ins that make Claude Code, Codex, and other AI coding agents dramatically more useful. And yes, all of them are free.
That’s a wrap! See you next week for more.
